Mercurial > repos > pointsource_algs / changeset
changeset
Merge dev to default
Mon, 17 Feb 2025 13:54:53 -0500
- author
- Tuomo Valkonen <tuomov@iki.fi>
- date
- Mon, 17 Feb 2025 13:54:53 -0500
- changeset 52
- f0e8704d3f0e
- parent 31
- 6105b5cd8d89 (current diff)
- parent 51
- 0693cc9ba9f0 (diff)
- child 53
- 92cae2e8f598
Merge dev to default
--- /dev/null Thu Jan 01 00:00:00 1970 +0000 +++ b/.cargo/config.toml Mon Feb 17 13:54:53 2025 -0500 @@ -0,0 +1,2 @@ +[target.'cfg(all(target_os = "macos"))'] +rustflags = ["-L", "/opt/homebrew/include"]
--- a/Cargo.lock Tue Aug 01 10:25:09 2023 +0300 +++ b/Cargo.lock Mon Feb 17 13:54:53 2025 -0500 @@ -1,12 +1,12 @@ # This file is automatically @generated by Cargo. # It is not intended for manual editing. -version = 3 +version = 4 [[package]] name = "GSL" -version = "6.0.0" +version = "7.0.0" source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "c9becaf6d7d1ba36a457288e661fa6a0472e8328629276f45369eafcd48ef1ce" +checksum = "db3943d5a15b5c46e991124abee6a1bc89c7c9ffb25dbb8aeb4eab926fd9b307" dependencies = [ "GSL-sys", "paste", @@ -23,24 +23,19 @@ ] [[package]] -name = "adler" -version = "1.0.2" +name = "aho-corasick" +version = "1.1.3" source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "f26201604c87b1e01bd3d98f8d5d9a8fcbb815e8cedb41ffccbeb4bf593a35fe" - -[[package]] -name = "aho-corasick" -version = "0.7.20" -source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "cc936419f96fa211c1b9166887b38e5e40b19958e5b895be7c1f93adec7071ac" +checksum = "8e60d3430d3a69478ad0993f19238d2df97c507009a52b3c10addcd7f6bcb916" dependencies = [ "memchr", ] [[package]] name = "alg_tools" -version = "0.1.0" +version = "0.3.0" dependencies = [ + "anyhow", "colored", "cpu-time", "csv", @@ -52,10 +47,16 @@ "rayon", "serde", "serde_json", - "trait-set", + "simba", ] [[package]] +name = "android-tzdata" +version = "0.1.1" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "e999941b234f3131b00bc13c22d06e8c5ff726d1b6318ac7eb276997bbb4fef0" + +[[package]] name = "android_system_properties" version = "0.1.5" source = "registry+https://github.com/rust-lang/crates.io-index" @@ -65,6 +66,61 @@ ] [[package]] +name = "anstream" +version = "0.6.18" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "8acc5369981196006228e28809f761875c0327210a891e941f4c683b3a99529b" +dependencies = [ + "anstyle", + "anstyle-parse", + "anstyle-query", + "anstyle-wincon", + "colorchoice", + "is_terminal_polyfill", + "utf8parse", +] + +[[package]] +name = "anstyle" +version = "1.0.10" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "55cc3b69f167a1ef2e161439aa98aed94e6028e5f9a59be9a6ffb47aef1651f9" + +[[package]] +name = "anstyle-parse" +version = "0.2.6" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "3b2d16507662817a6a20a9ea92df6652ee4f94f914589377d69f3b21bc5798a9" +dependencies = [ + "utf8parse", +] + +[[package]] +name = "anstyle-query" +version = "1.1.2" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "79947af37f4177cfead1110013d678905c37501914fba0efea834c3fe9a8d60c" +dependencies = [ + "windows-sys 0.59.0", +] + +[[package]] +name = "anstyle-wincon" +version = "3.0.6" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "2109dbce0e72be3ec00bed26e6a7479ca384ad226efdd66db8fa2e3a38c83125" +dependencies = [ + "anstyle", + "windows-sys 0.59.0", +] + +[[package]] +name = "anyhow" +version = "1.0.95" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "34ac096ce696dc2fcabef30516bb13c0a68a11d30131d3df6f04711467681b04" + +[[package]] name = "approx" version = "0.5.1" source = "registry+https://github.com/rust-lang/crates.io-index" @@ -74,69 +130,43 @@ ] [[package]] -name = "atty" -version = "0.2.14" -source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "d9b39be18770d11421cdb1b9947a45dd3f37e93092cbf377614828a319d5fee8" -dependencies = [ - "hermit-abi 0.1.19", - "libc", - "winapi", -] - -[[package]] name = "autocfg" version = "1.1.0" source = "registry+https://github.com/rust-lang/crates.io-index" checksum = "d468802bab17cbc0cc575e9b053f41e72aa36bfa6b7f55e3529ffa43161b97fa" [[package]] -name = "bit_field" -version = "0.10.1" +name = "base64" +version = "0.22.1" source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "dcb6dd1c2376d2e096796e234a70e17e94cc2d5d54ff8ce42b28cef1d0d359a4" +checksum = "72b3254f16251a8381aa12e40e3c4d2f0199f8c6508fbecb9d91f575e0fbb8c6" [[package]] name = "bitflags" -version = "1.3.2" -source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "bef38d45163c2f1dde094a7dfd33ccf595c92905c8f8f4fdc18d06fb1037718a" - -[[package]] -name = "bstr" -version = "0.2.17" +version = "2.4.0" source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "ba3569f383e8f1598449f1a423e72e99569137b47740b1da11ef19af3d5c3223" -dependencies = [ - "lazy_static", - "memchr", - "regex-automata", - "serde", -] +checksum = "b4682ae6287fcf752ecaabbfcc7b6f9b72aa33933dc23a554d853aea8eea8635" [[package]] name = "bumpalo" -version = "3.11.1" +version = "3.16.0" source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "572f695136211188308f16ad2ca5c851a712c464060ae6974944458eb83880ba" +checksum = "79296716171880943b8470b5f8d03aa55eb2e645a4874bdbb28adb49162e012c" [[package]] name = "bytemuck" -version = "1.12.3" +version = "1.20.0" source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "aaa3a8d9a1ca92e282c96a32d6511b695d7d994d1d102ba85d279f9b2756947f" - -[[package]] -name = "byteorder" -version = "1.4.3" -source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "14c189c53d098945499cdfa7ecc63567cf3886b3332b312a5b4585d8d3a6a610" +checksum = "8b37c88a63ffd85d15b406896cc343916d7cf57838a847b3a6f2ca5d39a5695a" [[package]] name = "cc" -version = "1.0.77" +version = "1.2.0" source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "e9f73505338f7d905b19d18738976aae232eb46b8efc15554ffc56deb5d9ebe4" +checksum = "1aeb932158bd710538c73702db6945cb68a8fb08c519e6e12706b94263b36db8" +dependencies = [ + "shlex", +] [[package]] name = "cfg-if" @@ -146,33 +176,39 @@ [[package]] name = "chrono" -version = "0.4.23" +version = "0.4.39" source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "16b0a3d9ed01224b22057780a37bb8c5dbfe1be8ba48678e7bf57ec4b385411f" +checksum = "7e36cc9d416881d2e24f9a963be5fb1cd90966419ac844274161d10488b3e825" dependencies = [ + "android-tzdata", "iana-time-zone", "js-sys", - "num-integer", "num-traits", "serde", - "time", "wasm-bindgen", - "winapi", + "windows-targets 0.52.6", ] [[package]] name = "clap" -version = "4.0.27" +version = "4.5.23" source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "0acbd8d28a0a60d7108d7ae850af6ba34cf2d1257fc646980e5f97ce14275966" +checksum = "3135e7ec2ef7b10c6ed8950f0f792ed96ee093fa088608f1c76e569722700c84" dependencies = [ - "bitflags", + "clap_builder", "clap_derive", +] + +[[package]] +name = "clap_builder" +version = "4.5.23" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "30582fc632330df2bd26877bde0c1f4470d57c582bbc070376afcd04d8cb4838" +dependencies = [ + "anstream", + "anstyle", "clap_lex", - "is-terminal", - "once_cell", "strsim", - "termcolor", "terminal_size", "unicase", "unicode-width", @@ -180,67 +216,43 @@ [[package]] name = "clap_derive" -version = "4.0.21" +version = "4.5.18" source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "0177313f9f02afc995627906bbd8967e2be069f5261954222dac78290c2b9014" +checksum = "4ac6a0c7b1a9e9a5186361f67dfa1b88213572f427fb9ab038efb2bd8c582dab" dependencies = [ "heck", - "proc-macro-error", "proc-macro2", "quote", - "syn", + "syn 2.0.93", ] [[package]] name = "clap_lex" -version = "0.3.0" +version = "0.7.4" source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "0d4198f73e42b4936b35b5bb248d81d2b595ecb170da0bac7655c54eedfa8da8" -dependencies = [ - "os_str_bytes", -] +checksum = "f46ad14479a25103f283c0f10005961cf086d8dc42205bb44c46ac563475dca6" [[package]] -name = "codespan-reporting" -version = "0.11.1" +name = "colorchoice" +version = "1.0.3" source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "3538270d33cc669650c4b093848450d380def10c331d38c768e34cac80576e6e" -dependencies = [ - "termcolor", - "unicode-width", -] - -[[package]] -name = "color_quant" -version = "1.1.0" -source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "3d7b894f5411737b7867f4827955924d7c254fc9f4d91a6aad6b097804b1018b" - -[[package]] -name = "colorbrewer" -version = "0.2.0" -source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "525be5012d97bc222e124ded87f18601e6fbd24a406761bcb1664475663919a6" -dependencies = [ - "rgb", -] +checksum = "5b63caa9aa9397e2d9480a9b13673856c78d8ac123288526c37d7839f2a86990" [[package]] name = "colored" -version = "2.0.0" +version = "2.1.0" source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "b3616f750b84d8f0de8a58bda93e08e2a81ad3f523089b05f1dffecab48c6cbd" +checksum = "cbf2150cce219b664a8a70df7a1f933836724b503f8a413af9365b4dcc4d90b8" dependencies = [ - "atty", "lazy_static", - "winapi", + "windows-sys 0.48.0", ] [[package]] name = "core-foundation-sys" -version = "0.8.3" +version = "0.8.7" source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "5827cebf4670468b8772dd191856768aedcb1b0278a04f989f7766351917b9dc" +checksum = "773648b94d0e5d620f64f280777445740e61fe701025087ec8b57f45c791888b" [[package]] name = "cpu-time" @@ -253,179 +265,116 @@ ] [[package]] -name = "crc32fast" -version = "1.3.2" -source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "b540bd8bc810d3885c6ea91e2018302f68baba2129ab3e88f32389ee9370880d" -dependencies = [ - "cfg-if", -] - -[[package]] -name = "crossbeam-channel" -version = "0.5.6" +name = "crossbeam-deque" +version = "0.8.6" source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "c2dd04ddaf88237dc3b8d8f9a3c1004b506b54b3313403944054d23c0870c521" +checksum = "9dd111b7b7f7d55b72c0a6ae361660ee5853c9af73f70c3c2ef6858b950e2e51" dependencies = [ - "cfg-if", - "crossbeam-utils", -] - -[[package]] -name = "crossbeam-deque" -version = "0.8.2" -source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "715e8152b692bba2d374b53d4875445368fdf21a94751410af607a5ac677d1fc" -dependencies = [ - "cfg-if", "crossbeam-epoch", "crossbeam-utils", ] [[package]] name = "crossbeam-epoch" -version = "0.9.13" +version = "0.9.18" source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "01a9af1f4c2ef74bb8aa1f7e19706bc72d03598c8a570bb5de72243c7a9d9d5a" +checksum = "5b82ac4a3c2ca9c3460964f020e1402edd5753411d7737aa39c3714ad1b5420e" dependencies = [ - "autocfg", - "cfg-if", "crossbeam-utils", - "memoffset", - "scopeguard", ] [[package]] name = "crossbeam-utils" -version = "0.8.14" +version = "0.8.21" source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "4fb766fa798726286dbbb842f174001dab8abc7b627a1dd86e0b7222a95d929f" -dependencies = [ - "cfg-if", -] - -[[package]] -name = "crunchy" -version = "0.2.2" -source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "7a81dae078cea95a014a339291cec439d2f232ebe854a9d672b796c6afafa9b7" +checksum = "d0a5c400df2834b80a4c3327b3aad3a4c4cd4de0629063962b03235697506a28" [[package]] name = "csv" -version = "1.1.6" +version = "1.3.1" source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "22813a6dc45b335f9bade10bf7271dc477e81113e89eb251a0bc2a8a81c536e1" +checksum = "acdc4883a9c96732e4733212c01447ebd805833b7275a73ca3ee080fd77afdaf" dependencies = [ - "bstr", "csv-core", - "itoa 0.4.8", + "itoa", "ryu", "serde", ] [[package]] name = "csv-core" -version = "0.1.10" +version = "0.1.11" source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "2b2466559f260f48ad25fe6317b3c8dac77b5bdb5763ac7d9d6103530663bc90" +checksum = "5efa2b3d7902f4b634a20cae3c9c4e6209dc4779feb6863329607560143efa70" dependencies = [ "memchr", ] [[package]] -name = "cxx" -version = "1.0.82" +name = "darling" +version = "0.20.10" source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "d4a41a86530d0fe7f5d9ea779916b7cadd2d4f9add748b99c2c029cbbdfaf453" +checksum = "6f63b86c8a8826a49b8c21f08a2d07338eec8d900540f8630dc76284be802989" dependencies = [ - "cc", - "cxxbridge-flags", - "cxxbridge-macro", - "link-cplusplus", + "darling_core", + "darling_macro", ] [[package]] -name = "cxx-build" -version = "1.0.82" +name = "darling_core" +version = "0.20.10" source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "06416d667ff3e3ad2df1cd8cd8afae5da26cf9cec4d0825040f88b5ca659a2f0" +checksum = "95133861a8032aaea082871032f5815eb9e98cef03fa916ab4500513994df9e5" dependencies = [ - "cc", - "codespan-reporting", - "once_cell", + "fnv", + "ident_case", "proc-macro2", "quote", - "scratch", - "syn", + "strsim", + "syn 2.0.93", ] [[package]] -name = "cxxbridge-flags" -version = "1.0.82" +name = "darling_macro" +version = "0.20.10" source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "820a9a2af1669deeef27cb271f476ffd196a2c4b6731336011e0ba63e2c7cf71" +checksum = "d336a2a514f6ccccaa3e09b02d41d35330c07ddf03a62165fcec10bb561c7806" +dependencies = [ + "darling_core", + "quote", + "syn 2.0.93", +] [[package]] -name = "cxxbridge-macro" -version = "1.0.82" +name = "deranged" +version = "0.3.11" source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "a08a6e2fcc370a089ad3b4aaf54db3b1b4cee38ddabce5896b33eb693275f470" +checksum = "b42b6fa04a440b495c8b04d0e71b707c585f83cb9cb28cf8cd0d976c315e31b4" dependencies = [ - "proc-macro2", - "quote", - "syn", + "powerfmt", + "serde", ] [[package]] name = "either" -version = "1.8.0" +version = "1.13.0" source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "90e5c1c8368803113bf0c9584fc495a58b86dc8a29edbf8fe877d21d9507e797" +checksum = "60b1af1c220855b6ceac025d3f6ecdd2b7c4894bfe9cd9bda4fbb4bc7c0d4cf0" + +[[package]] +name = "equivalent" +version = "1.0.1" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "5443807d6dff69373d433ab9ef5378ad8df50ca6298caf15de6e52e24aaf54d5" [[package]] name = "errno" -version = "0.2.8" +version = "0.3.5" source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "f639046355ee4f37944e44f60642c6f3a7efa3cf6b78c78a0d989a8ce6c396a1" +checksum = "ac3e13f66a2f95e32a39eaa81f6b95d42878ca0e1db0c7543723dfe12557e860" dependencies = [ - "errno-dragonfly", - "libc", - "winapi", -] - -[[package]] -name = "errno-dragonfly" -version = "0.1.2" -source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "aa68f1b12764fab894d2755d2518754e71b4fd80ecfb822714a1206c2aab39bf" -dependencies = [ - "cc", "libc", -] - -[[package]] -name = "exr" -version = "1.5.2" -source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "8eb5f255b5980bb0c8cf676b675d1a99be40f316881444f44e0462eaf5df5ded" -dependencies = [ - "bit_field", - "flume", - "half", - "lebe", - "miniz_oxide", - "smallvec", - "threadpool", -] - -[[package]] -name = "flate2" -version = "1.0.25" -source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "a8a2db397cb1c8772f31494cb8917e48cd1e64f0fa7efac59fbd741a0a8ce841" -dependencies = [ - "crc32fast", - "miniz_oxide", + "windows-sys 0.48.0", ] [[package]] @@ -438,193 +387,125 @@ ] [[package]] -name = "flume" -version = "0.10.14" +name = "fnv" +version = "1.0.7" source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "1657b4441c3403d9f7b3409e47575237dac27b1b5726df654a6ecbf92f0f7577" -dependencies = [ - "futures-core", - "futures-sink", - "nanorand", - "pin-project", - "spin", -] - -[[package]] -name = "futures-core" -version = "0.3.25" -source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "04909a7a7e4633ae6c4a9ab280aeb86da1236243a77b694a49eacd659a4bd3ac" - -[[package]] -name = "futures-sink" -version = "0.3.25" -source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "39c15cf1a4aa79df40f1bb462fb39676d0ad9e366c2a33b590d7c66f4f81fcf9" +checksum = "3f9eec918d3f24069decb9af1554cad7c880e2da24a9afd88aca000531ab82c1" [[package]] name = "getrandom" -version = "0.2.8" +version = "0.2.10" source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "c05aeb6a22b8f62540c194aac980f2115af067bfe15a0734d7277a768d396b31" +checksum = "be4136b2a15dd319360be1c07d9933517ccf0be8f16bf62a3bee4f0d618df427" dependencies = [ "cfg-if", - "js-sys", "libc", - "wasi 0.11.0+wasi-snapshot-preview1", - "wasm-bindgen", + "wasi", ] [[package]] -name = "gif" -version = "0.11.4" +name = "hashbrown" +version = "0.12.3" source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "3edd93c6756b4dfaf2709eafcc345ba2636565295c198a9cfbf75fa5e3e00b06" -dependencies = [ - "color_quant", - "weezl", -] +checksum = "8a9ee70c43aaf417c914396645a0fa852624801b24ebb7ae78fe8272889ac888" [[package]] -name = "half" -version = "2.1.0" +name = "hashbrown" +version = "0.15.2" source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "ad6a9459c9c30b177b925162351f97e7d967c7ea8bab3b8352805327daf45554" -dependencies = [ - "crunchy", -] +checksum = "bf151400ff0baff5465007dd2f3e717f3fe502074ca563069ce3a6629d07b289" [[package]] name = "heck" -version = "0.4.0" +version = "0.5.0" source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "2540771e65fc8cb83cd6e8a237f70c319bd5c29f78ed1084ba5d50eeac86f7f9" +checksum = "2304e00983f87ffb38b55b444b5e3b60a884b5d30c0fca7d82fe33449bbe55ea" [[package]] -name = "hermit-abi" -version = "0.1.19" +name = "hex" +version = "0.4.3" source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "62b467343b94ba476dcb2500d242dadbb39557df889310ac77c5d99100aaac33" -dependencies = [ - "libc", -] - -[[package]] -name = "hermit-abi" -version = "0.2.6" -source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "ee512640fe35acbfb4bb779db6f0d80704c2cacfa2e39b601ef3e3f47d1ae4c7" -dependencies = [ - "libc", -] +checksum = "7f24254aa9a54b5c858eaee2f5bccdb46aaf0e486a595ed5fd8f86ba55232a70" [[package]] name = "iana-time-zone" -version = "0.1.53" +version = "0.1.61" source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "64c122667b287044802d6ce17ee2ddf13207ed924c712de9a66a5814d5b64765" +checksum = "235e081f3925a06703c2d0117ea8b91f042756fd6e7a6e5d901e8ca1a996b220" dependencies = [ "android_system_properties", "core-foundation-sys", "iana-time-zone-haiku", "js-sys", "wasm-bindgen", - "winapi", + "windows-core", ] [[package]] name = "iana-time-zone-haiku" -version = "0.1.1" -source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "0703ae284fc167426161c2e3f1da3ea71d94b21bedbcc9494e92b28e334e3dca" -dependencies = [ - "cxx", - "cxx-build", -] - -[[package]] -name = "image" -version = "0.24.5" +version = "0.1.2" source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "69b7ea949b537b0fd0af141fff8c77690f2ce96f4f41f042ccb6c69c6c965945" +checksum = "f31827a206f56af32e590ba56d5d2d085f558508192593743f16b2306495269f" dependencies = [ - "bytemuck", - "byteorder", - "color_quant", - "exr", - "gif", - "jpeg-decoder", - "num-rational", - "num-traits", - "png", - "scoped_threadpool", - "tiff", + "cc", ] [[package]] -name = "io-lifetimes" -version = "0.7.5" +name = "ident_case" +version = "1.0.1" source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "59ce5ef949d49ee85593fc4d3f3f95ad61657076395cbbce23e2121fc5542074" +checksum = "b9e0384b61958566e926dc50660321d12159025e767c18e043daf26b70104c39" [[package]] -name = "io-lifetimes" -version = "1.0.2" +name = "indexmap" +version = "1.9.3" source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "e394faa0efb47f9f227f1cd89978f854542b318a6f64fa695489c9c993056656" +checksum = "bd070e393353796e801d209ad339e89596eb4c8d430d18ede6a1cced8fafbd99" dependencies = [ - "libc", - "windows-sys", + "autocfg", + "hashbrown 0.12.3", + "serde", ] [[package]] -name = "is-terminal" -version = "0.4.0" +name = "indexmap" +version = "2.7.0" source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "aae5bc6e2eb41c9def29a3e0f1306382807764b9b53112030eff57435667352d" +checksum = "62f822373a4fe84d4bb149bf54e584a7f4abec90e072ed49cda0edea5b95471f" dependencies = [ - "hermit-abi 0.2.6", - "io-lifetimes 1.0.2", - "rustix 0.36.3", - "windows-sys", + "equivalent", + "hashbrown 0.15.2", + "serde", ] [[package]] +name = "is_terminal_polyfill" +version = "1.70.1" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "7943c866cc5cd64cbc25b2e01621d07fa8eb2a1a23160ee81ce38704e97b8ecf" + +[[package]] name = "itertools" -version = "0.10.5" +version = "0.13.0" source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "b0fd2260e829bddf4cb6ea802289de2f86d6a7a690192fbe91b3f46e0f2c8473" +checksum = "413ee7dfc52ee1a4949ceeb7dbc8a33f2d6c088194d9f922fb8318faf1f01186" dependencies = [ "either", ] [[package]] name = "itoa" -version = "0.4.8" -source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "b71991ff56294aa922b450139ee08b3bfc70982c6b2c7562771375cf73542dd4" - -[[package]] -name = "itoa" -version = "1.0.4" +version = "1.0.9" source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "4217ad341ebadf8d8e724e264f13e593e0648f5b3e94b3896a5df283be015ecc" - -[[package]] -name = "jpeg-decoder" -version = "0.3.0" -source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "bc0000e42512c92e31c2252315bda326620a4e034105e900c98ec492fa077b3e" -dependencies = [ - "rayon", -] +checksum = "af150ab688ff2122fcef229be89cb50dd66af9e01a4ff320cc137eecc9bacc38" [[package]] name = "js-sys" -version = "0.3.60" +version = "0.3.76" source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "49409df3e3bf0856b916e2ceaca09ee28e6871cf7d9ce97a692cacfdb2a25a47" +checksum = "6717b6b5b077764fb5966237269cb3c64edddde4b14ce42647430a78ced9e7b7" dependencies = [ + "once_cell", "wasm-bindgen", ] @@ -635,101 +516,50 @@ checksum = "e2abad23fbc42b3700f2f279844dc832adb2b2eb069b2df918f455c4e18cc646" [[package]] -name = "lebe" -version = "0.5.2" +name = "libc" +version = "0.2.149" source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "03087c2bad5e1034e8cace5926dec053fb3790248370865f5117a7d0213354c8" - -[[package]] -name = "libc" -version = "0.2.137" -source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "fc7fcc620a3bff7cdd7a365be3376c97191aeaccc2a603e600951e452615bf89" +checksum = "a08173bc88b7955d1b3145aa561539096c421ac8debde8cbc3612ec635fee29b" [[package]] name = "libm" -version = "0.2.6" +version = "0.2.8" source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "348108ab3fba42ec82ff6e9564fc4ca0247bdccdc68dd8af9764bbc79c3c8ffb" - -[[package]] -name = "link-cplusplus" -version = "1.0.7" -source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "9272ab7b96c9046fbc5bc56c06c117cb639fe2d509df0c421cad82d2915cf369" -dependencies = [ - "cc", -] +checksum = "4ec2a862134d2a7d32d7983ddcdd1c4923530833c9f2ea1a44fc5fa473989058" [[package]] name = "linux-raw-sys" -version = "0.0.46" -source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "d4d2456c373231a208ad294c33dc5bff30051eafd954cd4caae83a712b12854d" - -[[package]] -name = "linux-raw-sys" -version = "0.1.3" +version = "0.4.10" source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "8f9f08d8963a6c613f4b1a78f4f4a4dbfadf8e6545b2d72861731e4858b8b47f" - -[[package]] -name = "lock_api" -version = "0.4.9" -source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "435011366fe56583b16cf956f9df0095b405b82d76425bc8981c0e22e60ec4df" -dependencies = [ - "autocfg", - "scopeguard", -] +checksum = "da2479e8c062e40bf0066ffa0bc823de0a9368974af99c9f6df941d2c231e03f" [[package]] name = "log" -version = "0.4.17" +version = "0.4.22" source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "abb12e687cfb44aa40f41fc3978ef76448f9b6038cad6aef4259d3c095a2382e" -dependencies = [ - "cfg-if", -] +checksum = "a7a70ba024b9dc04c27ea2f0c0548feb474ec5c54bba33a7f72f873a39d07b24" [[package]] name = "matrixmultiply" -version = "0.3.2" +version = "0.3.8" source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "add85d4dd35074e6fedc608f8c8f513a3548619a9024b751949ef0e8e45a4d84" +checksum = "7574c1cf36da4798ab73da5b215bbf444f50718207754cb522201d78d1cd0ff2" dependencies = [ + "autocfg", "rawpointer", ] [[package]] name = "memchr" -version = "2.5.0" -source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "2dffe52ecf27772e601905b7522cb4ef790d2cc203488bbd0e2fe85fcb74566d" - -[[package]] -name = "memoffset" -version = "0.7.1" +version = "2.6.4" source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "5de893c32cde5f383baa4c04c5d6dbdd735cfd4a794b0debdb2bb1b421da5ff4" -dependencies = [ - "autocfg", -] - -[[package]] -name = "miniz_oxide" -version = "0.6.2" -source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "b275950c28b37e794e8c55d88aeb5e139d0ce23fdbbeda68f8d7174abdf9e8fa" -dependencies = [ - "adler", -] +checksum = "f665ee40bc4a3c5590afb1e9677db74a508659dfd71e126420da8274909a0167" [[package]] name = "nalgebra" -version = "0.31.4" +version = "0.33.2" source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "20bd243ab3dbb395b39ee730402d2e5405e448c75133ec49cc977762c4cba3d1" +checksum = "26aecdf64b707efd1310e3544d709c5c0ac61c13756046aaaba41be5c4f66a3b" dependencies = [ "approx", "matrixmultiply", @@ -744,29 +574,20 @@ [[package]] name = "nalgebra-macros" -version = "0.1.0" +version = "0.2.2" source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "01fcc0b8149b4632adc89ac3b7b31a12fb6099a0317a4eb2ebff574ef7de7218" +checksum = "254a5372af8fc138e36684761d3c0cdb758a4410e938babcff1c860ce14ddbfc" dependencies = [ "proc-macro2", "quote", - "syn", -] - -[[package]] -name = "nanorand" -version = "0.7.0" -source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "6a51313c5820b0b02bd422f4b44776fbf47961755c74ce64afc73bfad10226c3" -dependencies = [ - "getrandom", + "syn 2.0.93", ] [[package]] name = "num" -version = "0.4.0" +version = "0.4.2" source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "43db66d1170d347f9a065114077f7dccb00c1b9478c89384490a3425279a4606" +checksum = "3135b08af27d103b0a51f2ae0f8632117b7b185ccf931445affa8df530576a41" dependencies = [ "num-bigint", "num-complex", @@ -778,39 +599,43 @@ [[package]] name = "num-bigint" -version = "0.4.3" +version = "0.4.6" source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "f93ab6289c7b344a8a9f60f88d80aa20032336fe78da341afc91c8a2341fc75f" +checksum = "a5e44f723f1133c9deac646763579fdb3ac745e418f2a7af9cd0c431da1f20b9" dependencies = [ - "autocfg", "num-integer", "num-traits", ] [[package]] name = "num-complex" -version = "0.4.2" +version = "0.4.6" source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "7ae39348c8bc5fbd7f40c727a9925f03517afd2ab27d46702108b6a7e5414c19" +checksum = "73f88a1307638156682bada9d7604135552957b7818057dcef22705b4d509495" dependencies = [ "num-traits", ] [[package]] -name = "num-integer" -version = "0.1.45" +name = "num-conv" +version = "0.1.0" source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "225d3389fb3509a24c93f5c29eb6bde2586b98d9f016636dff58d7c6f7569cd9" +checksum = "51d515d32fb182ee37cda2ccdcb92950d6a3c2893aa280e540671c2cd0f3b1d9" + +[[package]] +name = "num-integer" +version = "0.1.46" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "7969661fd2958a5cb096e56c8e1ad0444ac2bbcd0061bd28660485a44879858f" dependencies = [ - "autocfg", "num-traits", ] [[package]] name = "num-iter" -version = "0.1.43" +version = "0.1.45" source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "7d03e6c028c5dc5cac6e2dec0efda81fc887605bb3d884578bb6d6bf7514e252" +checksum = "1429034a0490724d0075ebb2bc9e875d6503c3cf69e235a8941aa757d83ef5bf" dependencies = [ "autocfg", "num-integer", @@ -831,124 +656,71 @@ [[package]] name = "num-traits" -version = "0.2.15" +version = "0.2.19" source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "578ede34cf02f8924ab9447f50c28075b4d3e5b269972345e7e0372b38c6cdcd" +checksum = "071dfc062690e90b734c0b2273ce72ad0ffa95f0c74596bc250dcfd960262841" dependencies = [ "autocfg", "libm", ] [[package]] -name = "num_cpus" -version = "1.14.0" -source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "f6058e64324c71e02bc2b150e4f3bc8286db6c83092132ffa3f6b1eab0f9def5" -dependencies = [ - "hermit-abi 0.1.19", - "libc", -] - -[[package]] name = "numeric_literals" version = "0.2.0" source = "registry+https://github.com/rust-lang/crates.io-index" checksum = "095aa67b0b9f2081746998f4f17106bdb51d56dc8c211afca5531b92b83bf98a" dependencies = [ "quote", - "syn", + "syn 1.0.109", ] [[package]] name = "once_cell" -version = "1.16.0" +version = "1.20.2" source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "86f0b0d4bf799edbc74508c1e8bf170ff5f41238e5f8225603ca7caaae2b7860" - -[[package]] -name = "os_str_bytes" -version = "6.4.1" -source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "9b7820b9daea5457c9f21c69448905d723fbd21136ccf521748f23fd49e723ee" +checksum = "1261fe7e33c73b354eab43b1273a57c8f967d0391e80353e51f764ac02cf6775" [[package]] name = "paste" -version = "1.0.9" -source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "b1de2e551fb905ac83f73f7aedf2f0cb4a0da7e35efa24a202a936269f1f18e1" - -[[package]] -name = "pin-project" -version = "1.0.12" +version = "1.0.14" source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "ad29a609b6bcd67fee905812e544992d216af9d755757c05ed2d0e15a74c6ecc" -dependencies = [ - "pin-project-internal", -] - -[[package]] -name = "pin-project-internal" -version = "1.0.12" -source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "069bdb1e05adc7a8990dce9cc75370895fbe4e3d58b9b73bf1aee56359344a55" -dependencies = [ - "proc-macro2", - "quote", - "syn", -] +checksum = "de3145af08024dea9fa9914f381a17b8fc6034dfb00f3a84013f7ff43f29ed4c" [[package]] name = "pkg-config" -version = "0.3.26" -source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "6ac9a59f73473f1b8d852421e59e64809f025994837ef743615c6d0c5b305160" - -[[package]] -name = "png" -version = "0.17.7" +version = "0.3.27" source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "5d708eaf860a19b19ce538740d2b4bdeeb8337fa53f7738455e706623ad5c638" -dependencies = [ - "bitflags", - "crc32fast", - "flate2", - "miniz_oxide", -] +checksum = "26072860ba924cbfa98ea39c8c19b4dd6a4a25423dbdf219c1eca91aa0cf6964" [[package]] name = "pointsource_algs" -version = "1.0.1" +version = "2.0.0-pre" dependencies = [ "GSL", "alg_tools", + "anyhow", "chrono", "clap", - "colorbrewer", "colored", "cpu-time", "float_extras", - "image", "itertools", "nalgebra", "num-traits", "numeric_literals", - "poloto", "rand", "rand_distr", "regex", - "rgb", "serde", "serde_json", + "serde_with", ] [[package]] -name = "poloto" -version = "3.13.1" +name = "powerfmt" +version = "0.2.0" source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "a2541c28c0622b297e342444bd8b1d87b02c8478dd3ed0ecc3eee47dc4d13282" -dependencies = [ - "tagger", -] +checksum = "439ee305def115ba05938db6eb1644ff94165c5ab5e9420d1c1bcedbba909391" [[package]] name = "ppv-lite86" @@ -957,43 +729,19 @@ checksum = "5b40af805b3121feab8a3c29f04d8ad262fa8e0561883e7653e024ae4479e6de" [[package]] -name = "proc-macro-error" -version = "1.0.4" +name = "proc-macro2" +version = "1.0.92" source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "da25490ff9892aab3fcf7c36f08cfb902dd3e71ca0f9f9517bea02a73a5ce38c" -dependencies = [ - "proc-macro-error-attr", - "proc-macro2", - "quote", - "syn", - "version_check", -] - -[[package]] -name = "proc-macro-error-attr" -version = "1.0.4" -source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "a1be40180e52ecc98ad80b184934baf3d0d29f979574e439af5a55274b35f869" -dependencies = [ - "proc-macro2", - "quote", - "version_check", -] - -[[package]] -name = "proc-macro2" -version = "1.0.47" -source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "5ea3d908b0e36316caf9e9e2c4625cdde190a7e6f440d794667ed17a1855e725" +checksum = "37d3544b3f2748c54e147655edb5025752e2303145b5aefb3c3ea2c78b973bb0" dependencies = [ "unicode-ident", ] [[package]] name = "quote" -version = "1.0.21" +version = "1.0.38" source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "bbe448f377a7d6961e30f5955f9b8d106c3f5e449d493ee1b125c1d43c2b5179" +checksum = "0e4dccaaaf89514f546c693ddc140f729f958c247918a13380cccc6078391acc" dependencies = [ "proc-macro2", ] @@ -1046,32 +794,41 @@ [[package]] name = "rayon" -version = "1.6.0" +version = "1.10.0" source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "1e060280438193c554f654141c9ea9417886713b7acd75974c85b18a69a88e0b" +checksum = "b418a60154510ca1a002a752ca9714984e21e4241e804d32555251faf8b78ffa" dependencies = [ - "crossbeam-deque", "either", "rayon-core", ] [[package]] name = "rayon-core" -version = "1.10.1" +version = "1.12.1" source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "cac410af5d00ab6884528b4ab69d1e8e146e8d471201800fa1b4524126de6ad3" +checksum = "1465873a3dfdaa8ae7cb14b4383657caab0b3e8a0aa9ae8e04b044854c8dfce2" dependencies = [ - "crossbeam-channel", "crossbeam-deque", "crossbeam-utils", - "num_cpus", ] [[package]] name = "regex" -version = "1.7.0" +version = "1.11.1" source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "e076559ef8e241f2ae3479e36f97bd5741c0330689e217ad51ce2c76808b868a" +checksum = "b544ef1b4eac5dc2db33ea63606ae9ffcfac26c1416a2806ae0bf5f56b201191" +dependencies = [ + "aho-corasick", + "memchr", + "regex-automata", + "regex-syntax", +] + +[[package]] +name = "regex-automata" +version = "0.4.9" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "809e8dc61f6de73b46c85f4c96486310fe304c434cfa43669d7b40f711150908" dependencies = [ "aho-corasick", "memchr", @@ -1079,123 +836,111 @@ ] [[package]] -name = "regex-automata" -version = "0.1.10" +name = "regex-syntax" +version = "0.8.5" source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "6c230d73fb8d8c1b9c0b3135c5142a8acee3a0558fb8db5cf1cb65f8d7862132" +checksum = "2b15c43186be67a4fd63bee50d0303afffcef381492ebe2c5d87f324e1b8815c" [[package]] -name = "regex-syntax" -version = "0.6.28" +name = "rustix" +version = "0.38.19" source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "456c603be3e8d448b072f410900c09faf164fbce2d480456f50eea6e25f9c848" +checksum = "745ecfa778e66b2b63c88a61cb36e0eea109e803b0b86bf9879fbc77c70e86ed" +dependencies = [ + "bitflags", + "errno", + "libc", + "linux-raw-sys", + "windows-sys 0.48.0", +] [[package]] -name = "rgb" -version = "0.8.34" +name = "ryu" +version = "1.0.15" source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "3603b7d71ca82644f79b5a06d1220e9a58ede60bd32255f698cb1af8838b8db3" +checksum = "1ad4cc8da4ef723ed60bced201181d83791ad433213d8c24efffda1eec85d741" + +[[package]] +name = "safe_arch" +version = "0.7.1" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "f398075ce1e6a179b46f51bd88d0598b92b00d3551f1a2d4ac49e771b56ac354" dependencies = [ "bytemuck", ] [[package]] -name = "rustix" -version = "0.35.13" -source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "727a1a6d65f786ec22df8a81ca3121107f235970dc1705ed681d3e6e8b9cd5f9" -dependencies = [ - "bitflags", - "errno", - "io-lifetimes 0.7.5", - "libc", - "linux-raw-sys 0.0.46", - "windows-sys", -] - -[[package]] -name = "rustix" -version = "0.36.3" -source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "0b1fbb4dfc4eb1d390c02df47760bb19a84bb80b301ecc947ab5406394d8223e" -dependencies = [ - "bitflags", - "errno", - "io-lifetimes 1.0.2", - "libc", - "linux-raw-sys 0.1.3", - "windows-sys", -] - -[[package]] -name = "ryu" -version = "1.0.11" +name = "serde" +version = "1.0.189" source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "4501abdff3ae82a1c1b477a17252eb69cee9e66eb915c1abaa4f44d873df9f09" - -[[package]] -name = "safe_arch" -version = "0.6.0" -source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "794821e4ccb0d9f979512f9c1973480123f9bd62a90d74ab0f9426fcf8f4a529" -dependencies = [ - "bytemuck", -] - -[[package]] -name = "scoped_threadpool" -version = "0.1.9" -source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "1d51f5df5af43ab3f1360b429fa5e0152ac5ce8c0bd6485cae490332e96846a8" - -[[package]] -name = "scopeguard" -version = "1.1.0" -source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "d29ab0c6d3fc0ee92fe66e2d99f700eab17a8d57d1c1d3b748380fb20baa78cd" - -[[package]] -name = "scratch" -version = "1.0.2" -source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "9c8132065adcfd6e02db789d9285a0deb2f3fcb04002865ab67d5fb103533898" - -[[package]] -name = "serde" -version = "1.0.148" -source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "e53f64bb4ba0191d6d0676e1b141ca55047d83b74f5607e6d8eb88126c52c2dc" +checksum = "8e422a44e74ad4001bdc8eede9a4570ab52f71190e9c076d14369f38b9200537" dependencies = [ "serde_derive", ] [[package]] name = "serde_derive" -version = "1.0.148" +version = "1.0.189" source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "a55492425aa53521babf6137309e7d34c20bbfbbfcfe2c7f3a047fd1f6b92c0c" +checksum = "1e48d1f918009ce3145511378cf68d613e3b3d9137d67272562080d68a2b32d5" dependencies = [ "proc-macro2", "quote", - "syn", + "syn 2.0.93", ] [[package]] name = "serde_json" -version = "1.0.89" +version = "1.0.107" source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "020ff22c755c2ed3f8cf162dbb41a7268d934702f3ed3631656ea597e08fc3db" +checksum = "6b420ce6e3d8bd882e9b243c6eed35dbc9a6110c9769e74b584e0d68d1f20c65" dependencies = [ - "itoa 1.0.4", + "itoa", "ryu", "serde", ] [[package]] +name = "serde_with" +version = "3.12.0" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "d6b6f7f2fcb69f747921f79f3926bd1e203fce4fef62c268dd3abfb6d86029aa" +dependencies = [ + "base64", + "chrono", + "hex", + "indexmap 1.9.3", + "indexmap 2.7.0", + "serde", + "serde_derive", + "serde_json", + "serde_with_macros", + "time", +] + +[[package]] +name = "serde_with_macros" +version = "3.12.0" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "8d00caa5193a3c8362ac2b73be6b9e768aa5a4b2f721d8f4b339600c3cb51f8e" +dependencies = [ + "darling", + "proc-macro2", + "quote", + "syn 2.0.93", +] + +[[package]] +name = "shlex" +version = "1.3.0" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "0fda2ff0d084019ba4d7c6f371c95d8fd75ce3524c3cb8fb653a3023f6323e64" + +[[package]] name = "simba" -version = "0.7.3" +version = "0.9.0" source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "2f3fd720c48c53cace224ae62bef1bbff363a70c68c4802a78b5cc6159618176" +checksum = "b3a386a501cd104797982c15ae17aafe8b9261315b5d07e3ec803f2ea26be0fa" dependencies = [ "approx", "num-complex", @@ -1205,31 +950,27 @@ ] [[package]] -name = "smallvec" -version = "1.10.0" +name = "strsim" +version = "0.11.1" source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "a507befe795404456341dfab10cef66ead4c041f62b8b11bbb92bffe5d0953e0" +checksum = "7da8b5736845d9f2fcb837ea5d9e2628564b3b043a70948a3f0b778838c5fb4f" [[package]] -name = "spin" -version = "0.9.4" +name = "syn" +version = "1.0.109" source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "7f6002a767bff9e83f8eeecf883ecb8011875a21ae8da43bffb817a57e78cc09" +checksum = "72b64191b275b66ffe2469e8af2c1cfe3bafa67b529ead792a6d0160888b4237" dependencies = [ - "lock_api", + "proc-macro2", + "quote", + "unicode-ident", ] [[package]] -name = "strsim" -version = "0.10.0" +name = "syn" +version = "2.0.93" source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "73473c0e59e6d5812c5dfe2a064a6444949f089e20eec9a2e5506596494e4623" - -[[package]] -name = "syn" -version = "1.0.104" -source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "4ae548ec36cf198c0ef7710d3c230987c2d6d7bd98ad6edc0274462724c585ce" +checksum = "9c786062daee0d6db1132800e623df74274a0a87322d8e183338e01b3d98d058" dependencies = [ "proc-macro2", "quote", @@ -1237,98 +978,78 @@ ] [[package]] -name = "tagger" -version = "4.3.4" -source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "6aaa6f5d645d1dae4cd0286e9f8bf15b75a31656348e5e106eb1a940abd34b63" - -[[package]] -name = "termcolor" -version = "1.1.3" +name = "terminal_size" +version = "0.4.1" source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "bab24d30b911b2376f3a13cc2cd443142f0c81dda04c118693e35b3835757755" -dependencies = [ - "winapi-util", -] - -[[package]] -name = "terminal_size" -version = "0.2.2" -source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "40ca90c434fd12083d1a6bdcbe9f92a14f96c8a1ba600ba451734ac334521f7a" +checksum = "5352447f921fda68cf61b4101566c0bdb5104eff6804d0678e5227580ab6a4e9" dependencies = [ - "rustix 0.35.13", - "windows-sys", -] - -[[package]] -name = "threadpool" -version = "1.8.1" -source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "d050e60b33d41c19108b32cea32164033a9013fe3b46cbd4457559bfbf77afaa" -dependencies = [ - "num_cpus", -] - -[[package]] -name = "tiff" -version = "0.8.0" -source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "f17def29300a156c19ae30814710d9c63cd50288a49c6fd3a10ccfbe4cf886fd" -dependencies = [ - "flate2", - "jpeg-decoder", - "weezl", + "rustix", + "windows-sys 0.59.0", ] [[package]] name = "time" -version = "0.1.45" +version = "0.3.37" source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "1b797afad3f312d1c66a56d11d0316f916356d11bd158fbc6ca6389ff6bf805a" +checksum = "35e7868883861bd0e56d9ac6efcaaca0d6d5d82a2a7ec8209ff492c07cf37b21" dependencies = [ - "libc", - "wasi 0.10.0+wasi-snapshot-preview1", - "winapi", + "deranged", + "itoa", + "num-conv", + "powerfmt", + "serde", + "time-core", + "time-macros", ] [[package]] -name = "trait-set" -version = "0.2.0" +name = "time-core" +version = "0.1.2" source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "875c4c873cc824e362fa9a9419ffa59807244824275a44ad06fec9684fff08f2" +checksum = "ef927ca75afb808a4d64dd374f00a2adf8d0fcff8e7b184af886c3c87ec4a3f3" + +[[package]] +name = "time-macros" +version = "0.2.19" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "2834e6017e3e5e4b9834939793b282bc03b37a3336245fa820e35e233e2a85de" dependencies = [ - "proc-macro2", - "quote", - "syn", + "num-conv", + "time-core", ] [[package]] name = "typenum" -version = "1.15.0" +version = "1.17.0" source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "dcf81ac59edc17cc8697ff311e8f5ef2d99fcbd9817b34cec66f90b6c3dfd987" +checksum = "42ff0bf0c66b8238c6f3b578df37d0b7848e55df8577b3f74f92a69acceeb825" [[package]] name = "unicase" -version = "2.6.0" +version = "2.7.0" source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "50f37be617794602aabbeee0be4f259dc1778fabe05e2d67ee8f79326d5cb4f6" +checksum = "f7d2d4dafb69621809a81864c9c1b864479e1235c0dd4e199924b9742439ed89" dependencies = [ "version_check", ] [[package]] name = "unicode-ident" -version = "1.0.5" +version = "1.0.12" source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "6ceab39d59e4c9499d4e5a8ee0e2735b891bb7308ac83dfb4e80cad195c9f6f3" +checksum = "3354b9ac3fae1ff6755cb6db53683adb661634f67557942dea4facebec0fee4b" [[package]] name = "unicode-width" -version = "0.1.10" +version = "0.2.0" source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "c0edd1e5b14653f783770bce4a4dabb4a5108a5370a5f5d8cfe8710c361f6c8b" +checksum = "1fc81956842c57dac11422a97c3b8195a1ff727f06e85c84ed2e8aa277c9a0fd" + +[[package]] +name = "utf8parse" +version = "0.2.2" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "06abde3611657adf66d383f00b093d7faecc7fa57071cce2578660c9f1010821" [[package]] name = "version_check" @@ -1338,46 +1059,40 @@ [[package]] name = "wasi" -version = "0.10.0+wasi-snapshot-preview1" -source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "1a143597ca7c7793eff794def352d41792a93c481eb1042423ff7ff72ba2c31f" - -[[package]] -name = "wasi" version = "0.11.0+wasi-snapshot-preview1" source = "registry+https://github.com/rust-lang/crates.io-index" checksum = "9c8d87e72b64a3b4db28d11ce29237c246188f4f51057d65a7eab63b7987e423" [[package]] name = "wasm-bindgen" -version = "0.2.83" +version = "0.2.99" source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "eaf9f5aceeec8be17c128b2e93e031fb8a4d469bb9c4ae2d7dc1888b26887268" +checksum = "a474f6281d1d70c17ae7aa6a613c87fce69a127e2624002df63dcb39d6cf6396" dependencies = [ "cfg-if", + "once_cell", "wasm-bindgen-macro", ] [[package]] name = "wasm-bindgen-backend" -version = "0.2.83" +version = "0.2.99" source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "4c8ffb332579b0557b52d268b91feab8df3615f265d5270fec2a8c95b17c1142" +checksum = "5f89bb38646b4f81674e8f5c3fb81b562be1fd936d84320f3264486418519c79" dependencies = [ "bumpalo", "log", - "once_cell", "proc-macro2", "quote", - "syn", + "syn 2.0.93", "wasm-bindgen-shared", ] [[package]] name = "wasm-bindgen-macro" -version = "0.2.83" +version = "0.2.99" source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "052be0f94026e6cbc75cdefc9bae13fd6052cdcaf532fa6c45e7ae33a1e6c810" +checksum = "2cc6181fd9a7492eef6fef1f33961e3695e4579b9872a6f7c83aee556666d4fe" dependencies = [ "quote", "wasm-bindgen-macro-support", @@ -1385,34 +1100,28 @@ [[package]] name = "wasm-bindgen-macro-support" -version = "0.2.83" +version = "0.2.99" source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "07bc0c051dc5f23e307b13285f9d75df86bfdf816c5721e573dec1f9b8aa193c" +checksum = "30d7a95b763d3c45903ed6c81f156801839e5ee968bb07e534c44df0fcd330c2" dependencies = [ "proc-macro2", "quote", - "syn", + "syn 2.0.93", "wasm-bindgen-backend", "wasm-bindgen-shared", ] [[package]] name = "wasm-bindgen-shared" -version = "0.2.83" +version = "0.2.99" source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "1c38c045535d93ec4f0b4defec448e4291638ee608530863b1e2ba115d4fff7f" - -[[package]] -name = "weezl" -version = "0.1.7" -source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "9193164d4de03a926d909d3bc7c30543cecb35400c02114792c2cae20d5e2dbb" +checksum = "943aab3fdaaa029a6e0271b35ea10b72b943135afe9bffca82384098ad0e06a6" [[package]] name = "wide" -version = "0.7.5" +version = "0.7.12" source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "ae41ecad2489a1655c8ef8489444b0b113c0a0c795944a3572a0931cf7d2525c" +checksum = "ebecebefc38ff1860b4bc47550bbfa63af5746061cf0d29fcd7fa63171602598" dependencies = [ "bytemuck", "safe_arch", @@ -1435,73 +1144,155 @@ checksum = "ac3b87c63620426dd9b991e5ce0329eff545bccbbb34f3be09ff6fb6ab51b7b6" [[package]] -name = "winapi-util" -version = "0.1.5" -source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "70ec6ce85bb158151cae5e5c87f95a8e97d2c0c4b001223f33a334e3ce5de178" -dependencies = [ - "winapi", -] - -[[package]] name = "winapi-x86_64-pc-windows-gnu" version = "0.4.0" source = "registry+https://github.com/rust-lang/crates.io-index" checksum = "712e227841d057c1ee1cd2fb22fa7e5a5461ae8e48fa2ca79ec42cfc1931183f" [[package]] +name = "windows-core" +version = "0.52.0" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "33ab640c8d7e35bf8ba19b884ba838ceb4fba93a4e8c65a9059d08afcfc683d9" +dependencies = [ + "windows-targets 0.52.6", +] + +[[package]] name = "windows-sys" -version = "0.42.0" +version = "0.48.0" source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "5a3e1820f08b8513f676f7ab6c1f99ff312fb97b553d30ff4dd86f9f15728aa7" +checksum = "677d2418bec65e3338edb076e806bc1ec15693c5d0104683f2efe857f61056a9" +dependencies = [ + "windows-targets 0.48.5", +] + +[[package]] +name = "windows-sys" +version = "0.59.0" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "1e38bc4d79ed67fd075bcc251a1c39b32a1776bbe92e5bef1f0bf1f8c531853b" dependencies = [ - "windows_aarch64_gnullvm", - "windows_aarch64_msvc", - "windows_i686_gnu", - "windows_i686_msvc", - "windows_x86_64_gnu", - "windows_x86_64_gnullvm", - "windows_x86_64_msvc", + "windows-targets 0.52.6", +] + +[[package]] +name = "windows-targets" +version = "0.48.5" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "9a2fa6e2155d7247be68c096456083145c183cbbbc2764150dda45a87197940c" +dependencies = [ + "windows_aarch64_gnullvm 0.48.5", + "windows_aarch64_msvc 0.48.5", + "windows_i686_gnu 0.48.5", + "windows_i686_msvc 0.48.5", + "windows_x86_64_gnu 0.48.5", + "windows_x86_64_gnullvm 0.48.5", + "windows_x86_64_msvc 0.48.5", +] + +[[package]] +name = "windows-targets" +version = "0.52.6" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "9b724f72796e036ab90c1021d4780d4d3d648aca59e491e6b98e725b84e99973" +dependencies = [ + "windows_aarch64_gnullvm 0.52.6", + "windows_aarch64_msvc 0.52.6", + "windows_i686_gnu 0.52.6", + "windows_i686_gnullvm", + "windows_i686_msvc 0.52.6", + "windows_x86_64_gnu 0.52.6", + "windows_x86_64_gnullvm 0.52.6", + "windows_x86_64_msvc 0.52.6", ] [[package]] name = "windows_aarch64_gnullvm" -version = "0.42.0" +version = "0.48.5" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "2b38e32f0abccf9987a4e3079dfb67dcd799fb61361e53e2882c3cbaf0d905d8" + +[[package]] +name = "windows_aarch64_gnullvm" +version = "0.52.6" source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "41d2aa71f6f0cbe00ae5167d90ef3cfe66527d6f613ca78ac8024c3ccab9a19e" +checksum = "32a4622180e7a0ec044bb555404c800bc9fd9ec262ec147edd5989ccd0c02cd3" + +[[package]] +name = "windows_aarch64_msvc" +version = "0.48.5" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "dc35310971f3b2dbbf3f0690a219f40e2d9afcf64f9ab7cc1be722937c26b4bc" [[package]] name = "windows_aarch64_msvc" -version = "0.42.0" +version = "0.52.6" source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "dd0f252f5a35cac83d6311b2e795981f5ee6e67eb1f9a7f64eb4500fbc4dcdb4" +checksum = "09ec2a7bb152e2252b53fa7803150007879548bc709c039df7627cabbd05d469" [[package]] name = "windows_i686_gnu" -version = "0.42.0" +version = "0.48.5" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "a75915e7def60c94dcef72200b9a8e58e5091744960da64ec734a6c6e9b3743e" + +[[package]] +name = "windows_i686_gnu" +version = "0.52.6" source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "fbeae19f6716841636c28d695375df17562ca208b2b7d0dc47635a50ae6c5de7" +checksum = "8e9b5ad5ab802e97eb8e295ac6720e509ee4c243f69d781394014ebfe8bbfa0b" + +[[package]] +name = "windows_i686_gnullvm" +version = "0.52.6" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "0eee52d38c090b3caa76c563b86c3a4bd71ef1a819287c19d586d7334ae8ed66" [[package]] name = "windows_i686_msvc" -version = "0.42.0" +version = "0.48.5" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "8f55c233f70c4b27f66c523580f78f1004e8b5a8b659e05a4eb49d4166cca406" + +[[package]] +name = "windows_i686_msvc" +version = "0.52.6" source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "84c12f65daa39dd2babe6e442988fc329d6243fdce47d7d2d155b8d874862246" +checksum = "240948bc05c5e7c6dabba28bf89d89ffce3e303022809e73deaefe4f6ec56c66" + +[[package]] +name = "windows_x86_64_gnu" +version = "0.48.5" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "53d40abd2583d23e4718fddf1ebec84dbff8381c07cae67ff7768bbf19c6718e" [[package]] name = "windows_x86_64_gnu" -version = "0.42.0" +version = "0.52.6" source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "bf7b1b21b5362cbc318f686150e5bcea75ecedc74dd157d874d754a2ca44b0ed" +checksum = "147a5c80aabfbf0c7d901cb5895d1de30ef2907eb21fbbab29ca94c5b08b1a78" + +[[package]] +name = "windows_x86_64_gnullvm" +version = "0.48.5" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "0b7b52767868a23d5bab768e390dc5f5c55825b6d30b86c844ff2dc7414044cc" [[package]] name = "windows_x86_64_gnullvm" -version = "0.42.0" +version = "0.52.6" source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "09d525d2ba30eeb3297665bd434a54297e4170c7f1a44cad4ef58095b4cd2028" +checksum = "24d5b23dc417412679681396f2b49f3de8c1473deb516bd34410872eff51ed0d" [[package]] name = "windows_x86_64_msvc" -version = "0.42.0" +version = "0.48.5" source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "f40009d85759725a34da6d89a94e63d7bdc50a862acf0dbc7c8e488f1edcb6f5" +checksum = "ed94fce61571a4006852b7389a063ab983c02eb1bb37b47f8272ce92d06d9538" + +[[package]] +name = "windows_x86_64_msvc" +version = "0.52.6" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "589f6da84c646204747d1270a2a5661ea66ed1cced2631d546fdfb155959f9ec"
--- a/Cargo.toml Tue Aug 01 10:25:09 2023 +0300 +++ b/Cargo.toml Mon Feb 17 13:54:53 2025 -0500 @@ -1,8 +1,8 @@ [package] name = "pointsource_algs" -version = "1.0.1" +version = "2.0.0-pre" edition = "2021" -rust-version = "1.67" +rust-version = "1.85" authors = ["Tuomo Valkonen <tuomov@iki.fi>"] description = "Algorithms for point source localisation" homepage = "https://tuomov.iki.fi/software/pointsource_algs/" @@ -21,29 +21,32 @@ ] categories = ["mathematics", "science", "computer-vision"] +[dependencies.alg_tools] +version = "~0.3.0-dev" +path = "../alg_tools" +default-features = false +features = ["nightly"] + [dependencies] -alg_tools = { version = "~0.1.0", path = "../alg_tools", default-features = false } serde = { version = "1.0", features = ["derive"] } num-traits = { version = "~0.2.14", features = ["std"] } rand = "~0.8.5" -colored = "~2.0.0" +colored = "~2.1.0" rand_distr = "~0.4.3" -nalgebra = { version = "~0.31.0", features = ["rand-no-std"] } -itertools = "~0.10.3" +nalgebra = { version = "~0.33.0", features = ["rand-no-std"] } +itertools = "~0.13.0" numeric_literals = "~0.2.0" -poloto = "~3.13.1" -GSL = "~6.0.0" +GSL = "~7.0.0" float_extras = "~0.1.6" -clap = { version = "~4.0.27", features = ["derive", "unicode", "wrap_help"] } -image = "~0.24.3" +clap = { version = "~4.5.0", features = ["derive", "unicode", "wrap_help"] } cpu-time = "~1.0.0" -colorbrewer = "~0.2.0" -rgb = "~0.8.33" serde_json = "~1.0.85" chrono = { version = "~0.4.23", features = ["alloc", "std", "serde"] } +anyhow = "1.0.95" +serde_with = { version = "3.11.0", features = ["macros"] } [build-dependencies] -regex = "~1.7.0" +regex = "~1.11.0" [profile.release] debug = true
--- a/README.md Tue Aug 01 10:25:09 2023 +0300 +++ b/README.md Mon Feb 17 13:54:53 2025 -0500 @@ -1,19 +1,19 @@ # Proximal methods for point source localisation: the implementation -This package contains [Rust] codes for the manuscript “_Proximal methods for -point source localisation_” ([arXiv:2212.02991]) by Tuomo Valkonen -⟨tuomov@iki.fi⟩. It concerns solution of problems of the type +This package contains the [Rust] codes for the numerical experiments in the articles +* T. Valkonen, “_Proximal methods for +point source localisation_”, Journal of Nonsmooth Analysis and Optimization 4 (2023), 10433, [doi:10.46298/jnsao-2023-10433] ([arXiv:2212.02991]) +* T. Valkonen, “_Point source localisation with unbalanced optimal transport_” (2025), submitted. + +It concerns solution of problems of the type $$ \min_{μ ∈ ℳ(Ω)}~ F(μ) + λ \|μ\|_{ℳ(Ω)} + δ_{≥ 0}(μ), $$ where $F$ is a data term, and $ℳ(Ω)$ is the space of Radon measures on the (rectangular) domain $Ω ⊂ ℝ^n$. Implemented are $F(μ)=\frac12\|Aμ-b\|_2^2$ and $F(μ)=\|Aμ-b\|_1$ for the forward operator $A \in 𝕃(ℳ(Ω); ℝ^m)$ modelling a -simple sensor grid. For the 2-norm-squared data term implemented are the -algorithms μFB, μFISTA, and μPDPS from the aforementioned manuscript along with -comparison relaxed and fully corrective conditional gradient methods from the -literature. For the 1-norm data term only the μPDPS is applicable. +simple sensor grid. ## Installation and usage @@ -36,8 +36,8 @@ brew install gsl ``` For other operating systems, suggestions are available in the [rust-GSL] - crate documentation. On Windows, you will likely need to pass extra - `RUSTFLAGS` options to Cargo in the following steps to locate the library. + crate documentation. You may need to pass extra `RUSTFLAGS` options to + Cargo in the following steps to locate the library. 4. Download [alg_tools] and unpack it under the same directory as this package. @@ -49,38 +49,33 @@ [rust-GSL]: https://docs.rs/GSL/6.0.0/rgsl/ [Homebrew]: https://brew.sh [arXiv:2212.02991]: https://arxiv.org/abs/2212.02991 + [doi:10.46298/jnsao-2023-10433]: http://doi.org/10.46298/jnsao-2023-10433 ### Building and running the experiments -To compile the code and run the experiments in the manuscript, use +To compile and install the program, use +```console +cargo install --path=. +``` +When doing this for the first time, several dependencies will be downloaded. +Now you can run the default set of experiments with +``` +pointsource_algs -o results +``` +The `-o results` option tells `pointsource_algs` to write results in the +`results` directory. The option is required. + +Alternatively, you may build and run the program without installing with ```console cargo run --release -- -o results ``` -When doing this for the first time, several dependencies will be downloaded. -The double-dash (`--`) separates the arguments of Cargo and this software, -`pointsource_algs`. The `--release` option to Cargo is required for `rustc` to -build optimised high performance code. Without that flag the performance will -be significantly worse. The `-o results` option tells `pointsource_algs` to -write results in the `results` directory. The option is required. - -Alternatively, you may build the executable with -```console -cargo build --release -``` -and then run it with -``` -target/release/pointsource_algs -o results -``` +The double-dash separates the options for the Cargo build system +and `pointsource_algs`. ### Documentation Use the `--help` option to get an extensive listing of command line options to -customise algorithm parameters and the experiments performed. As above with -`-o`, if using `cargo` to run the executable, you have to pass any arguments -to `pointsource_algs` after a double-dash: -```console -cargo run --release -- --help -``` +customise algorithm parameters and the experiments performed. ## Internals
--- /dev/null Thu Jan 01 00:00:00 1970 +0000 +++ b/src/dataterm.rs Mon Feb 17 13:54:53 2025 -0500 @@ -0,0 +1,94 @@ +/*! +Basid definitions for data terms +*/ + +use numeric_literals::replace_float_literals; + +use alg_tools::euclidean::Euclidean; +use alg_tools::linops::GEMV; +pub use alg_tools::norms::L1; +use alg_tools::norms::Norm; +use alg_tools::instance::{Instance, Space}; + +use crate::types::*; +pub use crate::types::L2Squared; +use crate::measures::RNDM; + +/// Calculates the residual $Aμ-b$. +#[replace_float_literals(F::cast_from(literal))] +pub(crate) fn calculate_residual< + X : Space, + I : Instance<X>, + F : Float, + V : Euclidean<F> + Clone, + A : GEMV<F, X, Codomain = V>, +>( + μ : I, + opA : &A, + b : &V +) -> V { + let mut r = b.clone(); + opA.gemv(&mut r, 1.0, μ, -1.0); + r +} + +/// Calculates the residual $A(μ+μ_delta)-b$. +#[replace_float_literals(F::cast_from(literal))] +pub(crate) fn calculate_residual2< + F : Float, + X : Space, + I : Instance<X>, + J : Instance<X>, + V : Euclidean<F> + Clone, + A : GEMV<F, X, Codomain = V>, +>( + μ : I, + μ_delta : J, + opA : &A, + b : &V +) -> V { + let mut r = b.clone(); + opA.gemv(&mut r, 1.0, μ, -1.0); + opA.gemv(&mut r, 1.0, μ_delta, 1.0); + r +} + + +/// Trait for data terms +#[replace_float_literals(F::cast_from(literal))] +pub trait DataTerm<F : Float, V, const N : usize> { + /// Calculates $F(y)$, where $F$ is the data fidelity. + fn calculate_fit(&self, _residual : &V) -> F; + + /// Calculates $F(Aμ-b)$, where $F$ is the data fidelity. + fn calculate_fit_op<I, A : GEMV<F, RNDM<F, N>, Codomain = V>>( + &self, + μ : I, + opA : &A, + b : &V + ) -> F + where + V : Euclidean<F> + Clone, + I : Instance<RNDM<F, N>>, + { + let r = calculate_residual(μ, opA, b); + self.calculate_fit(&r) + } +} + +impl<F : Float, V : Euclidean<F>, const N : usize> +DataTerm<F, V, N> +for L2Squared { + fn calculate_fit(&self, residual : &V) -> F { + residual.norm2_squared_div2() + } +} + + +impl<F : Float, V : Euclidean<F> + Norm<F, L1>, const N : usize> +DataTerm<F, V, N> +for L1 { + fn calculate_fit(&self, residual : &V) -> F { + residual.norm(L1) + } +}
--- a/src/experiments.rs Tue Aug 01 10:25:09 2023 +0300 +++ b/src/experiments.rs Mon Feb 17 13:54:53 2025 -0500 @@ -13,21 +13,24 @@ use alg_tools::error::DynResult; use alg_tools::norms::Linfinity; -use crate::ExperimentOverrides; +use crate::{ExperimentOverrides, AlgorithmOverrides}; use crate::kernels::*; -use crate::kernels::{SupportProductFirst as Prod}; -use crate::pdps::PDPSConfig; +use crate::kernels::SupportProductFirst as Prod; use crate::types::*; use crate::run::{ RunnableExperiment, ExperimentV2, + ExperimentBiased, Named, DefaultAlgorithm, - AlgorithmConfig }; //use crate::fb::FBGenericConfig; use crate::rand_distr::{SerializableNormal, SaltAndPepper}; use crate::regularisation::Regularisation; +use alg_tools::euclidean::Euclidean; +use alg_tools::instance::Instance; +use alg_tools::mapping::Mapping; +use alg_tools::operator_arithmetic::{MappingSum, Weighted}; /// Experiments shorthands, to be used with the command line parser @@ -58,6 +61,12 @@ /// Two dimensions, “fast” spread, 1-norm data fidelity #[clap(name = "2d_l1_fast")] Experiment2D_L1_Fast, + /// One dimension, “fast” spread, 2-norm-squared data fidelity with extra TV-regularised bias + #[clap(name = "1d_tv_fast")] + Experiment1D_TV_Fast, + /// Two dimensions, “fast” spread, 2-norm-squared data fidelity with extra TV-regularised bias + #[clap(name = "2d_tv_fast")] + Experiment2D_TV_Fast, } macro_rules! make_float_constant { @@ -92,6 +101,25 @@ ([0.30, 0.70], 5.0) ]; +/// The $\{0,1\}$-valued characteristic function of a ball as a [`Mapping`]. +#[derive(Debug,Copy,Clone,Serialize,PartialEq)] +struct BallCharacteristic<F : Float, const N : usize> { + pub center : Loc<F, N>, + pub radius : F, +} + +impl<F : Float, const N : usize> Mapping<Loc<F, N>> for BallCharacteristic<F, N> { + type Codomain =F; + + fn apply<I : Instance<Loc<F, N>>>(&self, i : I) -> F { + if self.center.dist2(i) <= self.radius { + F::ONE + } else { + F::ZERO + } + } +} + //#[replace_float_literals(F::cast_from(literal))] impl DefaultExperiment { /// Convert the experiment shorthand into a runnable experiment configuration. @@ -115,23 +143,70 @@ make_float_constant!(Variance1 = 0.05.powi(2)); make_float_constant!(CutOff1 = 0.15); make_float_constant!(Hat1 = 0.16); + make_float_constant!(HatBias = 0.05); // We use a different step length for PDPS in 2D experiments - let pdps_2d = || { - let τ0 = 3.0; - PDPSConfig { - τ0, - σ0 : 0.99 / τ0, + // let pdps_2d = (DefaultAlgorithm::PDPS, + // AlgorithmOverrides { + // tau0 : Some(3.0), + // sigma0 : Some(0.99 / 3.0), + // .. Default::default() + // } + // ); + // let radon_pdps_2d = (DefaultAlgorithm::RadonPDPS, + // AlgorithmOverrides { + // tau0 : Some(3.0), + // sigma0 : Some(0.99 / 3.0), + // .. Default::default() + // } + // ); + let sliding_fb_cut_gaussian = (DefaultAlgorithm::SlidingFB, + AlgorithmOverrides { + theta0 : Some(0.3), .. Default::default() } - }; - + ); + // let higher_cpos = |alg| (alg, + // AlgorithmOverrides { + // transport_tolerance_pos : Some(1000.0), + // .. Default::default() + // } + // ); + let higher_cpos_merging = |alg| (alg, + AlgorithmOverrides { + transport_tolerance_pos : Some(1000.0), + merge : Some(true), + fitness_merging : Some(true), + .. Default::default() + } + ); + let higher_cpos_merging_steptune = |alg| (alg, + AlgorithmOverrides { + transport_tolerance_pos : Some(1000.0), + theta0 : Some(0.3), + merge : Some(true), + fitness_merging : Some(true), + .. Default::default() + } + ); + let much_higher_cpos_merging_steptune = |alg| (alg, + AlgorithmOverrides { + transport_tolerance_pos : Some(10000.0), + sigma0 : Some(0.15), + theta0 : Some(0.3), + merge : Some(true), + fitness_merging : Some(true), + .. Default::default() + } + ); // We add a hash of the experiment name to the configured // noise seed to not use the same noise for different experiments. let mut h = DefaultHasher::new(); name.hash(&mut h); let noise_seed = cli.noise_seed.unwrap_or(BASE_SEED) + h.finish(); + let default_merge_radius = 0.01; + use DefaultExperiment::*; Ok(match self { Experiment1D => { @@ -140,7 +215,7 @@ Box::new(Named { name, data : ExperimentV2 { domain : [[0.0, 1.0]].into(), sensor_count : [N_SENSORS_1D], - regularisation : Regularisation::NonnegRadon(cli.alpha.unwrap_or(0.09)), + regularisation : Regularisation::NonnegRadon(cli.alpha.unwrap_or(0.08)), noise_distr : SerializableNormal::new(0.0, cli.variance.unwrap_or(0.2))?, dataterm : DataTerm::L2Squared, μ_hat : MU_TRUE_1D_BASIC.into(), @@ -149,7 +224,12 @@ kernel : Prod(AutoConvolution(spread_cutoff), base_spread), kernel_plot_width, noise_seed, - algorithm_defaults: HashMap::new(), + default_merge_radius, + algorithm_overrides: HashMap::from([ + sliding_fb_cut_gaussian, + higher_cpos_merging(DefaultAlgorithm::RadonFB), + higher_cpos_merging(DefaultAlgorithm::RadonSlidingFB), + ]), }}) }, Experiment1DFast => { @@ -166,7 +246,11 @@ kernel : base_spread, kernel_plot_width, noise_seed, - algorithm_defaults: HashMap::new(), + default_merge_radius, + algorithm_overrides: HashMap::from([ + higher_cpos_merging(DefaultAlgorithm::RadonFB), + higher_cpos_merging(DefaultAlgorithm::RadonSlidingFB), + ]), }}) }, Experiment2D => { @@ -184,8 +268,11 @@ kernel : Prod(AutoConvolution(spread_cutoff), base_spread), kernel_plot_width, noise_seed, - algorithm_defaults: HashMap::from([ - (DefaultAlgorithm::PDPS, AlgorithmConfig::PDPS(pdps_2d())) + default_merge_radius, + algorithm_overrides: HashMap::from([ + sliding_fb_cut_gaussian, + higher_cpos_merging(DefaultAlgorithm::RadonFB), + higher_cpos_merging(DefaultAlgorithm::RadonSlidingFB), ]), }}) }, @@ -203,8 +290,10 @@ kernel : base_spread, kernel_plot_width, noise_seed, - algorithm_defaults: HashMap::from([ - (DefaultAlgorithm::PDPS, AlgorithmConfig::PDPS(pdps_2d())) + default_merge_radius, + algorithm_overrides: HashMap::from([ + higher_cpos_merging(DefaultAlgorithm::RadonFB), + higher_cpos_merging(DefaultAlgorithm::RadonSlidingFB), ]), }}) }, @@ -226,7 +315,8 @@ kernel : Prod(AutoConvolution(spread_cutoff), base_spread), kernel_plot_width, noise_seed, - algorithm_defaults: HashMap::new(), + default_merge_radius, + algorithm_overrides: HashMap::new(), }}) }, Experiment1D_L1_Fast => { @@ -246,7 +336,8 @@ kernel : base_spread, kernel_plot_width, noise_seed, - algorithm_defaults: HashMap::new(), + default_merge_radius, + algorithm_overrides: HashMap::new(), }}) }, Experiment2D_L1 => { @@ -267,8 +358,8 @@ kernel : Prod(AutoConvolution(spread_cutoff), base_spread), kernel_plot_width, noise_seed, - algorithm_defaults: HashMap::from([ - (DefaultAlgorithm::PDPS, AlgorithmConfig::PDPS(pdps_2d())) + default_merge_radius, + algorithm_overrides: HashMap::from([ ]), }}) }, @@ -289,9 +380,65 @@ kernel : base_spread, kernel_plot_width, noise_seed, - algorithm_defaults: HashMap::from([ - (DefaultAlgorithm::PDPS, AlgorithmConfig::PDPS(pdps_2d())) + default_merge_radius, + algorithm_overrides: HashMap::from([ + ]), + }}) + }, + Experiment1D_TV_Fast => { + let base_spread = HatConv { radius : HatBias }; + Box::new(Named { name, data : ExperimentBiased { + λ : 0.02, + bias : MappingSum::new([ + Weighted::new(1.0, BallCharacteristic{ center : 0.3.into(), radius : 0.2 }), + Weighted::new(0.5, BallCharacteristic{ center : 0.6.into(), radius : 0.3 }), ]), + base : ExperimentV2 { + domain : [[0.0, 1.0]].into(), + sensor_count : [N_SENSORS_1D], + regularisation : Regularisation::NonnegRadon(cli.alpha.unwrap_or(0.2)), + noise_distr : SerializableNormal::new(0.0, cli.variance.unwrap_or(0.1))?, + dataterm : DataTerm::L2Squared, + μ_hat : MU_TRUE_1D_BASIC.into(), + sensor : BallIndicator { r : SensorWidth1D, exponent : Linfinity }, + spread : base_spread, + kernel : base_spread, + kernel_plot_width, + noise_seed, + default_merge_radius, + algorithm_overrides: HashMap::from([ + higher_cpos_merging_steptune(DefaultAlgorithm::RadonForwardPDPS), + higher_cpos_merging_steptune(DefaultAlgorithm::RadonSlidingPDPS), + ]), + }, + }}) + }, + Experiment2D_TV_Fast => { + let base_spread = HatConv { radius : Hat1 }; + Box::new(Named { name, data : ExperimentBiased { + λ : 0.005, + bias : MappingSum::new([ + Weighted::new(1.0, BallCharacteristic{ center : [0.3, 0.3].into(), radius : 0.2 }), + Weighted::new(0.5, BallCharacteristic{ center : [0.6, 0.6].into(), radius : 0.3 }), + ]), + base : ExperimentV2 { + domain : [[0.0, 1.0]; 2].into(), + sensor_count : [N_SENSORS_2D; 2], + regularisation : Regularisation::NonnegRadon(cli.alpha.unwrap_or(0.06)), + noise_distr : SerializableNormal::new(0.0, cli.variance.unwrap_or(0.15))?, //0.25 + dataterm : DataTerm::L2Squared, + μ_hat : MU_TRUE_2D_BASIC.into(), + sensor : BallIndicator { r : SensorWidth2D, exponent : Linfinity }, + spread : base_spread, + kernel : base_spread, + kernel_plot_width, + noise_seed, + default_merge_radius, + algorithm_overrides: HashMap::from([ + much_higher_cpos_merging_steptune(DefaultAlgorithm::RadonForwardPDPS), + much_higher_cpos_merging_steptune(DefaultAlgorithm::RadonSlidingPDPS), + ]), + }, }}) }, })
--- a/src/fb.rs Tue Aug 01 10:25:09 2023 +0300 +++ b/src/fb.rs Mon Feb 17 13:54:53 2025 -0500 @@ -6,10 +6,7 @@ * Valkonen T. - _Proximal methods for point source localisation_, [arXiv:2212.02991](https://arxiv.org/abs/2212.02991). -The main routine is [`pointsource_fb_reg`]. It is based on [`generic_pointsource_fb_reg`], which is -also used by our [primal-dual proximal splitting][crate::pdps] implementation. - -FISTA-type inertia can also be enabled through [`FBConfig::meta`]. +The main routine is [`pointsource_fb_reg`]. ## Problem @@ -76,650 +73,94 @@ $$ </p> -We solve this with either SSN or FB via [`quadratic_nonneg`] as determined by -[`InnerSettings`] in [`FBGenericConfig::inner`]. +We solve this with either SSN or FB as determined by +[`crate::subproblem::InnerSettings`] in [`FBGenericConfig::inner`]. */ +use colored::Colorize; use numeric_literals::replace_float_literals; -use serde::{Serialize, Deserialize}; -use colored::Colorize; -use nalgebra::{DVector, DMatrix}; +use serde::{Deserialize, Serialize}; -use alg_tools::iterate::{ - AlgIteratorFactory, - AlgIteratorState, -}; use alg_tools::euclidean::Euclidean; -use alg_tools::linops::Apply; -use alg_tools::sets::Cube; -use alg_tools::loc::Loc; -use alg_tools::mapping::Mapping; -use alg_tools::bisection_tree::{ - BTFN, - PreBTFN, - Bounds, - BTNodeLookup, - BTNode, - BTSearch, - P2Minimise, - SupportGenerator, - LocalAnalysis, - Bounded, -}; +use alg_tools::instance::Instance; +use alg_tools::iterate::AlgIteratorFactory; +use alg_tools::linops::{Mapping, GEMV}; use alg_tools::mapping::RealMapping; use alg_tools::nalgebra_support::ToNalgebraRealField; +use crate::dataterm::{calculate_residual, DataTerm, L2Squared}; +use crate::forward_model::{AdjointProductBoundedBy, ForwardModel}; +use crate::measures::merging::SpikeMerging; +use crate::measures::{DiscreteMeasure, RNDM}; +use crate::plot::{PlotLookup, Plotting, SeqPlotter}; +pub use crate::prox_penalty::{FBGenericConfig, ProxPenalty}; +use crate::regularisation::RegTerm; use crate::types::*; -use crate::measures::{ - DiscreteMeasure, - DeltaMeasure, -}; -use crate::measures::merging::{ - SpikeMergingMethod, - SpikeMerging, -}; -use crate::forward_model::ForwardModel; -use crate::seminorms::{ - DiscreteMeasureOp, Lipschitz -}; -use crate::subproblem::{ - nonneg::quadratic_nonneg, - unconstrained::quadratic_unconstrained, - InnerSettings, - InnerMethod, -}; -use crate::tolerance::Tolerance; -use crate::plot::{ - SeqPlotter, - Plotting, - PlotLookup -}; -use crate::regularisation::{ - NonnegRadonRegTerm, - RadonRegTerm, -}; - -/// Method for constructing $μ$ on each iteration -#[derive(Clone, Copy, Eq, PartialEq, Serialize, Deserialize, Debug)] -#[allow(dead_code)] -pub enum InsertionStyle { - /// Resuse previous $μ$ from previous iteration, optimising weights - /// before inserting new spikes. - Reuse, - /// Start each iteration with $μ=0$. - Zero, -} - -/// Meta-algorithm type -#[derive(Clone, Copy, Eq, PartialEq, Serialize, Deserialize, Debug)] -#[allow(dead_code)] -pub enum FBMetaAlgorithm { - /// No meta-algorithm - None, - /// FISTA-style inertia - InertiaFISTA, -} /// Settings for [`pointsource_fb_reg`]. #[derive(Clone, Copy, Eq, PartialEq, Serialize, Deserialize, Debug)] #[serde(default)] -pub struct FBConfig<F : Float> { +pub struct FBConfig<F: Float> { /// Step length scaling - pub τ0 : F, - /// Meta-algorithm to apply - pub meta : FBMetaAlgorithm, + pub τ0: F, /// Generic parameters - pub insertion : FBGenericConfig<F>, -} - -/// Settings for the solution of the stepwise optimality condition in algorithms based on -/// [`generic_pointsource_fb_reg`]. -#[derive(Clone, Copy, Eq, PartialEq, Serialize, Deserialize, Debug)] -#[serde(default)] -pub struct FBGenericConfig<F : Float> { - /// Method for constructing $μ$ on each iteration; see [`InsertionStyle`]. - pub insertion_style : InsertionStyle, - /// Tolerance for point insertion. - pub tolerance : Tolerance<F>, - /// Stop looking for predual maximum (where to isert a new point) below - /// `tolerance` multiplied by this factor. - pub insertion_cutoff_factor : F, - /// Settings for branch and bound refinement when looking for predual maxima - pub refinement : RefinementSettings<F>, - /// Maximum insertions within each outer iteration - pub max_insertions : usize, - /// Pair `(n, m)` for maximum insertions `m` on first `n` iterations. - pub bootstrap_insertions : Option<(usize, usize)>, - /// Inner method settings - pub inner : InnerSettings<F>, - /// Spike merging method - pub merging : SpikeMergingMethod<F>, - /// Tolerance multiplier for merges - pub merge_tolerance_mult : F, - /// Spike merging method after the last step - pub final_merging : SpikeMergingMethod<F>, - /// Iterations between merging heuristic tries - pub merge_every : usize, - /// Save $μ$ for postprocessing optimisation - pub postprocessing : bool + pub generic: FBGenericConfig<F>, } #[replace_float_literals(F::cast_from(literal))] -impl<F : Float> Default for FBConfig<F> { +impl<F: Float> Default for FBConfig<F> { fn default() -> Self { FBConfig { - τ0 : 0.99, - meta : FBMetaAlgorithm::None, - insertion : Default::default() - } - } -} - -#[replace_float_literals(F::cast_from(literal))] -impl<F : Float> Default for FBGenericConfig<F> { - fn default() -> Self { - FBGenericConfig { - insertion_style : InsertionStyle::Reuse, - tolerance : Default::default(), - insertion_cutoff_factor : 1.0, - refinement : Default::default(), - max_insertions : 100, - //bootstrap_insertions : None, - bootstrap_insertions : Some((10, 1)), - inner : InnerSettings { - method : InnerMethod::SSN, - .. Default::default() - }, - merging : SpikeMergingMethod::None, - //merging : Default::default(), - final_merging : Default::default(), - merge_every : 10, - merge_tolerance_mult : 2.0, - postprocessing : false, + τ0: 0.99, + generic: Default::default(), } } } -/// Trait for specialisation of [`generic_pointsource_fb_reg`] to basic FB, FISTA. -/// -/// The idea is that the residual $Aμ - b$ in the forward step can be replaced by an arbitrary -/// value. For example, to implement [primal-dual proximal splitting][crate::pdps] we replace it -/// with the dual variable $y$. We can then also implement alternative data terms, as the -/// (pre)differential of $F(μ)=F\_0(Aμ-b)$ is $F\'(μ) = A\_*F\_0\'(Aμ-b)$. In the case of the -/// quadratic fidelity $F_0(y)=\frac{1}{2}\\|y\\|_2^2$ in a Hilbert space, of course, -/// $F\_0\'(Aμ-b)=Aμ-b$ is the residual. -pub trait FBSpecialisation<F : Float, Observable : Euclidean<F>, const N : usize> : Sized { - /// Updates the residual and does any necessary pruning of `μ`. - /// - /// Returns the new residual and possibly a new step length. - /// - /// The measure `μ` may also be modified to apply, e.g., inertia to it. - /// The updated residual should correspond to the residual at `μ`. - /// See the [trait documentation][FBSpecialisation] for the use and meaning of the residual. - /// - /// The parameter `μ_base` is the base point of the iteration, typically the previous iterate, - /// but for, e.g., FISTA has inertia applied to it. - fn update( - &mut self, - μ : &mut DiscreteMeasure<Loc<F, N>, F>, - μ_base : &DiscreteMeasure<Loc<F, N>, F>, - ) -> (Observable, Option<F>); - - /// Calculates the data term value corresponding to iterate `μ` and available residual. - /// - /// Inertia and other modifications, as deemed, necessary, should be applied to `μ`. - /// - /// The blanket implementation correspondsn to the 2-norm-squared data fidelity - /// $\\|\text{residual}\\|\_2^2/2$. - fn calculate_fit( - &self, - _μ : &DiscreteMeasure<Loc<F, N>, F>, - residual : &Observable - ) -> F { - residual.norm2_squared_div2() - } - - /// Calculates the data term value at $μ$. - /// - /// Unlike [`Self::calculate_fit`], no inertia, etc., should be applied to `μ`. - fn calculate_fit_simple( - &self, - μ : &DiscreteMeasure<Loc<F, N>, F>, - ) -> F; - - /// Returns the final iterate after any necessary postprocess pruning, merging, etc. - fn postprocess(self, mut μ : DiscreteMeasure<Loc<F, N>, F>, merging : SpikeMergingMethod<F>) - -> DiscreteMeasure<Loc<F, N>, F> - where DiscreteMeasure<Loc<F, N>, F> : SpikeMerging<F> { - μ.merge_spikes_fitness(merging, - |μ̃| self.calculate_fit_simple(μ̃), - |&v| v); - μ.prune(); - μ - } - - /// Returns measure to be used for value calculations, which may differ from μ. - fn value_μ<'c, 'b : 'c>(&'b self, μ : &'c DiscreteMeasure<Loc<F, N>, F>) - -> &'c DiscreteMeasure<Loc<F, N>, F> { - μ - } -} - -/// Specialisation of [`generic_pointsource_fb_reg`] to basic μFB. -struct BasicFB< - 'a, - F : Float + ToNalgebraRealField, - A : ForwardModel<Loc<F, N>, F>, - const N : usize -> { - /// The data - b : &'a A::Observable, - /// The forward operator - opA : &'a A, -} - -/// Implementation of [`FBSpecialisation`] for basic μFB forward-backward splitting. -#[replace_float_literals(F::cast_from(literal))] -impl<'a, F : Float + ToNalgebraRealField , A : ForwardModel<Loc<F, N>, F>, const N : usize> -FBSpecialisation<F, A::Observable, N> for BasicFB<'a, F, A, N> { - fn update( - &mut self, - μ : &mut DiscreteMeasure<Loc<F, N>, F>, - _μ_base : &DiscreteMeasure<Loc<F, N>, F> - ) -> (A::Observable, Option<F>) { - μ.prune(); - //*residual = self.opA.apply(μ) - self.b; - let mut residual = self.b.clone(); - self.opA.gemv(&mut residual, 1.0, μ, -1.0); - (residual, None) - } - - fn calculate_fit_simple( - &self, - μ : &DiscreteMeasure<Loc<F, N>, F>, - ) -> F { - let mut residual = self.b.clone(); - self.opA.gemv(&mut residual, 1.0, μ, -1.0); - residual.norm2_squared_div2() - } -} - -/// Specialisation of [`generic_pointsource_fb_reg`] to FISTA. -struct FISTA< - 'a, - F : Float + ToNalgebraRealField, - A : ForwardModel<Loc<F, N>, F>, - const N : usize -> { - /// The data - b : &'a A::Observable, - /// The forward operator - opA : &'a A, - /// Current inertial parameter - λ : F, - /// Previous iterate without inertia applied. - /// We need to store this here because `μ_base` passed to [`FBSpecialisation::update`] will - /// have inertia applied to it, so is not useful to use. - μ_prev : DiscreteMeasure<Loc<F, N>, F>, -} - -/// Implementation of [`FBSpecialisation`] for μFISTA inertial forward-backward splitting. -#[replace_float_literals(F::cast_from(literal))] -impl<'a, F : Float + ToNalgebraRealField, A : ForwardModel<Loc<F, N>, F>, const N : usize> -FBSpecialisation<F, A::Observable, N> for FISTA<'a, F, A, N> { - fn update( - &mut self, - μ : &mut DiscreteMeasure<Loc<F, N>, F>, - _μ_base : &DiscreteMeasure<Loc<F, N>, F> - ) -> (A::Observable, Option<F>) { - // Update inertial parameters - let λ_prev = self.λ; - self.λ = 2.0 * λ_prev / ( λ_prev + (4.0 + λ_prev * λ_prev).sqrt() ); - let θ = self.λ / λ_prev - self.λ; - // Perform inertial update on μ. - // This computes μ ← (1 + θ) * μ - θ * μ_prev, pruning spikes where both μ - // and μ_prev have zero weight. Since both have weights from the finite-dimensional - // subproblem with a proximal projection step, this is likely to happen when the - // spike is not needed. A copy of the pruned μ without artithmetic performed is - // stored in μ_prev. - μ.pruning_sub(1.0 + θ, θ, &mut self.μ_prev); - - //*residual = self.opA.apply(μ) - self.b; - let mut residual = self.b.clone(); - self.opA.gemv(&mut residual, 1.0, μ, -1.0); - (residual, None) - } - - fn calculate_fit_simple( - &self, - μ : &DiscreteMeasure<Loc<F, N>, F>, - ) -> F { - let mut residual = self.b.clone(); - self.opA.gemv(&mut residual, 1.0, μ, -1.0); - residual.norm2_squared_div2() - } - - fn calculate_fit( - &self, - _μ : &DiscreteMeasure<Loc<F, N>, F>, - _residual : &A::Observable - ) -> F { - self.calculate_fit_simple(&self.μ_prev) - } - - // For FISTA we need to do a final pruning as well, due to the limited - // pruning that can be done on each step. - fn postprocess(mut self, μ_base : DiscreteMeasure<Loc<F, N>, F>, merging : SpikeMergingMethod<F>) - -> DiscreteMeasure<Loc<F, N>, F> - where DiscreteMeasure<Loc<F, N>, F> : SpikeMerging<F> { - let mut μ = self.μ_prev; - self.μ_prev = μ_base; - μ.merge_spikes_fitness(merging, - |μ̃| self.calculate_fit_simple(μ̃), - |&v| v); - μ.prune(); - μ - } - - fn value_μ<'c, 'b : 'c>(&'c self, _μ : &'c DiscreteMeasure<Loc<F, N>, F>) - -> &'c DiscreteMeasure<Loc<F, N>, F> { - &self.μ_prev - } -} - - -/// Abstraction of regularisation terms for [`generic_pointsource_fb_reg`]. -pub trait RegTerm<F : Float + ToNalgebraRealField, const N : usize> -: for<'a> Apply<&'a DiscreteMeasure<Loc<F, N>, F>, Output = F> { - /// Approximately solve the problem - /// <div>$$ - /// \min_{x ∈ ℝ^n} \frac{1}{2} x^⊤Ax - g^⊤ x + τ G(x) - /// $$</div> - /// for $G$ depending on the trait implementation. - /// - /// The parameter `mA` is $A$. An estimate for its opeator norm should be provided in - /// `mA_normest`. The initial iterate and output is `x`. The current main tolerance is `ε`. - /// - /// Returns the number of iterations taken. - fn solve_findim( - &self, - mA : &DMatrix<F::MixedType>, - g : &DVector<F::MixedType>, - τ : F, - x : &mut DVector<F::MixedType>, - mA_normest : F, - ε : F, - config : &FBGenericConfig<F> - ) -> usize; - - /// Find a point where `d` may violate the tolerance `ε`. - /// - /// If `skip_by_rough_check` is set, do not find the point if a rough check indicates that we - /// are in bounds. `ε` is the current main tolerance and `τ` a scaling factor for the - /// regulariser. - /// - /// Returns `None` if `d` is in bounds either based on the rough check, or a more precise check - /// terminating early. Otherwise returns a possibly violating point, the value of `d` there, - /// and a boolean indicating whether the found point is in bounds. - fn find_tolerance_violation<G, BT>( - &self, - d : &mut BTFN<F, G, BT, N>, - τ : F, - ε : F, - skip_by_rough_check : bool, - config : &FBGenericConfig<F>, - ) -> Option<(Loc<F, N>, F, bool)> - where BT : BTSearch<F, N, Agg=Bounds<F>>, - G : SupportGenerator<F, N, Id=BT::Data>, - G::SupportType : Mapping<Loc<F, N>,Codomain=F> - + LocalAnalysis<F, Bounds<F>, N>; - - /// Verify that `d` is in bounds `ε` for a merge candidate `μ` - /// - /// `ε` is the current main tolerance and `τ` a scaling factor for the regulariser. - fn verify_merge_candidate<G, BT>( - &self, - d : &mut BTFN<F, G, BT, N>, - μ : &DiscreteMeasure<Loc<F, N>, F>, - τ : F, - ε : F, - config : &FBGenericConfig<F>, - ) -> bool - where BT : BTSearch<F, N, Agg=Bounds<F>>, - G : SupportGenerator<F, N, Id=BT::Data>, - G::SupportType : Mapping<Loc<F, N>,Codomain=F> - + LocalAnalysis<F, Bounds<F>, N>; - - fn target_bounds(&self, τ : F, ε : F) -> Option<Bounds<F>>; - - /// Returns a scaling factor for the tolerance sequence. - /// - /// Typically this is the regularisation parameter. - fn tolerance_scaling(&self) -> F; +pub(crate) fn prune_with_stats<F: Float, const N: usize>(μ: &mut RNDM<F, N>) -> usize { + let n_before_prune = μ.len(); + μ.prune(); + debug_assert!(μ.len() <= n_before_prune); + n_before_prune - μ.len() } #[replace_float_literals(F::cast_from(literal))] -impl<F : Float + ToNalgebraRealField, const N : usize> RegTerm<F, N> for NonnegRadonRegTerm<F> -where Cube<F, N> : P2Minimise<Loc<F, N>, F> { - fn solve_findim( - &self, - mA : &DMatrix<F::MixedType>, - g : &DVector<F::MixedType>, - τ : F, - x : &mut DVector<F::MixedType>, - mA_normest : F, - ε : F, - config : &FBGenericConfig<F> - ) -> usize { - let inner_tolerance = ε * config.inner.tolerance_mult; - let inner_it = config.inner.iterator_options.stop_target(inner_tolerance); - let inner_τ = config.inner.τ0 / mA_normest; - quadratic_nonneg(config.inner.method, mA, g, τ * self.α(), x, - inner_τ, inner_it) - } - - #[inline] - fn find_tolerance_violation<G, BT>( - &self, - d : &mut BTFN<F, G, BT, N>, - τ : F, - ε : F, - skip_by_rough_check : bool, - config : &FBGenericConfig<F>, - ) -> Option<(Loc<F, N>, F, bool)> - where BT : BTSearch<F, N, Agg=Bounds<F>>, - G : SupportGenerator<F, N, Id=BT::Data>, - G::SupportType : Mapping<Loc<F, N>,Codomain=F> - + LocalAnalysis<F, Bounds<F>, N> { - let τα = τ * self.α(); - let keep_below = τα + ε; - let maximise_above = τα + ε * config.insertion_cutoff_factor; - let refinement_tolerance = ε * config.refinement.tolerance_mult; - - // If preliminary check indicates that we are in bonds, and if it otherwise matches - // the insertion strategy, skip insertion. - if skip_by_rough_check && d.bounds().upper() <= keep_below { - None - } else { - // If the rough check didn't indicate no insertion needed, find maximising point. - d.maximise_above(maximise_above, refinement_tolerance, config.refinement.max_steps) - .map(|(ξ, v_ξ)| (ξ, v_ξ, v_ξ <= keep_below)) - } - } - - fn verify_merge_candidate<G, BT>( - &self, - d : &mut BTFN<F, G, BT, N>, - μ : &DiscreteMeasure<Loc<F, N>, F>, - τ : F, - ε : F, - config : &FBGenericConfig<F>, - ) -> bool - where BT : BTSearch<F, N, Agg=Bounds<F>>, - G : SupportGenerator<F, N, Id=BT::Data>, - G::SupportType : Mapping<Loc<F, N>,Codomain=F> - + LocalAnalysis<F, Bounds<F>, N> { - let τα = τ * self.α(); - let refinement_tolerance = ε * config.refinement.tolerance_mult; - let merge_tolerance = config.merge_tolerance_mult * ε; - let keep_below = τα + merge_tolerance; - let keep_supp_above = τα - merge_tolerance; - let bnd = d.bounds(); - - return ( - bnd.lower() >= keep_supp_above - || - μ.iter_spikes().map(|&DeltaMeasure{ α : β, ref x }| { - (β == 0.0) || d.apply(x) >= keep_supp_above - }).all(std::convert::identity) - ) && ( - bnd.upper() <= keep_below - || - d.has_upper_bound(keep_below, refinement_tolerance, config.refinement.max_steps) - ) - } - - fn target_bounds(&self, τ : F, ε : F) -> Option<Bounds<F>> { - let τα = τ * self.α(); - Some(Bounds(τα - ε, τα + ε)) - } - - fn tolerance_scaling(&self) -> F { - self.α() - } +pub(crate) fn postprocess< + F: Float, + V: Euclidean<F> + Clone, + A: GEMV<F, RNDM<F, N>, Codomain = V>, + D: DataTerm<F, V, N>, + const N: usize, +>( + mut μ: RNDM<F, N>, + config: &FBGenericConfig<F>, + dataterm: D, + opA: &A, + b: &V, +) -> RNDM<F, N> +where + RNDM<F, N>: SpikeMerging<F>, + for<'a> &'a RNDM<F, N>: Instance<RNDM<F, N>>, +{ + μ.merge_spikes_fitness( + config.final_merging_method(), + |μ̃| dataterm.calculate_fit_op(μ̃, opA, b), + |&v| v, + ); + μ.prune(); + μ } -#[replace_float_literals(F::cast_from(literal))] -impl<F : Float + ToNalgebraRealField, const N : usize> RegTerm<F, N> for RadonRegTerm<F> -where Cube<F, N> : P2Minimise<Loc<F, N>, F> { - fn solve_findim( - &self, - mA : &DMatrix<F::MixedType>, - g : &DVector<F::MixedType>, - τ : F, - x : &mut DVector<F::MixedType>, - mA_normest: F, - ε : F, - config : &FBGenericConfig<F> - ) -> usize { - let inner_tolerance = ε * config.inner.tolerance_mult; - let inner_it = config.inner.iterator_options.stop_target(inner_tolerance); - let inner_τ = config.inner.τ0 / mA_normest; - quadratic_unconstrained(config.inner.method, mA, g, τ * self.α(), x, - inner_τ, inner_it) - } - - fn find_tolerance_violation<G, BT>( - &self, - d : &mut BTFN<F, G, BT, N>, - τ : F, - ε : F, - skip_by_rough_check : bool, - config : &FBGenericConfig<F>, - ) -> Option<(Loc<F, N>, F, bool)> - where BT : BTSearch<F, N, Agg=Bounds<F>>, - G : SupportGenerator<F, N, Id=BT::Data>, - G::SupportType : Mapping<Loc<F, N>,Codomain=F> - + LocalAnalysis<F, Bounds<F>, N> { - let τα = τ * self.α(); - let keep_below = τα + ε; - let keep_above = -τα - ε; - let maximise_above = τα + ε * config.insertion_cutoff_factor; - let minimise_below = -τα - ε * config.insertion_cutoff_factor; - let refinement_tolerance = ε * config.refinement.tolerance_mult; - - // If preliminary check indicates that we are in bonds, and if it otherwise matches - // the insertion strategy, skip insertion. - if skip_by_rough_check && Bounds(keep_above, keep_below).superset(&d.bounds()) { - None - } else { - // If the rough check didn't indicate no insertion needed, find maximising point. - let mx = d.maximise_above(maximise_above, refinement_tolerance, - config.refinement.max_steps); - let mi = d.minimise_below(minimise_below, refinement_tolerance, - config.refinement.max_steps); - - match (mx, mi) { - (None, None) => None, - (Some((ξ, v_ξ)), None) => Some((ξ, v_ξ, keep_below >= v_ξ)), - (None, Some((ζ, v_ζ))) => Some((ζ, v_ζ, keep_above <= v_ζ)), - (Some((ξ, v_ξ)), Some((ζ, v_ζ))) => { - if v_ξ - τα > τα - v_ζ { - Some((ξ, v_ξ, keep_below >= v_ξ)) - } else { - Some((ζ, v_ζ, keep_above <= v_ζ)) - } - } - } - } - } - - fn verify_merge_candidate<G, BT>( - &self, - d : &mut BTFN<F, G, BT, N>, - μ : &DiscreteMeasure<Loc<F, N>, F>, - τ : F, - ε : F, - config : &FBGenericConfig<F>, - ) -> bool - where BT : BTSearch<F, N, Agg=Bounds<F>>, - G : SupportGenerator<F, N, Id=BT::Data>, - G::SupportType : Mapping<Loc<F, N>,Codomain=F> - + LocalAnalysis<F, Bounds<F>, N> { - let τα = τ * self.α(); - let refinement_tolerance = ε * config.refinement.tolerance_mult; - let merge_tolerance = config.merge_tolerance_mult * ε; - let keep_below = τα + merge_tolerance; - let keep_above = -τα - merge_tolerance; - let keep_supp_pos_above = τα - merge_tolerance; - let keep_supp_neg_below = -τα + merge_tolerance; - let bnd = d.bounds(); - - return ( - (bnd.lower() >= keep_supp_pos_above && bnd.upper() <= keep_supp_neg_below) - || - μ.iter_spikes().map(|&DeltaMeasure{ α : β, ref x }| { - use std::cmp::Ordering::*; - match β.partial_cmp(&0.0) { - Some(Greater) => d.apply(x) >= keep_supp_pos_above, - Some(Less) => d.apply(x) <= keep_supp_neg_below, - _ => true, - } - }).all(std::convert::identity) - ) && ( - bnd.upper() <= keep_below - || - d.has_upper_bound(keep_below, refinement_tolerance, - config.refinement.max_steps) - ) && ( - bnd.lower() >= keep_above - || - d.has_lower_bound(keep_above, refinement_tolerance, - config.refinement.max_steps) - ) - } - - fn target_bounds(&self, τ : F, ε : F) -> Option<Bounds<F>> { - let τα = τ * self.α(); - Some(Bounds(-τα - ε, τα + ε)) - } - - fn tolerance_scaling(&self) -> F { - self.α() - } -} - - -/// Generic implementation of [`pointsource_fb_reg`]. +/// Iteratively solve the pointsource localisation problem using forward-backward splitting. /// -/// The method can be specialised to even primal-dual proximal splitting through the -/// [`FBSpecialisation`] parameter `specialisation`. -/// The settings in `config` have their [respective documentation](FBGenericConfig). `opA` is the +/// The settings in `config` have their [respective documentation](FBConfig). `opA` is the /// forward operator $A$, $b$ the observable, and $\lambda$ the regularisation weight. /// The operator `op𝒟` is used for forming the proximal term. Typically it is a convolution /// operator. Finally, the `iterator` is an outer loop verbosity and iteration count control /// as documented in [`alg_tools::iterate`]. /// +/// For details on the mathematical formulation, see the [module level](self) documentation. +/// /// The implementation relies on [`alg_tools::bisection_tree::BTFN`] presentations of /// sums of simple functions usign bisection trees, and the related /// [`alg_tools::bisection_tree::Aggregator`]s, to efficiently search for component functions @@ -729,233 +170,103 @@ /// /// Returns the final iterate. #[replace_float_literals(F::cast_from(literal))] -pub fn generic_pointsource_fb_reg< - 'a, F, I, A, GA, 𝒟, BTA, G𝒟, S, K, Spec, Reg, const N : usize ->( - opA : &'a A, - reg : Reg, - op𝒟 : &'a 𝒟, - mut τ : F, - config : &FBGenericConfig<F>, - iterator : I, - mut plotter : SeqPlotter<F, N>, - mut residual : A::Observable, - mut specialisation : Spec -) -> DiscreteMeasure<Loc<F, N>, F> -where F : Float + ToNalgebraRealField, - I : AlgIteratorFactory<IterInfo<F, N>>, - Spec : FBSpecialisation<F, A::Observable, N>, - A::Observable : std::ops::MulAssign<F>, - GA : SupportGenerator<F, N, SupportType = S, Id = usize> + Clone, - A : ForwardModel<Loc<F, N>, F, PreadjointCodomain = BTFN<F, GA, BTA, N>> - + Lipschitz<𝒟, FloatType=F>, - BTA : BTSearch<F, N, Data=usize, Agg=Bounds<F>>, - G𝒟 : SupportGenerator<F, N, SupportType = K, Id = usize> + Clone, - 𝒟 : DiscreteMeasureOp<Loc<F, N>, F, PreCodomain = PreBTFN<F, G𝒟, N>>, - 𝒟::Codomain : RealMapping<F, N>, - S: RealMapping<F, N> + LocalAnalysis<F, Bounds<F>, N>, - K: RealMapping<F, N> + LocalAnalysis<F, Bounds<F>, N>, - BTNodeLookup: BTNode<F, usize, Bounds<F>, N>, - PlotLookup : Plotting<N>, - DiscreteMeasure<Loc<F, N>, F> : SpikeMerging<F>, - Reg : RegTerm<F, N> { - +pub fn pointsource_fb_reg<F, I, A, Reg, P, const N: usize>( + opA: &A, + b: &A::Observable, + reg: Reg, + prox_penalty: &P, + fbconfig: &FBConfig<F>, + iterator: I, + mut plotter: SeqPlotter<F, N>, +) -> RNDM<F, N> +where + F: Float + ToNalgebraRealField, + I: AlgIteratorFactory<IterInfo<F, N>>, + for<'b> &'b A::Observable: std::ops::Neg<Output = A::Observable>, + A: ForwardModel<RNDM<F, N>, F> + AdjointProductBoundedBy<RNDM<F, N>, P, FloatType = F>, + A::PreadjointCodomain: RealMapping<F, N>, + PlotLookup: Plotting<N>, + RNDM<F, N>: SpikeMerging<F>, + Reg: RegTerm<F, N>, + P: ProxPenalty<F, A::PreadjointCodomain, Reg, N>, +{ // Set up parameters - let quiet = iterator.is_quiet(); - let op𝒟norm = op𝒟.opnorm_bound(); + let config = &fbconfig.generic; + let τ = fbconfig.τ0 / opA.adjoint_product_bound(prox_penalty).unwrap(); // We multiply tolerance by τ for FB since our subproblems depending on tolerances are scaled // by τ compared to the conditional gradient approach. let tolerance = config.tolerance * τ * reg.tolerance_scaling(); let mut ε = tolerance.initial(); - // Initialise operators - let preadjA = opA.preadjoint(); - // Initialise iterates let mut μ = DiscreteMeasure::new(); - - let mut inner_iters = 0; - let mut this_iters = 0; - let mut pruned = 0; - let mut merged = 0; + let mut residual = -b; - let μ_diff = |μ_new : &DiscreteMeasure<Loc<F, N>, F>, - μ_base : &DiscreteMeasure<Loc<F, N>, F>| { - let mut ν : DiscreteMeasure<Loc<F, N>, F> = match config.insertion_style { - InsertionStyle::Reuse => { - μ_new.iter_spikes() - .zip(μ_base.iter_masses().chain(std::iter::repeat(0.0))) - .map(|(δ, α_base)| (δ.x, α_base - δ.α)) - .collect() - }, - InsertionStyle::Zero => { - μ_new.iter_spikes() - .map(|δ| -δ) - .chain(μ_base.iter_spikes().copied()) - .collect() - } - }; - ν.prune(); // Potential small performance improvement - ν + // Statistics + let full_stats = |residual: &A::Observable, μ: &RNDM<F, N>, ε, stats| IterInfo { + value: residual.norm2_squared_div2() + reg.apply(μ), + n_spikes: μ.len(), + ε, + //postprocessing: config.postprocessing.then(|| μ.clone()), + ..stats }; + let mut stats = IterInfo::new(); // Run the algorithm - iterator.iterate(|state| { - // Maximum insertion count and measure difference calculation depend on insertion style. - let (m, warn_insertions) = match (state.iteration(), config.bootstrap_insertions) { - (i, Some((l, k))) if i <= l => (k, false), - _ => (config.max_insertions, !quiet), - }; - let max_insertions = match config.insertion_style { - InsertionStyle::Zero => { - todo!("InsertionStyle::Zero does not currently work with FISTA, so diabled."); - // let n = μ.len(); - // μ = DiscreteMeasure::new(); - // n + m - }, - InsertionStyle::Reuse => m, - }; - + for state in iterator.iter_init(|| full_stats(&residual, &μ, ε, stats.clone())) { // Calculate smooth part of surrogate model. - // Using `std::mem::replace` here is not ideal, and expects that `empty_observable` - // has no significant overhead. For some reosn Rust doesn't allow us simply moving - // the residual and replacing it below before the end of this closure. - residual *= -τ; - let r = std::mem::replace(&mut residual, opA.empty_observable()); - let minus_τv = preadjA.apply(r); // minus_τv = -τA^*(Aμ^k-b) - // TODO: should avoid a second copy of μ here; μ_base already stores a copy. - let ω0 = op𝒟.apply(μ.clone()); // 𝒟μ^k - //let g = &minus_τv + ω0; // Linear term of surrogate model + let mut τv = opA.preadjoint().apply(residual * τ); // Save current base point let μ_base = μ.clone(); - - // Add points to support until within error tolerance or maximum insertion count reached. - let mut count = 0; - let (within_tolerances, d) = 'insertion: loop { - if μ.len() > 0 { - // Form finite-dimensional subproblem. The subproblem references to the original μ^k - // from the beginning of the iteration are all contained in the immutable c and g. - let à = op𝒟.findim_matrix(μ.iter_locations()); - let g̃ = DVector::from_iterator(μ.len(), - μ.iter_locations() - .map(|ζ| minus_τv.apply(ζ) + ω0.apply(ζ)) - .map(F::to_nalgebra_mixed)); - let mut x = μ.masses_dvector(); - // The gradient of the forward component of the inner objective is C^*𝒟Cx - g̃. - // We have |C^*𝒟Cx|_2 = sup_{|z|_2 ≤ 1} ⟨z, C^*𝒟Cx⟩ = sup_{|z|_2 ≤ 1} ⟨Cz|𝒟Cx⟩ - // ≤ sup_{|z|_2 ≤ 1} |Cz|_ℳ |𝒟Cx|_∞ ≤ sup_{|z|_2 ≤ 1} |Cz|_ℳ |𝒟| |Cx|_ℳ - // ≤ sup_{|z|_2 ≤ 1} |z|_1 |𝒟| |x|_1 ≤ sup_{|z|_2 ≤ 1} n |z|_2 |𝒟| |x|_2 - // = n |𝒟| |x|_2, where n is the number of points. Therefore - let Ã_normest = op𝒟norm * F::cast_from(μ.len()); - - // Solve finite-dimensional subproblem. - inner_iters += reg.solve_findim(&Ã, &g̃, τ, &mut x, Ã_normest, ε, config); - - // Update masses of μ based on solution of finite-dimensional subproblem. - μ.set_masses_dvector(&x); - } - - // Form d = ω0 - τv - 𝒟μ = -𝒟(μ - μ^k) - τv for checking the proximate optimality - // conditions in the predual space, and finding new points for insertion, if necessary. - let mut d = &minus_τv + op𝒟.preapply(μ_diff(&μ, &μ_base)); + // Insert and reweigh + let (maybe_d, _within_tolerances) = prox_penalty.insert_and_reweigh( + &mut μ, &mut τv, &μ_base, None, τ, ε, config, ®, &state, &mut stats, + ); - // If no merging heuristic is used, let's be more conservative about spike insertion, - // and skip it after first round. If merging is done, being more greedy about spike - // insertion also seems to improve performance. - let skip_by_rough_check = if let SpikeMergingMethod::None = config.merging { - false - } else { - count > 0 - }; - - // Find a spike to insert, if needed - let (ξ, _v_ξ, in_bounds) = match reg.find_tolerance_violation( - &mut d, τ, ε, skip_by_rough_check, config - ) { - None => break 'insertion (true, d), - Some(res) => res, - }; - - // Break if maximum insertion count reached - if count >= max_insertions { - break 'insertion (in_bounds, d) - } - - // No point in optimising the weight here; the finite-dimensional algorithm is fast. - μ += DeltaMeasure { x : ξ, α : 0.0 }; - count += 1; - }; - - if !within_tolerances && warn_insertions { - // Complain (but continue) if we failed to get within tolerances - // by inserting more points. - let err = format!("Maximum insertions reached without achieving \ - subproblem solution tolerance"); - println!("{}", err.red()); + // Prune and possibly merge spikes + if config.merge_now(&state) { + stats.merged += prox_penalty.merge_spikes( + &mut μ, + &mut τv, + &μ_base, + None, + τ, + ε, + config, + ®, + Some(|μ̃: &RNDM<F, N>| L2Squared.calculate_fit_op(μ̃, opA, b)), + ); } - // Merge spikes - if state.iteration() % config.merge_every == 0 { - let n_before_merge = μ.len(); - μ.merge_spikes(config.merging, |μ_candidate| { - let mut d = &minus_τv + op𝒟.preapply(μ_diff(&μ_candidate, &μ_base)); + stats.pruned += prune_with_stats(&mut μ); + + // Update residual + residual = calculate_residual(&μ, opA, b); + + let iter = state.iteration(); + stats.this_iters += 1; - reg.verify_merge_candidate(&mut d, μ_candidate, τ, ε, &config) - .then_some(()) - }); - debug_assert!(μ.len() >= n_before_merge); - merged += μ.len() - n_before_merge; - } - - let n_before_prune = μ.len(); - (residual, τ) = match specialisation.update(&mut μ, &μ_base) { - (r, None) => (r, τ), - (r, Some(new_τ)) => (r, new_τ) - }; - debug_assert!(μ.len() <= n_before_prune); - pruned += n_before_prune - μ.len(); - - this_iters += 1; + // Give statistics if needed + state.if_verbose(|| { + plotter.plot_spikes(iter, maybe_d.as_ref(), Some(&τv), &μ); + full_stats( + &residual, + &μ, + ε, + std::mem::replace(&mut stats, IterInfo::new()), + ) + }); // Update main tolerance for next iteration - let ε_prev = ε; - ε = tolerance.update(ε, state.iteration()); + ε = tolerance.update(ε, iter); + } - // Give function value if needed - state.if_verbose(|| { - let value_μ = specialisation.value_μ(&μ); - // Plot if so requested - plotter.plot_spikes( - format!("iter {} end; {}", state.iteration(), within_tolerances), &d, - "start".to_string(), Some(&minus_τv), - reg.target_bounds(τ, ε_prev), value_μ, - ); - // Calculate mean inner iterations and reset relevant counters. - // Return the statistics - let res = IterInfo { - value : specialisation.calculate_fit(&μ, &residual) + reg.apply(value_μ), - n_spikes : value_μ.len(), - inner_iters, - this_iters, - merged, - pruned, - ε : ε_prev, - postprocessing: config.postprocessing.then(|| value_μ.clone()), - }; - inner_iters = 0; - this_iters = 0; - merged = 0; - pruned = 0; - res - }) - }); - - specialisation.postprocess(μ, config.final_merging) + postprocess(μ, config, L2Squared, opA, b) } -/// Iteratively solve the pointsource localisation problem using forward-backward splitting +/// Iteratively solve the pointsource localisation problem using inertial forward-backward splitting. /// /// The settings in `config` have their [respective documentation](FBConfig). `opA` is the /// forward operator $A$, $b$ the observable, and $\lambda$ the regularisation weight. @@ -965,118 +276,114 @@ /// /// For details on the mathematical formulation, see the [module level](self) documentation. /// +/// The implementation relies on [`alg_tools::bisection_tree::BTFN`] presentations of +/// sums of simple functions usign bisection trees, and the related +/// [`alg_tools::bisection_tree::Aggregator`]s, to efficiently search for component functions +/// active at a specific points, and to maximise their sums. Through the implementation of the +/// [`alg_tools::bisection_tree::BT`] bisection trees, it also relies on the copy-on-write features +/// of [`std::sync::Arc`] to only update relevant parts of the bisection tree when adding functions. +/// /// Returns the final iterate. #[replace_float_literals(F::cast_from(literal))] -pub fn pointsource_fb_reg<'a, F, I, A, GA, 𝒟, BTA, G𝒟, S, K, Reg, const N : usize>( - opA : &'a A, - b : &A::Observable, - reg : Reg, - op𝒟 : &'a 𝒟, - config : &FBConfig<F>, - iterator : I, - plotter : SeqPlotter<F, N>, -) -> DiscreteMeasure<Loc<F, N>, F> -where F : Float + ToNalgebraRealField, - I : AlgIteratorFactory<IterInfo<F, N>>, - for<'b> &'b A::Observable : std::ops::Neg<Output=A::Observable>, - //+ std::ops::Mul<F, Output=A::Observable>, <-- FIXME: compiler overflow - A::Observable : std::ops::MulAssign<F>, - GA : SupportGenerator<F, N, SupportType = S, Id = usize> + Clone, - A : ForwardModel<Loc<F, N>, F, PreadjointCodomain = BTFN<F, GA, BTA, N>> - + Lipschitz<𝒟, FloatType=F>, - BTA : BTSearch<F, N, Data=usize, Agg=Bounds<F>>, - G𝒟 : SupportGenerator<F, N, SupportType = K, Id = usize> + Clone, - 𝒟 : DiscreteMeasureOp<Loc<F, N>, F, PreCodomain = PreBTFN<F, G𝒟, N>>, - 𝒟::Codomain : RealMapping<F, N>, - S: RealMapping<F, N> + LocalAnalysis<F, Bounds<F>, N>, - K: RealMapping<F, N> + LocalAnalysis<F, Bounds<F>, N>, - BTNodeLookup: BTNode<F, usize, Bounds<F>, N>, - Cube<F, N>: P2Minimise<Loc<F, N>, F>, - PlotLookup : Plotting<N>, - DiscreteMeasure<Loc<F, N>, F> : SpikeMerging<F>, - Reg : RegTerm<F, N> { +pub fn pointsource_fista_reg<F, I, A, Reg, P, const N: usize>( + opA: &A, + b: &A::Observable, + reg: Reg, + prox_penalty: &P, + fbconfig: &FBConfig<F>, + iterator: I, + mut plotter: SeqPlotter<F, N>, +) -> RNDM<F, N> +where + F: Float + ToNalgebraRealField, + I: AlgIteratorFactory<IterInfo<F, N>>, + for<'b> &'b A::Observable: std::ops::Neg<Output = A::Observable>, + A: ForwardModel<RNDM<F, N>, F> + AdjointProductBoundedBy<RNDM<F, N>, P, FloatType = F>, + A::PreadjointCodomain: RealMapping<F, N>, + PlotLookup: Plotting<N>, + RNDM<F, N>: SpikeMerging<F>, + Reg: RegTerm<F, N>, + P: ProxPenalty<F, A::PreadjointCodomain, Reg, N>, +{ + // Set up parameters + let config = &fbconfig.generic; + let τ = fbconfig.τ0 / opA.adjoint_product_bound(prox_penalty).unwrap(); + let mut λ = 1.0; + // We multiply tolerance by τ for FB since our subproblems depending on tolerances are scaled + // by τ compared to the conditional gradient approach. + let tolerance = config.tolerance * τ * reg.tolerance_scaling(); + let mut ε = tolerance.initial(); - let initial_residual = -b; - let τ = config.τ0/opA.lipschitz_factor(&op𝒟).unwrap(); + // Initialise iterates + let mut μ = DiscreteMeasure::new(); + let mut μ_prev = DiscreteMeasure::new(); + let mut residual = -b; + let mut warned_merging = false; - match config.meta { - FBMetaAlgorithm::None => generic_pointsource_fb_reg( - opA, reg, op𝒟, τ, &config.insertion, iterator, plotter, initial_residual, - BasicFB{ b, opA }, - ), - FBMetaAlgorithm::InertiaFISTA => generic_pointsource_fb_reg( - opA, reg, op𝒟, τ, &config.insertion, iterator, plotter, initial_residual, - FISTA{ b, opA, λ : 1.0, μ_prev : DiscreteMeasure::new() }, - ), - } -} + // Statistics + let full_stats = |ν: &RNDM<F, N>, ε, stats| IterInfo { + value: L2Squared.calculate_fit_op(ν, opA, b) + reg.apply(ν), + n_spikes: ν.len(), + ε, + // postprocessing: config.postprocessing.then(|| ν.clone()), + ..stats + }; + let mut stats = IterInfo::new(); -// -// Deprecated interfaces -// + // Run the algorithm + for state in iterator.iter_init(|| full_stats(&μ, ε, stats.clone())) { + // Calculate smooth part of surrogate model. + let mut τv = opA.preadjoint().apply(residual * τ); -#[deprecated(note = "Use `pointsource_fb_reg`")] -pub fn pointsource_fb<'a, F, I, A, GA, 𝒟, BTA, G𝒟, S, K, const N : usize>( - opA : &'a A, - b : &A::Observable, - α : F, - op𝒟 : &'a 𝒟, - config : &FBConfig<F>, - iterator : I, - plotter : SeqPlotter<F, N> -) -> DiscreteMeasure<Loc<F, N>, F> -where F : Float + ToNalgebraRealField, - I : AlgIteratorFactory<IterInfo<F, N>>, - for<'b> &'b A::Observable : std::ops::Neg<Output=A::Observable>, - A::Observable : std::ops::MulAssign<F>, - GA : SupportGenerator<F, N, SupportType = S, Id = usize> + Clone, - A : ForwardModel<Loc<F, N>, F, PreadjointCodomain = BTFN<F, GA, BTA, N>> - + Lipschitz<𝒟, FloatType=F>, - BTA : BTSearch<F, N, Data=usize, Agg=Bounds<F>>, - G𝒟 : SupportGenerator<F, N, SupportType = K, Id = usize> + Clone, - 𝒟 : DiscreteMeasureOp<Loc<F, N>, F, PreCodomain = PreBTFN<F, G𝒟, N>>, - 𝒟::Codomain : RealMapping<F, N>, - S: RealMapping<F, N> + LocalAnalysis<F, Bounds<F>, N>, - K: RealMapping<F, N> + LocalAnalysis<F, Bounds<F>, N>, - BTNodeLookup: BTNode<F, usize, Bounds<F>, N>, - Cube<F, N>: P2Minimise<Loc<F, N>, F>, - PlotLookup : Plotting<N>, - DiscreteMeasure<Loc<F, N>, F> : SpikeMerging<F> { + // Save current base point + let μ_base = μ.clone(); + + // Insert new spikes and reweigh + let (maybe_d, _within_tolerances) = prox_penalty.insert_and_reweigh( + &mut μ, &mut τv, &μ_base, None, τ, ε, config, ®, &state, &mut stats, + ); - pointsource_fb_reg(opA, b, NonnegRadonRegTerm(α), op𝒟, config, iterator, plotter) -} + // (Do not) merge spikes. + if config.merge_now(&state) && !warned_merging { + let err = format!("Merging not supported for μFISTA"); + println!("{}", err.red()); + warned_merging = true; + } + // Update inertial prameters + let λ_prev = λ; + λ = 2.0 * λ_prev / (λ_prev + (4.0 + λ_prev * λ_prev).sqrt()); + let θ = λ / λ_prev - λ; -#[deprecated(note = "Use `generic_pointsource_fb_reg`")] -pub fn generic_pointsource_fb<'a, F, I, A, GA, 𝒟, BTA, G𝒟, S, K, Spec, const N : usize>( - opA : &'a A, - α : F, - op𝒟 : &'a 𝒟, - τ : F, - config : &FBGenericConfig<F>, - iterator : I, - plotter : SeqPlotter<F, N>, - residual : A::Observable, - specialisation : Spec, -) -> DiscreteMeasure<Loc<F, N>, F> -where F : Float + ToNalgebraRealField, - I : AlgIteratorFactory<IterInfo<F, N>>, - Spec : FBSpecialisation<F, A::Observable, N>, - A::Observable : std::ops::MulAssign<F>, - GA : SupportGenerator<F, N, SupportType = S, Id = usize> + Clone, - A : ForwardModel<Loc<F, N>, F, PreadjointCodomain = BTFN<F, GA, BTA, N>> - + Lipschitz<𝒟, FloatType=F>, - BTA : BTSearch<F, N, Data=usize, Agg=Bounds<F>>, - G𝒟 : SupportGenerator<F, N, SupportType = K, Id = usize> + Clone, - 𝒟 : DiscreteMeasureOp<Loc<F, N>, F, PreCodomain = PreBTFN<F, G𝒟, N>>, - 𝒟::Codomain : RealMapping<F, N>, - S: RealMapping<F, N> + LocalAnalysis<F, Bounds<F>, N>, - K: RealMapping<F, N> + LocalAnalysis<F, Bounds<F>, N>, - BTNodeLookup: BTNode<F, usize, Bounds<F>, N>, - Cube<F, N>: P2Minimise<Loc<F, N>, F>, - PlotLookup : Plotting<N>, - DiscreteMeasure<Loc<F, N>, F> : SpikeMerging<F> { + // Perform inertial update on μ. + // This computes μ ← (1 + θ) * μ - θ * μ_prev, pruning spikes where both μ + // and μ_prev have zero weight. Since both have weights from the finite-dimensional + // subproblem with a proximal projection step, this is likely to happen when the + // spike is not needed. A copy of the pruned μ without artithmetic performed is + // stored in μ_prev. + let n_before_prune = μ.len(); + μ.pruning_sub(1.0 + θ, θ, &mut μ_prev); + //let μ_new = (&μ * (1.0 + θ)).sub_matching(&(&μ_prev * θ)); + // μ_prev = μ; + // μ = μ_new; + debug_assert!(μ.len() <= n_before_prune); + stats.pruned += n_before_prune - μ.len(); - generic_pointsource_fb_reg(opA, NonnegRadonRegTerm(α), op𝒟, τ, config, iterator, plotter, - residual, specialisation) + // Update residual + residual = calculate_residual(&μ, opA, b); + + let iter = state.iteration(); + stats.this_iters += 1; + + // Give statistics if needed + state.if_verbose(|| { + plotter.plot_spikes(iter, maybe_d.as_ref(), Some(&τv), &μ_prev); + full_stats(&μ_prev, ε, std::mem::replace(&mut stats, IterInfo::new())) + }); + + // Update main tolerance for next iteration + ε = tolerance.update(ε, iter); + } + + postprocess(μ_prev, config, L2Squared, opA, b) }
--- a/src/forward_model.rs Tue Aug 01 10:25:09 2023 +0300 +++ b/src/forward_model.rs Mon Feb 17 13:54:53 2025 -0500 @@ -2,677 +2,82 @@ Forward models from discrete measures to observations. */ -use numeric_literals::replace_float_literals; -use nalgebra::base::{ - DMatrix, - DVector -}; -use std::iter::Zip; -use std::ops::RangeFrom; -use std::marker::PhantomData; - -pub use alg_tools::linops::*; +use alg_tools::error::DynError; use alg_tools::euclidean::Euclidean; -use alg_tools::norms::{ - L1, Linfinity, Norm -}; -use alg_tools::bisection_tree::*; -use alg_tools::mapping::RealMapping; -use alg_tools::lingrid::*; -use alg_tools::iter::{MapX, Mappable}; -use alg_tools::nalgebra_support::ToNalgebraRealField; -use alg_tools::tabledump::write_csv; -use alg_tools::error::DynError; +use alg_tools::instance::Instance; +pub use alg_tools::linops::*; +use alg_tools::norms::{Norm, NormExponent, L2}; +use crate::measures::Radon; use crate::types::*; -use crate::measures::*; -use crate::seminorms::{ - Lipschitz, - ConvolutionOp, - SimpleConvolutionKernel, -}; -use crate::kernels::{ - Convolution, - AutoConvolution, - BoundedBy, -}; - -pub type RNDM<F, const N : usize> = DiscreteMeasure<Loc<F,N>, F>; +pub mod bias; +pub mod sensor_grid; /// `ForwardeModel`s are bounded preadjointable linear operators $A ∈ 𝕃(𝒵(Ω); E)$ /// where $𝒵(Ω) ⊂ ℳ(Ω)$ is the space of sums of delta measures, presented by -/// [`DiscreteMeasure`], and $E$ is a [`Euclidean`] space. -pub trait ForwardModel<Domain, F : Float + ToNalgebraRealField> -: BoundedLinear<DiscreteMeasure<Domain, F>, Codomain=Self::Observable, FloatType=F> -+ GEMV<F, DiscreteMeasure<Domain, F>, Self::Observable> -+ Linear<DeltaMeasure<Domain, F>, Codomain=Self::Observable> -+ Preadjointable<DiscreteMeasure<Domain, F>, Self::Observable> { +/// [`crate::measures::DiscreteMeasure`], and $E$ is a [`Euclidean`] space. +pub trait ForwardModel<Domain: Space, F: Float = f64, E: NormExponent = Radon>: + BoundedLinear<Domain, E, L2, F, Codomain = Self::Observable> + + GEMV<F, Domain, Self::Observable> + + Preadjointable<Domain, Self::Observable> +where + for<'a> Self::Observable: Instance<Self::Observable>, + Domain: Norm<F, E>, +{ /// The codomain or value space (of “observables”) for this operator. /// It is assumed to be a [`Euclidean`] space, and therefore also (identified with) /// the domain of the preadjoint. - type Observable : Euclidean<F, Output=Self::Observable> - + AXPY<F> - + Clone; - - /// Return A_*A and A_* b - fn findim_quadratic_model( - &self, - μ : &DiscreteMeasure<Domain, F>, - b : &Self::Observable - ) -> (DMatrix<F::MixedType>, DVector<F::MixedType>); + type Observable: Euclidean<F, Output = Self::Observable> + AXPY<F> + Space + Clone; /// Write an observable into a file. - fn write_observable(&self, b : &Self::Observable, prefix : String) -> DynError; + fn write_observable(&self, b: &Self::Observable, prefix: String) -> DynError; /// Returns a zero observable fn zero_observable(&self) -> Self::Observable; - - /// Returns an empty (uninitialised) observable. - /// - /// This is used as a placeholder for temporary [`std::mem::replace`] move operations. - fn empty_observable(&self) -> Self::Observable; -} - -pub type ShiftedSensor<F, S, P, const N : usize> = Shift<Convolution<S, P>, F, N>; - -/// Trait for physical convolution models. Has blanket implementation for all cases. -pub trait Spread<F : Float, const N : usize> -: 'static + Clone + Support<F, N> + RealMapping<F, N> + Bounded<F> {} - -impl<F, T, const N : usize> Spread<F, N> for T -where F : Float, - T : 'static + Clone + Support<F, N> + Bounded<F> + RealMapping<F, N> {} - -/// Trait for compactly supported sensors. Has blanket implementation for all cases. -pub trait Sensor<F : Float, const N : usize> : Spread<F, N> + Norm<F, L1> + Norm<F, Linfinity> {} - -impl<F, T, const N : usize> Sensor<F, N> for T -where F : Float, - T : Spread<F, N> + Norm<F, L1> + Norm<F, Linfinity> {} - - -pub trait SensorGridBT<F, S, P, const N : usize> : -Clone + BTImpl<F, N, Data=usize, Agg=Bounds<F>> -where F : Float, - S : Sensor<F, N>, - P : Spread<F, N> {} - -impl<F, S, P, T, const N : usize> -SensorGridBT<F, S, P, N> -for T -where T : Clone + BTImpl<F, N, Data=usize, Agg=Bounds<F>>, - F : Float, - S : Sensor<F, N>, - P : Spread<F, N> {} - -// We need type alias bounds to access associated types -#[allow(type_alias_bounds)] -type SensorGridBTFN<F, S, P, BT : SensorGridBT<F, S, P, N>, const N : usize> -= BTFN<F, SensorGridSupportGenerator<F, S, P, N>, BT, N>; - -/// Sensor grid forward model -#[derive(Clone)] -pub struct SensorGrid<F, S, P, BT, const N : usize> -where F : Float, - S : Sensor<F, N>, - P : Spread<F, N>, - Convolution<S, P> : Spread<F, N>, - BT : SensorGridBT<F, S, P, N>, { - domain : Cube<F, N>, - sensor_count : [usize; N], - sensor : S, - spread : P, - base_sensor : Convolution<S, P>, - bt : BT, } -impl<F, S, P, BT, const N : usize> SensorGrid<F, S, P, BT, N> -where F : Float, - BT : SensorGridBT<F, S, P, N>, - S : Sensor<F, N>, - P : Spread<F, N>, - Convolution<S, P> : Spread<F, N>, - ShiftedSensor<F, S, P, N> : LocalAnalysis<F, BT::Agg, N> { - - pub fn new( - domain : Cube<F, N>, - sensor_count : [usize; N], - sensor : S, - spread : P, - depth : BT::Depth - ) -> Self { - let base_sensor = Convolution(sensor.clone(), spread.clone()); - let bt = BT::new(domain, depth); - let mut sensorgrid = SensorGrid { - domain, - sensor_count, - sensor, - spread, - base_sensor, - bt, - }; - - for (x, id) in sensorgrid.grid().into_iter().zip(0usize..) { - let s = sensorgrid.shifted_sensor(x); - sensorgrid.bt.insert(id, &s); - } - - sensorgrid - } - - pub fn grid(&self) -> LinGrid<F, N> { - lingrid_centered(&self.domain, &self.sensor_count) - } - - pub fn n_sensors(&self) -> usize { - self.sensor_count.iter().product() - } - - #[inline] - fn shifted_sensor(&self, x : Loc<F, N>) -> ShiftedSensor<F, S, P, N> { - self.base_sensor.clone().shift(x) - } - - #[inline] - fn _zero_observable(&self) -> DVector<F> { - DVector::zeros(self.n_sensors()) - } -} - -impl<F, S, P, BT, const N : usize> Apply<RNDM<F, N>> for SensorGrid<F, S, P, BT, N> -where F : Float, - BT : SensorGridBT<F, S, P, N>, - S : Sensor<F, N>, - P : Spread<F, N>, - Convolution<S, P> : Spread<F, N>, - ShiftedSensor<F, S, P, N> : LocalAnalysis<F, BT::Agg, N> { - - type Output = DVector<F>; - - #[inline] - fn apply(&self, μ : RNDM<F, N>) -> DVector<F> { - self.apply(&μ) - } -} - -impl<'a, F, S, P, BT, const N : usize> Apply<&'a RNDM<F, N>> for SensorGrid<F, S, P, BT, N> -where F : Float, - BT : SensorGridBT<F, S, P, N>, - S : Sensor<F, N>, - P : Spread<F, N>, - Convolution<S, P> : Spread<F, N>, - ShiftedSensor<F, S, P, N> : LocalAnalysis<F, BT::Agg, N> { - - type Output = DVector<F>; - - fn apply(&self, μ : &'a RNDM<F, N>) -> DVector<F> { - let mut res = self._zero_observable(); - self.apply_add(&mut res, μ); - res - } +/// Trait for operators $A$ for which $A_*A$ is bounded by some other operator. +pub trait AdjointProductBoundedBy<Domain: Space, D>: Linear<Domain> { + type FloatType: Float; + /// Return $L$ such that $A_*A ≤ LD$. + fn adjoint_product_bound(&self, other: &D) -> Option<Self::FloatType>; } -impl<F, S, P, BT, const N : usize> Linear<RNDM<F, N>> for SensorGrid<F, S, P, BT, N> -where F : Float, - BT : SensorGridBT<F, S, P, N>, - S : Sensor<F, N>, - P : Spread<F, N>, - Convolution<S, P> : Spread<F, N>, - ShiftedSensor<F, S, P, N> : LocalAnalysis<F, BT::Agg, N> { - type Codomain = DVector<F>; -} - - -#[replace_float_literals(F::cast_from(literal))] -impl<F, S, P, BT, const N : usize> GEMV<F, RNDM<F, N>, DVector<F>> for SensorGrid<F, S, P, BT, N> -where F : Float, - BT : SensorGridBT<F, S, P, N>, - S : Sensor<F, N>, - P : Spread<F, N>, - Convolution<S, P> : Spread<F, N>, - ShiftedSensor<F, S, P, N> : LocalAnalysis<F, BT::Agg, N> { - - fn gemv(&self, y : &mut DVector<F>, α : F, μ : &RNDM<F, N>, β : F) { - let grid = self.grid(); - if β == 0.0 { - y.fill(0.0) - } else if β != 1.0 { - *y *= β; // Need to multiply first, as we have to be able to add to y. - } - if α == 1.0 { - self.apply_add(y, μ) - } else { - for δ in μ.iter_spikes() { - for &d in self.bt.iter_at(&δ.x) { - let sensor = self.shifted_sensor(grid.entry_linear_unchecked(d)); - y[d] += sensor.apply(&δ.x) * (α * δ.α); - } - } - } - } - - fn apply_add(&self, y : &mut DVector<F>, μ : &RNDM<F, N>) { - let grid = self.grid(); - for δ in μ.iter_spikes() { - for &d in self.bt.iter_at(&δ.x) { - let sensor = self.shifted_sensor(grid.entry_linear_unchecked(d)); - y[d] += sensor.apply(&δ.x) * δ.α; - } - } - } - -} - -impl<F, S, P, BT, const N : usize> Apply<DeltaMeasure<Loc<F, N>, F>> -for SensorGrid<F, S, P, BT, N> -where F : Float, - BT : SensorGridBT<F, S, P, N>, - S : Sensor<F, N>, - P : Spread<F, N>, - Convolution<S, P> : Spread<F, N>, - ShiftedSensor<F, S, P, N> : LocalAnalysis<F, BT::Agg, N> { - - type Output = DVector<F>; - - #[inline] - fn apply(&self, δ : DeltaMeasure<Loc<F, N>, F>) -> DVector<F> { - self.apply(&δ) - } -} - -impl<'a, F, S, P, BT, const N : usize> Apply<&'a DeltaMeasure<Loc<F, N>, F>> -for SensorGrid<F, S, P, BT, N> -where F : Float, - BT : SensorGridBT<F, S, P, N>, - S : Sensor<F, N>, - P : Spread<F, N>, - Convolution<S, P> : Spread<F, N>, - ShiftedSensor<F, S, P, N> : LocalAnalysis<F, BT::Agg, N> { - - type Output = DVector<F>; - - fn apply(&self, δ : &DeltaMeasure<Loc<F, N>, F>) -> DVector<F> { - let mut res = DVector::zeros(self.n_sensors()); - let grid = self.grid(); - for &d in self.bt.iter_at(&δ.x) { - let sensor = self.shifted_sensor(grid.entry_linear_unchecked(d)); - res[d] += sensor.apply(&δ.x) * δ.α; - } - res - } -} - -impl<F, S, P, BT, const N : usize> Linear<DeltaMeasure<Loc<F, N>, F>> for SensorGrid<F, S, P, BT, N> -where F : Float, - BT : SensorGridBT<F, S, P, N>, - S : Sensor<F, N>, - P : Spread<F, N>, - Convolution<S, P> : Spread<F, N>, - ShiftedSensor<F, S, P, N> : LocalAnalysis<F, BT::Agg, N> { - type Codomain = DVector<F>; -} - -impl<F, S, P, BT, const N : usize> BoundedLinear<RNDM<F, N>> for SensorGrid<F, S, P, BT, N> -where F : Float, - BT : SensorGridBT<F, S, P, N, Agg=Bounds<F>>, - S : Sensor<F, N>, - P : Spread<F, N>, - Convolution<S, P> : Spread<F, N>, - ShiftedSensor<F, S, P, N> : LocalAnalysis<F, BT::Agg, N> { - type FloatType = F; - - /// An estimate on the operator norm in $𝕃(ℳ(Ω); ℝ^n)$ with $ℳ(Ω)$ equipped - /// with the Radon norm, and $ℝ^n$ with the Euclidean norm. - fn opnorm_bound(&self) -> F { - // With {x_i}_{i=1}^n the grid centres and φ the kernel, we have - // |Aμ|_2 = sup_{|z|_2 ≤ 1} ⟨z,Αμ⟩ = sup_{|z|_2 ≤ 1} ⟨A^*z|μ⟩ - // ≤ sup_{|z|_2 ≤ 1} |A^*z|_∞ |μ|_ℳ - // = sup_{|z|_2 ≤ 1} |∑ φ(· - x_i)z_i|_∞ |μ|_ℳ - // ≤ sup_{|z|_2 ≤ 1} |φ|_∞ ∑ |z_i| |μ|_ℳ - // ≤ sup_{|z|_2 ≤ 1} |φ|_∞ √n |z|_2 |μ|_ℳ - // = |φ|_∞ √n |μ|_ℳ. - // Hence - let n = F::cast_from(self.n_sensors()); - self.base_sensor.bounds().uniform() * n.sqrt() - } -} - -type SensorGridPreadjoint<'a, A, F, const N : usize> = PreadjointHelper<'a, A, RNDM<F,N>>; - - -impl<F, S, P, BT, const N : usize> -Preadjointable<RNDM<F, N>, DVector<F>> -for SensorGrid<F, S, P, BT, N> -where F : Float, - BT : SensorGridBT<F, S, P, N>, - S : Sensor<F, N>, - P : Spread<F, N>, - Convolution<S, P> : Spread<F, N>, - ShiftedSensor<F, S, P, N> : LocalAnalysis<F, BT::Agg, N>, - Weighted<ShiftedSensor<F, S, P, N>, F> : LocalAnalysis<F, BT::Agg, N> { - type PreadjointCodomain = BTFN<F, SensorGridSupportGenerator<F, S, P, N>, BT, N>; - type Preadjoint<'a> = SensorGridPreadjoint<'a, Self, F, N> where Self : 'a; - - fn preadjoint(&self) -> Self::Preadjoint<'_> { - PreadjointHelper::new(self) - } -} - -#[derive(Clone,Debug)] -pub struct SensorGridSupportGenerator<F, S, P, const N : usize> -where F : Float, - S : Sensor<F, N>, - P : Spread<F, N> { - base_sensor : Convolution<S, P>, - grid : LinGrid<F, N>, - weights : DVector<F> +/// Trait for operators $A$ for which $A_*A$ is bounded by a diagonal operator. +pub trait AdjointProductPairBoundedBy<Domain: Space, D1, D2>: Linear<Domain> { + type FloatType: Float; + /// Return $(L, L_z)$ such that $A_*A ≤ (L_1 D_1, L_2 D_2)$. + fn adjoint_product_pair_bound( + &self, + other1: &D1, + other_2: &D2, + ) -> Option<(Self::FloatType, Self::FloatType)>; } -impl<F, S, P, const N : usize> SensorGridSupportGenerator<F, S, P, N> -where F : Float, - S : Sensor<F, N>, - P : Spread<F, N>, - Convolution<S, P> : Spread<F, N> { - - #[inline] - fn construct_sensor(&self, id : usize, w : F) -> Weighted<ShiftedSensor<F, S, P, N>, F> { - let x = self.grid.entry_linear_unchecked(id); - self.base_sensor.clone().shift(x).weigh(w) - } - - #[inline] - fn construct_sensor_and_id<'a>(&'a self, (id, w) : (usize, &'a F)) - -> (usize, Weighted<ShiftedSensor<F, S, P, N>, F>) { - (id.into(), self.construct_sensor(id, *w)) - } -} - -impl<F, S, P, const N : usize> SupportGenerator<F, N> -for SensorGridSupportGenerator<F, S, P, N> -where F : Float, - S : Sensor<F, N>, - P : Spread<F, N>, - Convolution<S, P> : Spread<F, N> { - type Id = usize; - type SupportType = Weighted<ShiftedSensor<F, S, P, N>, F>; - type AllDataIter<'a> = MapX<'a, Zip<RangeFrom<usize>, - std::slice::Iter<'a, F>>, - Self, - (Self::Id, Self::SupportType)> - where Self : 'a; - - #[inline] - fn support_for(&self, d : Self::Id) -> Self::SupportType { - self.construct_sensor(d, self.weights[d]) - } - - #[inline] - fn support_count(&self) -> usize { - self.weights.len() +/* +/// Trait for [`ForwardModel`]s whose preadjoint has Lipschitz values. +pub trait LipschitzValues { + type FloatType : Float; + /// Return (if one exists) a factor $L$ such that $A_*z$ is $L$-Lipschitz for all + /// $z$ in the unit ball. + fn value_unit_lipschitz_factor(&self) -> Option<Self::FloatType> { + None } - #[inline] - fn all_data(&self) -> Self::AllDataIter<'_> { - (0..).zip(self.weights.as_slice().iter()).mapX(self, Self::construct_sensor_and_id) - } -} - -/// Helper structure for constructing preadjoints of `S` where `S : Linear<X>`. -/// [`Linear`] needs to be implemented for each instance, but [`Adjointable`] -/// and [`BoundedLinear`] have blanket implementations. -#[derive(Clone,Debug)] -pub struct PreadjointHelper<'a, S : 'a, X> { - forward_op : &'a S, - _domain : PhantomData<X> -} - -impl<'a, S : 'a, X> PreadjointHelper<'a, S, X> { - pub fn new(forward_op : &'a S) -> Self { - PreadjointHelper { forward_op, _domain: PhantomData } - } -} - -impl<'a, X, Ypre, S> Adjointable<Ypre, X> -for PreadjointHelper<'a, S, X> -where Self : Linear<Ypre>, - S : Clone + Linear<X> { - type AdjointCodomain = S::Codomain; - type Adjoint<'b> = S where Self : 'b; - fn adjoint(&self) -> Self::Adjoint<'_> { - self.forward_op.clone() - } -} - -impl<'a, X, Ypre, S> BoundedLinear<Ypre> -for PreadjointHelper<'a, S, X> -where Self : Linear<Ypre>, - S : 'a + Clone + BoundedLinear<X> { - type FloatType = S::FloatType; - fn opnorm_bound(&self) -> Self::FloatType { - self.forward_op.opnorm_bound() - } -} - - -impl<'a, 'b, F, S, P, BT, const N : usize> Apply<&'b DVector<F>> -for PreadjointHelper<'a, SensorGrid<F, S, P, BT, N>, RNDM<F,N>> -where F : Float, - BT : SensorGridBT<F, S, P, N>, - S : Sensor<F, N>, - P : Spread<F, N>, - Convolution<S, P> : Spread<F, N>, - ShiftedSensor<F, S, P, N> : LocalAnalysis<F, BT::Agg, N>, - Weighted<ShiftedSensor<F, S, P, N>, F> : LocalAnalysis<F, BT::Agg, N> { - - type Output = SensorGridBTFN<F, S, P, BT, N>; - - fn apply(&self, x : &'b DVector<F>) -> Self::Output { - self.apply(x.clone()) - } -} - -impl<'a, F, S, P, BT, const N : usize> Apply<DVector<F>> -for PreadjointHelper<'a, SensorGrid<F, S, P, BT, N>, RNDM<F,N>> -where F : Float, - BT : SensorGridBT<F, S, P, N>, - S : Sensor<F, N>, - P : Spread<F, N>, - Convolution<S, P> : Spread<F, N>, - ShiftedSensor<F, S, P, N> : LocalAnalysis<F, BT::Agg, N>, - Weighted<ShiftedSensor<F, S, P, N>, F> : LocalAnalysis<F, BT::Agg, N> { - - type Output = SensorGridBTFN<F, S, P, BT, N>; - - fn apply(&self, x : DVector<F>) -> Self::Output { - let fwd = &self.forward_op; - let generator = SensorGridSupportGenerator{ - base_sensor : fwd.base_sensor.clone(), - grid : fwd.grid(), - weights : x - }; - BTFN::new_refresh(&fwd.bt, generator) + /// Return (if one exists) a factor $L$ such that $∇A_*z$ is $L$-Lipschitz for all + /// $z$ in the unit ball. + fn value_diff_unit_lipschitz_factor(&self) -> Option<Self::FloatType> { + None } } - -impl<'a, F, S, P, BT, const N : usize> Linear<DVector<F>> -for PreadjointHelper<'a, SensorGrid<F, S, P, BT, N>, RNDM<F,N>> -where F : Float, - BT : SensorGridBT<F, S, P, N>, - S : Sensor<F, N>, - P : Spread<F, N>, - Convolution<S, P> : Spread<F, N>, - ShiftedSensor<F, S, P, N> : LocalAnalysis<F, BT::Agg, N>, - Weighted<ShiftedSensor<F, S, P, N>, F> : LocalAnalysis<F, BT::Agg, N> { - - type Codomain = SensorGridBTFN<F, S, P, BT, N>; -} - -impl<F, S, P, BT, const N : usize> ForwardModel<Loc<F, N>, F> -for SensorGrid<F, S, P, BT, N> -where F : Float + ToNalgebraRealField<MixedType=F> + nalgebra::RealField, - BT : SensorGridBT<F, S, P, N>, - S : Sensor<F, N>, - P : Spread<F, N>, - Convolution<S, P> : Spread<F, N>, - ShiftedSensor<F, S, P, N> : LocalAnalysis<F, BT::Agg, N>, - Weighted<ShiftedSensor<F, S, P, N>, F> : LocalAnalysis<F, BT::Agg, N> { - type Observable = DVector<F>; - - fn findim_quadratic_model( - &self, - μ : &DiscreteMeasure<Loc<F, N>, F>, - b : &Self::Observable - ) -> (DMatrix<F::MixedType>, DVector<F::MixedType>) { - assert_eq!(b.len(), self.n_sensors()); - let mut mA = DMatrix::zeros(self.n_sensors(), μ.len()); - let grid = self.grid(); - for (mut mAcol, δ) in mA.column_iter_mut().zip(μ.iter_spikes()) { - for &d in self.bt.iter_at(&δ.x) { - let sensor = self.shifted_sensor(grid.entry_linear_unchecked(d)); - mAcol[d] += sensor.apply(&δ.x); - } - } - let mAt = mA.transpose(); - (&mAt * mA, &mAt * b) - } +*/ - fn write_observable(&self, b : &Self::Observable, prefix : String) -> DynError { - let it = self.grid().into_iter().zip(b.iter()).map(|(x, &v)| (x, v)); - write_csv(it, prefix + ".txt") - } - - #[inline] - fn zero_observable(&self) -> Self::Observable { - self._zero_observable() - } - - #[inline] - fn empty_observable(&self) -> Self::Observable { - DVector::zeros(0) - } - -} - -/// Implements the calculation a factor $L$ such that $A_*A ≤ L 𝒟$ for $A$ the forward model -/// and $𝒟$ a seminorm of suitable form. -/// -/// **This assumes (but does not check) that the sensors are not overlapping.** -#[replace_float_literals(F::cast_from(literal))] -impl<F, BT, S, P, K, const N : usize> Lipschitz<ConvolutionOp<F, K, BT, N>> -for SensorGrid<F, S, P, BT, N> -where F : Float + nalgebra::RealField + ToNalgebraRealField, - BT : SensorGridBT<F, S, P, N>, - S : Sensor<F, N>, - P : Spread<F, N>, - Convolution<S, P> : Spread<F, N>, - K : SimpleConvolutionKernel<F, N>, - AutoConvolution<P> : BoundedBy<F, K> { - - type FloatType = F; - - fn lipschitz_factor(&self, seminorm : &ConvolutionOp<F, K, BT, N>) -> Option<F> { - // Sensors should not take on negative values to allow - // A_*A to be upper bounded by a simple convolution of `spread`. - if self.sensor.bounds().lower() < 0.0 { - return None - } - - // Calculate the factor $L_1$ for betwee $ℱ[ψ * ψ] ≤ L_1 ℱ[ρ]$ for $ψ$ the base spread - // and $ρ$ the kernel of the seminorm. - let l1 = AutoConvolution(self.spread.clone()).bounding_factor(seminorm.kernel())?; - - // Calculate the factor for transitioning from $A_*A$ to `AutoConvolution<P>`, where A - // consists of several `Convolution<S, P>` for the physical model `P` and the sensor `S`. - let l0 = self.sensor.norm(Linfinity) * self.sensor.norm(L1); +/// Trait for [`ForwardModel`]s that satisfy bounds on curvature. +pub trait BoundedCurvature { + type FloatType: Float; - // The final transition factor is: - Some(l0 * l1) - } + /// Returns factor $ℓ_F$ and $ℓ_r$ such that + /// $B_{F'(μ)} dγ ≤ ℓ_F c_2$ and $⟨F'(μ)+F'(μ+Δ)|Δ⟩ ≤ ℓ_r|γ|(c_2)$, + /// where $Δ=(π_♯^1-π_♯^0)γ$. + fn curvature_bound_components(&self) -> (Option<Self::FloatType>, Option<Self::FloatType>); } - -macro_rules! make_sensorgridsupportgenerator_scalarop_rhs { - ($trait:ident, $fn:ident, $trait_assign:ident, $fn_assign:ident) => { - impl<F, S, P, const N : usize> - std::ops::$trait_assign<F> - for SensorGridSupportGenerator<F, S, P, N> - where F : Float, - S : Sensor<F, N>, - P : Spread<F, N>, - Convolution<S, P> : Spread<F, N> { - fn $fn_assign(&mut self, t : F) { - self.weights.$fn_assign(t); - } - } - - impl<F, S, P, const N : usize> - std::ops::$trait<F> - for SensorGridSupportGenerator<F, S, P, N> - where F : Float, - S : Sensor<F, N>, - P : Spread<F, N>, - Convolution<S, P> : Spread<F, N> { - type Output = SensorGridSupportGenerator<F, S, P, N>; - fn $fn(mut self, t : F) -> Self::Output { - std::ops::$trait_assign::$fn_assign(&mut self.weights, t); - self - } - } - - impl<'a, F, S, P, const N : usize> - std::ops::$trait<F> - for &'a SensorGridSupportGenerator<F, S, P, N> - where F : Float, - S : Sensor<F, N>, - P : Spread<F, N>, - Convolution<S, P> : Spread<F, N> { - type Output = SensorGridSupportGenerator<F, S, P, N>; - fn $fn(self, t : F) -> Self::Output { - SensorGridSupportGenerator{ - base_sensor : self.base_sensor.clone(), - grid : self.grid, - weights : (&self.weights).$fn(t) - } - } - } - } -} - -make_sensorgridsupportgenerator_scalarop_rhs!(Mul, mul, MulAssign, mul_assign); -make_sensorgridsupportgenerator_scalarop_rhs!(Div, div, DivAssign, div_assign); - -macro_rules! make_sensorgridsupportgenerator_unaryop { - ($trait:ident, $fn:ident) => { - impl<F, S, P, const N : usize> - std::ops::$trait - for SensorGridSupportGenerator<F, S, P, N> - where F : Float, - S : Sensor<F, N>, - P : Spread<F, N>, - Convolution<S, P> : Spread<F, N> { - type Output = SensorGridSupportGenerator<F, S, P, N>; - fn $fn(mut self) -> Self::Output { - self.weights = self.weights.$fn(); - self - } - } - - impl<'a, F, S, P, const N : usize> - std::ops::$trait - for &'a SensorGridSupportGenerator<F, S, P, N> - where F : Float, - S : Sensor<F, N>, - P : Spread<F, N>, - Convolution<S, P> : Spread<F, N> { - type Output = SensorGridSupportGenerator<F, S, P, N>; - fn $fn(self) -> Self::Output { - SensorGridSupportGenerator{ - base_sensor : self.base_sensor.clone(), - grid : self.grid, - weights : (&self.weights).$fn() - } - } - } - } -} - -make_sensorgridsupportgenerator_unaryop!(Neg, neg);
--- /dev/null Thu Jan 01 00:00:00 1970 +0000 +++ b/src/forward_model/bias.rs Mon Feb 17 13:54:53 2025 -0500 @@ -0,0 +1,108 @@ +/*! +Simple parametric forward model. + */ + +use super::{AdjointProductBoundedBy, AdjointProductPairBoundedBy, BoundedCurvature, ForwardModel}; +use crate::measures::RNDM; +use alg_tools::direct_product::Pair; +use alg_tools::error::DynError; +use alg_tools::linops::{IdOp, Linear, RowOp, ZeroOp, AXPY}; +use alg_tools::mapping::Space; +use alg_tools::norms::{Norm, NormExponent, PairNorm, L2}; +use alg_tools::types::{ClosedAdd, Float}; +use numeric_literals::replace_float_literals; + +impl<Domain, F, A, E> ForwardModel<Pair<Domain, A::Observable>, F, PairNorm<E, L2, L2>> + for RowOp<A, IdOp<A::Observable>> +where + E: NormExponent, + Domain: Space + Norm<F, E>, + F: Float, + A::Observable: ClosedAdd + Norm<F, L2> + 'static, + A: ForwardModel<Domain, F, E> + 'static, +{ + type Observable = A::Observable; + + fn write_observable(&self, b: &Self::Observable, prefix: String) -> DynError { + self.0.write_observable(b, prefix) + } + + /// Returns a zero observable + fn zero_observable(&self) -> Self::Observable { + self.0.zero_observable() + } +} + +#[replace_float_literals(F::cast_from(literal))] +impl<Domain, F, A, D, Z> AdjointProductPairBoundedBy<Pair<Domain, Z>, D, IdOp<Z>> + for RowOp<A, IdOp<Z>> +where + Domain: Space, + F: Float, + Z: Clone + Space + ClosedAdd, + A: AdjointProductBoundedBy<Domain, D, FloatType = F, Codomain = Z>, + A::Codomain: ClosedAdd, +{ + type FloatType = F; + + fn adjoint_product_pair_bound(&self, d: &D, _: &IdOp<Z>) -> Option<(F, F)> { + self.0.adjoint_product_bound(d).map(|l_0| { + // [A_*; B_*][A, B] = [A_*A, A_* B; B_* A, B_* B] ≤ diag(2A_*A, 2B_*B) + // ≤ diag(2l_A𝒟_A, 2l_B𝒟_B), where now 𝒟_B=Id and l_B=1. + (2.0 * l_0, 2.0) + }) + } +} + +/* +/// This `impl` is bit of an abuse as the codomain of `Apre` is a [`Pair`] of a measure predual, +/// to which this `impl` applies, and another space. +impl<F, Apre, Z> LipschitzValues +for ColOp<Apre, IdOp<Z>> +where + F : Float, + Z : Clone + Space + ClosedAdd, + Apre : LipschitzValues<FloatType = F>, +{ + type FloatType = F; + /// Return (if one exists) a factor $L$ such that $A_*z$ is $L$-Lipschitz for all + /// $z$ in the unit ball. + fn value_unit_lipschitz_factor(&self) -> Option<Self::FloatType> { + self.0.value_unit_lipschitz_factor() + } + + /// Return (if one exists) a factor $L$ such that $∇A_*z$ is $L$-Lipschitz for all + /// $z$ in the unit ball. + fn value_diff_unit_lipschitz_factor(&self) -> Option<Self::FloatType> { + self.0.value_diff_unit_lipschitz_factor() + } +} +*/ + +impl<F, A, Z> BoundedCurvature for RowOp<A, IdOp<Z>> +where + F: Float, + Z: Clone + Space + ClosedAdd, + A: BoundedCurvature<FloatType = F>, +{ + type FloatType = F; + + fn curvature_bound_components(&self) -> (Option<Self::FloatType>, Option<Self::FloatType>) { + self.0.curvature_bound_components() + } +} + +#[replace_float_literals(F::cast_from(literal))] +impl<'a, F, D, XD, Y, const N: usize> AdjointProductBoundedBy<RNDM<F, N>, D> + for ZeroOp<'a, RNDM<F, N>, XD, Y, F> +where + F: Float, + Y: AXPY<F> + Clone, + D: Linear<RNDM<F, N>>, +{ + type FloatType = F; + /// Return $L$ such that $A_*A ≤ L𝒟$ is bounded by some `other` operator $𝒟$. + fn adjoint_product_bound(&self, _: &D) -> Option<F> { + Some(0.0) + } +}
--- /dev/null Thu Jan 01 00:00:00 1970 +0000 +++ b/src/forward_model/sensor_grid.rs Mon Feb 17 13:54:53 2025 -0500 @@ -0,0 +1,645 @@ +/*! +Sensor grid forward model +*/ + +use nalgebra::base::{DMatrix, DVector}; +use numeric_literals::replace_float_literals; +use std::iter::Zip; +use std::ops::RangeFrom; + +use alg_tools::bisection_tree::*; +use alg_tools::error::DynError; +use alg_tools::instance::Instance; +use alg_tools::iter::{MapX, Mappable}; +use alg_tools::lingrid::*; +pub use alg_tools::linops::*; +use alg_tools::mapping::{DifferentiableMapping, RealMapping}; +use alg_tools::maputil::map2; +use alg_tools::nalgebra_support::ToNalgebraRealField; +use alg_tools::norms::{Linfinity, Norm, L1, L2}; +use alg_tools::tabledump::write_csv; + +use super::{AdjointProductBoundedBy, BoundedCurvature, ForwardModel}; +use crate::frank_wolfe::FindimQuadraticModel; +use crate::kernels::{AutoConvolution, BoundedBy, Convolution}; +use crate::measures::{DiscreteMeasure, Radon}; +use crate::preadjoint_helper::PreadjointHelper; +use crate::seminorms::{ConvolutionOp, SimpleConvolutionKernel}; +use crate::types::*; + +type RNDM<F, const N: usize> = DiscreteMeasure<Loc<F, N>, F>; + +pub type ShiftedSensor<F, S, P, const N: usize> = Shift<Convolution<S, P>, F, N>; + +/// Trait for physical convolution models. Has blanket implementation for all cases. +pub trait Spread<F: Float, const N: usize>: + 'static + Clone + Support<F, N> + RealMapping<F, N> + Bounded<F> +{ +} + +impl<F, T, const N: usize> Spread<F, N> for T +where + F: Float, + T: 'static + Clone + Support<F, N> + Bounded<F> + RealMapping<F, N>, +{ +} + +/// Trait for compactly supported sensors. Has blanket implementation for all cases. +pub trait Sensor<F: Float, const N: usize>: + Spread<F, N> + Norm<F, L1> + Norm<F, Linfinity> +{ +} + +impl<F, T, const N: usize> Sensor<F, N> for T +where + F: Float, + T: Spread<F, N> + Norm<F, L1> + Norm<F, Linfinity>, +{ +} + +pub trait SensorGridBT<F, S, P, const N: usize>: + Clone + BTImpl<F, N, Data = usize, Agg = Bounds<F>> +where + F: Float, + S: Sensor<F, N>, + P: Spread<F, N>, +{ +} + +impl<F, S, P, T, const N: usize> SensorGridBT<F, S, P, N> for T +where + T: Clone + BTImpl<F, N, Data = usize, Agg = Bounds<F>>, + F: Float, + S: Sensor<F, N>, + P: Spread<F, N>, +{ +} + +// We need type alias bounds to access associated types +#[allow(type_alias_bounds)] +pub type SensorGridBTFN<F, S, P, BT: SensorGridBT<F, S, P, N>, const N: usize> = + BTFN<F, SensorGridSupportGenerator<F, S, P, N>, BT, N>; + +/// Sensor grid forward model +#[derive(Clone)] +pub struct SensorGrid<F, S, P, BT, const N: usize> +where + F: Float, + S: Sensor<F, N>, + P: Spread<F, N>, + Convolution<S, P>: Spread<F, N>, + BT: SensorGridBT<F, S, P, N>, +{ + domain: Cube<F, N>, + sensor_count: [usize; N], + sensor: S, + spread: P, + base_sensor: Convolution<S, P>, + bt: BT, +} + +impl<F, S, P, BT, const N: usize> SensorGrid<F, S, P, BT, N> +where + F: Float, + BT: SensorGridBT<F, S, P, N>, + S: Sensor<F, N>, + P: Spread<F, N>, + Convolution<S, P>: Spread<F, N> + LocalAnalysis<F, BT::Agg, N>, +{ + /// Create a new sensor grid. + /// + /// The parameter `depth` indicates the search depth of the created [`BT`]s + /// for the adjoint values. + pub fn new( + domain: Cube<F, N>, + sensor_count: [usize; N], + sensor: S, + spread: P, + depth: BT::Depth, + ) -> Self { + let base_sensor = Convolution(sensor.clone(), spread.clone()); + let bt = BT::new(domain, depth); + let mut sensorgrid = SensorGrid { + domain, + sensor_count, + sensor, + spread, + base_sensor, + bt, + }; + + for (x, id) in sensorgrid.grid().into_iter().zip(0usize..) { + let s = sensorgrid.shifted_sensor(x); + sensorgrid.bt.insert(id, &s); + } + + sensorgrid + } +} + +impl<F, S, P, BT, const N: usize> SensorGrid<F, S, P, BT, N> +where + F: Float, + BT: SensorGridBT<F, S, P, N>, + S: Sensor<F, N>, + P: Spread<F, N>, + Convolution<S, P>: Spread<F, N>, +{ + /// Return the grid of sensor locations. + pub fn grid(&self) -> LinGrid<F, N> { + lingrid_centered(&self.domain, &self.sensor_count) + } + + /// Returns the number of sensors (number of grid points) + pub fn n_sensors(&self) -> usize { + self.sensor_count.iter().product() + } + + /// Constructs a sensor shifted by `x`. + #[inline] + fn shifted_sensor(&self, x: Loc<F, N>) -> ShiftedSensor<F, S, P, N> { + self.base_sensor.clone().shift(x) + } + + #[inline] + fn _zero_observable(&self) -> DVector<F> { + DVector::zeros(self.n_sensors()) + } + + /// Returns the maximum number of overlapping sensors $N_\psi$. + pub fn max_overlapping(&self) -> F { + let w = self.base_sensor.support_hint().width(); + let d = map2(self.domain.width(), &self.sensor_count, |wi, &i| { + wi / F::cast_from(i) + }); + w.iter() + .zip(d.iter()) + .map(|(&wi, &di)| (wi / di).ceil()) + .reduce(F::mul) + .unwrap() + } +} + +impl<F, S, P, BT, const N: usize> Mapping<RNDM<F, N>> for SensorGrid<F, S, P, BT, N> +where + F: Float, + BT: SensorGridBT<F, S, P, N>, + S: Sensor<F, N>, + P: Spread<F, N>, + Convolution<S, P>: Spread<F, N>, +{ + type Codomain = DVector<F>; + + #[inline] + fn apply<I: Instance<RNDM<F, N>>>(&self, μ: I) -> DVector<F> { + let mut y = self._zero_observable(); + self.apply_add(&mut y, μ); + y + } +} + +impl<F, S, P, BT, const N: usize> Linear<RNDM<F, N>> for SensorGrid<F, S, P, BT, N> +where + F: Float, + BT: SensorGridBT<F, S, P, N>, + S: Sensor<F, N>, + P: Spread<F, N>, + Convolution<S, P>: Spread<F, N>, +{ +} + +#[replace_float_literals(F::cast_from(literal))] +impl<F, S, P, BT, const N: usize> GEMV<F, RNDM<F, N>, DVector<F>> for SensorGrid<F, S, P, BT, N> +where + F: Float, + BT: SensorGridBT<F, S, P, N>, + S: Sensor<F, N>, + P: Spread<F, N>, + Convolution<S, P>: Spread<F, N>, +{ + fn gemv<I: Instance<RNDM<F, N>>>(&self, y: &mut DVector<F>, α: F, μ: I, β: F) { + let grid = self.grid(); + if β == 0.0 { + y.fill(0.0) + } else if β != 1.0 { + *y *= β; // Need to multiply first, as we have to be able to add to y. + } + if α == 1.0 { + self.apply_add(y, μ) + } else { + for δ in μ.ref_instance() { + for &d in self.bt.iter_at(&δ.x) { + let sensor = self.shifted_sensor(grid.entry_linear_unchecked(d)); + y[d] += sensor.apply(&δ.x) * (α * δ.α); + } + } + } + } + + fn apply_add<I: Instance<RNDM<F, N>>>(&self, y: &mut DVector<F>, μ: I) { + let grid = self.grid(); + for δ in μ.ref_instance() { + for &d in self.bt.iter_at(&δ.x) { + let sensor = self.shifted_sensor(grid.entry_linear_unchecked(d)); + y[d] += sensor.apply(&δ.x) * δ.α; + } + } + } +} + +impl<F, S, P, BT, const N: usize> BoundedLinear<RNDM<F, N>, Radon, L2, F> + for SensorGrid<F, S, P, BT, N> +where + F: Float, + BT: SensorGridBT<F, S, P, N, Agg = Bounds<F>>, + S: Sensor<F, N>, + P: Spread<F, N>, + Convolution<S, P>: Spread<F, N> + LocalAnalysis<F, BT::Agg, N>, +{ + /// An estimate on the operator norm in $𝕃(ℳ(Ω); ℝ^n)$ with $ℳ(Ω)$ equipped + /// with the Radon norm, and $ℝ^n$ with the Euclidean norm. + fn opnorm_bound(&self, _: Radon, _: L2) -> F { + // With {x_i}_{i=1}^n the grid centres and φ the kernel, we have + // |Aμ|_2 = sup_{|z|_2 ≤ 1} ⟨z,Αμ⟩ = sup_{|z|_2 ≤ 1} ⟨A^*z|μ⟩ + // ≤ sup_{|z|_2 ≤ 1} |A^*z|_∞ |μ|_ℳ + // = sup_{|z|_2 ≤ 1} |∑ φ(· - x_i)z_i|_∞ |μ|_ℳ + // ≤ sup_{|z|_2 ≤ 1} |φ(y)| ∑_{i:th sensor active at y}|z_i| |μ|_ℳ + // where the supremum of |∑ φ(· - x_i)z_i|_∞ is reached at y + // ≤ sup_{|z|_2 ≤ 1} |φ|_∞ √N_ψ |z|_2 |μ|_ℳ + // where N_ψ is the maximum number of sensors that overlap, and + // |z|_2 is restricted to the active sensors. + // = |φ|_∞ √N_ψ |μ|_ℳ. + // Hence + let n = self.max_overlapping(); + self.base_sensor.bounds().uniform() * n.sqrt() + } +} + +type SensorGridPreadjoint<'a, A, F, const N: usize> = PreadjointHelper<'a, A, RNDM<F, N>>; + +impl<F, S, P, BT, const N: usize> Preadjointable<RNDM<F, N>, DVector<F>> + for SensorGrid<F, S, P, BT, N> +where + F: Float, + BT: SensorGridBT<F, S, P, N>, + S: Sensor<F, N>, + P: Spread<F, N>, + Convolution<S, P>: Spread<F, N> + LocalAnalysis<F, BT::Agg, N>, +{ + type PreadjointCodomain = BTFN<F, SensorGridSupportGenerator<F, S, P, N>, BT, N>; + type Preadjoint<'a> + = SensorGridPreadjoint<'a, Self, F, N> + where + Self: 'a; + + fn preadjoint(&self) -> Self::Preadjoint<'_> { + PreadjointHelper::new(self) + } +} + +/* +#[replace_float_literals(F::cast_from(literal))] +impl<'a, F, S, P, BT, const N : usize> LipschitzValues +for SensorGridPreadjoint<'a, SensorGrid<F, S, P, BT, N>, F, N> +where F : Float, + BT : SensorGridBT<F, S, P, N>, + S : Sensor<F, N>, + P : Spread<F, N>, + Convolution<S, P> : Spread<F, N> + Lipschitz<L2, FloatType=F> + DifferentiableMapping<Loc<F,N>> + LocalAnalysis<F, BT::Agg, N>, + for<'b> <Convolution<S, P> as DifferentiableMapping<Loc<F,N>>>::Differential<'b> : Lipschitz<L2, FloatType=F>, +{ + + type FloatType = F; + + fn value_unit_lipschitz_factor(&self) -> Option<F> { + // The Lipschitz factor of the sensors has to be scaled by the square root of twice + // the number of overlapping sensors at a single ponit, as Lipschitz estimates involve + // two points. + let fw = self.forward_op; + let n = fw.max_overlapping(); + fw.base_sensor.lipschitz_factor(L2).map(|l| (2.0 * n).sqrt() * l) + } + + fn value_diff_unit_lipschitz_factor(&self) -> Option<F> { + // The Lipschitz factor of the sensors has to be scaled by the square root of twice + // the number of overlapping sensors at a single ponit, as Lipschitz estimates involve + // two points. + let fw = self.forward_op; + let n = fw.max_overlapping(); + fw.base_sensor.diff_ref().lipschitz_factor(L2).map(|l| (2.0 * n).sqrt() * l) + } +} +*/ + +#[replace_float_literals(F::cast_from(literal))] +impl<'a, F, S, P, BT, const N: usize> BoundedCurvature for SensorGrid<F, S, P, BT, N> +where + F: Float, + BT: SensorGridBT<F, S, P, N>, + S: Sensor<F, N>, + P: Spread<F, N>, + Convolution<S, P>: Spread<F, N> + + Lipschitz<L2, FloatType = F> + + DifferentiableMapping<Loc<F, N>> + + LocalAnalysis<F, BT::Agg, N>, + for<'b> <Convolution<S, P> as DifferentiableMapping<Loc<F, N>>>::Differential<'b>: + Lipschitz<L2, FloatType = F>, +{ + type FloatType = F; + + /// Returns factors $ℓ_F$ and $Θ²$ such that + /// $B_{F'(μ)} dγ ≤ ℓ_F c_2$ and $⟨F'(μ)+F'(μ+Δ)|Δ⟩ ≤ Θ²|γ|(c_2)‖γ‖$, + /// where $Δ=(π_♯^1-π_♯^0)γ$. + /// + /// See Lemma 3.8, Lemma 5.10, Remark 5.14, and Example 5.15. + fn curvature_bound_components(&self) -> (Option<Self::FloatType>, Option<Self::FloatType>) { + let n_ψ = self.max_overlapping(); + let ψ_diff_lip = self.base_sensor.diff_ref().lipschitz_factor(L2); + let ψ_lip = self.base_sensor.lipschitz_factor(L2); + let ℓ_F = ψ_diff_lip.map(|l| (2.0 * n_ψ).sqrt() * l); + let θ2 = ψ_lip.map(|l| 4.0 * n_ψ * l.powi(2)); + + (ℓ_F, θ2) + } +} + +#[derive(Clone, Debug)] +pub struct SensorGridSupportGenerator<F, S, P, const N: usize> +where + F: Float, + S: Sensor<F, N>, + P: Spread<F, N>, +{ + base_sensor: Convolution<S, P>, + grid: LinGrid<F, N>, + weights: DVector<F>, +} + +impl<F, S, P, const N: usize> SensorGridSupportGenerator<F, S, P, N> +where + F: Float, + S: Sensor<F, N>, + P: Spread<F, N>, + Convolution<S, P>: Spread<F, N>, +{ + #[inline] + fn construct_sensor(&self, id: usize, w: F) -> Weighted<ShiftedSensor<F, S, P, N>, F> { + let x = self.grid.entry_linear_unchecked(id); + self.base_sensor.clone().shift(x).weigh(w) + } + + #[inline] + fn construct_sensor_and_id<'a>( + &'a self, + (id, w): (usize, &'a F), + ) -> (usize, Weighted<ShiftedSensor<F, S, P, N>, F>) { + (id.into(), self.construct_sensor(id, *w)) + } +} + +impl<F, S, P, const N: usize> SupportGenerator<F, N> for SensorGridSupportGenerator<F, S, P, N> +where + F: Float, + S: Sensor<F, N>, + P: Spread<F, N>, + Convolution<S, P>: Spread<F, N>, +{ + type Id = usize; + type SupportType = Weighted<ShiftedSensor<F, S, P, N>, F>; + type AllDataIter<'a> + = MapX< + 'a, + Zip<RangeFrom<usize>, std::slice::Iter<'a, F>>, + Self, + (Self::Id, Self::SupportType), + > + where + Self: 'a; + + #[inline] + fn support_for(&self, d: Self::Id) -> Self::SupportType { + self.construct_sensor(d, self.weights[d]) + } + + #[inline] + fn support_count(&self) -> usize { + self.weights.len() + } + + #[inline] + fn all_data(&self) -> Self::AllDataIter<'_> { + (0..) + .zip(self.weights.as_slice().iter()) + .mapX(self, Self::construct_sensor_and_id) + } +} + +impl<F, S, P, BT, const N: usize> ForwardModel<DiscreteMeasure<Loc<F, N>, F>, F> + for SensorGrid<F, S, P, BT, N> +where + F: Float + ToNalgebraRealField<MixedType = F> + nalgebra::RealField, + BT: SensorGridBT<F, S, P, N>, + S: Sensor<F, N>, + P: Spread<F, N>, + Convolution<S, P>: Spread<F, N> + LocalAnalysis<F, BT::Agg, N>, +{ + type Observable = DVector<F>; + + fn write_observable(&self, b: &Self::Observable, prefix: String) -> DynError { + let it = self.grid().into_iter().zip(b.iter()).map(|(x, &v)| (x, v)); + write_csv(it, prefix + ".txt") + } + + #[inline] + fn zero_observable(&self) -> Self::Observable { + self._zero_observable() + } +} + +impl<F, S, P, BT, const N: usize> FindimQuadraticModel<Loc<F, N>, F> for SensorGrid<F, S, P, BT, N> +where + F: Float + ToNalgebraRealField<MixedType = F> + nalgebra::RealField, + BT: SensorGridBT<F, S, P, N>, + S: Sensor<F, N>, + P: Spread<F, N>, + Convolution<S, P>: Spread<F, N> + LocalAnalysis<F, BT::Agg, N>, +{ + fn findim_quadratic_model( + &self, + μ: &DiscreteMeasure<Loc<F, N>, F>, + b: &Self::Observable, + ) -> (DMatrix<F::MixedType>, DVector<F::MixedType>) { + assert_eq!(b.len(), self.n_sensors()); + let mut mA = DMatrix::zeros(self.n_sensors(), μ.len()); + let grid = self.grid(); + for (mut mAcol, δ) in mA.column_iter_mut().zip(μ.iter_spikes()) { + for &d in self.bt.iter_at(&δ.x) { + let sensor = self.shifted_sensor(grid.entry_linear_unchecked(d)); + mAcol[d] += sensor.apply(&δ.x); + } + } + let mAt = mA.transpose(); + (&mAt * mA, &mAt * b) + } +} + +/// Implements the calculation a factor $L$ such that $A_*A ≤ L 𝒟$ for $A$ the forward model +/// and $𝒟$ a seminorm of suitable form. +/// +/// **This assumes (but does not check) that the sensors are not overlapping.** +#[replace_float_literals(F::cast_from(literal))] +impl<F, BT, S, P, K, const N: usize> AdjointProductBoundedBy<RNDM<F, N>, ConvolutionOp<F, K, BT, N>> + for SensorGrid<F, S, P, BT, N> +where + F: Float + nalgebra::RealField + ToNalgebraRealField, + BT: SensorGridBT<F, S, P, N>, + S: Sensor<F, N>, + P: Spread<F, N>, + Convolution<S, P>: Spread<F, N>, + K: SimpleConvolutionKernel<F, N>, + AutoConvolution<P>: BoundedBy<F, K>, +{ + type FloatType = F; + + fn adjoint_product_bound(&self, seminorm: &ConvolutionOp<F, K, BT, N>) -> Option<F> { + // Sensors should not take on negative values to allow + // A_*A to be upper bounded by a simple convolution of `spread`. + if self.sensor.bounds().lower() < 0.0 { + return None; + } + + // Calculate the factor $L_1$ for betwee $ℱ[ψ * ψ] ≤ L_1 ℱ[ρ]$ for $ψ$ the base spread + // and $ρ$ the kernel of the seminorm. + let l1 = AutoConvolution(self.spread.clone()).bounding_factor(seminorm.kernel())?; + + // Calculate the factor for transitioning from $A_*A$ to `AutoConvolution<P>`, where A + // consists of several `Convolution<S, P>` for the physical model `P` and the sensor `S`. + let l0 = self.sensor.norm(Linfinity) * self.sensor.norm(L1); + + // The final transition factor is: + Some(l0 * l1) + } +} + +macro_rules! make_sensorgridsupportgenerator_scalarop_rhs { + ($trait:ident, $fn:ident, $trait_assign:ident, $fn_assign:ident) => { + impl<F, S, P, const N: usize> std::ops::$trait_assign<F> + for SensorGridSupportGenerator<F, S, P, N> + where + F: Float, + S: Sensor<F, N>, + P: Spread<F, N>, + Convolution<S, P>: Spread<F, N>, + { + fn $fn_assign(&mut self, t: F) { + self.weights.$fn_assign(t); + } + } + + impl<F, S, P, const N: usize> std::ops::$trait<F> for SensorGridSupportGenerator<F, S, P, N> + where + F: Float, + S: Sensor<F, N>, + P: Spread<F, N>, + Convolution<S, P>: Spread<F, N>, + { + type Output = SensorGridSupportGenerator<F, S, P, N>; + fn $fn(mut self, t: F) -> Self::Output { + std::ops::$trait_assign::$fn_assign(&mut self.weights, t); + self + } + } + + impl<'a, F, S, P, const N: usize> std::ops::$trait<F> + for &'a SensorGridSupportGenerator<F, S, P, N> + where + F: Float, + S: Sensor<F, N>, + P: Spread<F, N>, + Convolution<S, P>: Spread<F, N>, + { + type Output = SensorGridSupportGenerator<F, S, P, N>; + fn $fn(self, t: F) -> Self::Output { + SensorGridSupportGenerator { + base_sensor: self.base_sensor.clone(), + grid: self.grid, + weights: (&self.weights).$fn(t), + } + } + } + }; +} + +make_sensorgridsupportgenerator_scalarop_rhs!(Mul, mul, MulAssign, mul_assign); +make_sensorgridsupportgenerator_scalarop_rhs!(Div, div, DivAssign, div_assign); + +macro_rules! make_sensorgridsupportgenerator_unaryop { + ($trait:ident, $fn:ident) => { + impl<F, S, P, const N: usize> std::ops::$trait for SensorGridSupportGenerator<F, S, P, N> + where + F: Float, + S: Sensor<F, N>, + P: Spread<F, N>, + Convolution<S, P>: Spread<F, N>, + { + type Output = SensorGridSupportGenerator<F, S, P, N>; + fn $fn(mut self) -> Self::Output { + self.weights = self.weights.$fn(); + self + } + } + + impl<'a, F, S, P, const N: usize> std::ops::$trait + for &'a SensorGridSupportGenerator<F, S, P, N> + where + F: Float, + S: Sensor<F, N>, + P: Spread<F, N>, + Convolution<S, P>: Spread<F, N>, + { + type Output = SensorGridSupportGenerator<F, S, P, N>; + fn $fn(self) -> Self::Output { + SensorGridSupportGenerator { + base_sensor: self.base_sensor.clone(), + grid: self.grid, + weights: (&self.weights).$fn(), + } + } + } + }; +} + +make_sensorgridsupportgenerator_unaryop!(Neg, neg); + +impl<'a, F, S, P, BT, const N: usize> Mapping<DVector<F>> + for PreadjointHelper<'a, SensorGrid<F, S, P, BT, N>, RNDM<F, N>> +where + F: Float, + BT: SensorGridBT<F, S, P, N>, + S: Sensor<F, N>, + P: Spread<F, N>, + Convolution<S, P>: Spread<F, N> + LocalAnalysis<F, Bounds<F>, N>, +{ + type Codomain = SensorGridBTFN<F, S, P, BT, N>; + + fn apply<I: Instance<DVector<F>>>(&self, x: I) -> Self::Codomain { + let fwd = &self.forward_op; + let generator = SensorGridSupportGenerator { + base_sensor: fwd.base_sensor.clone(), + grid: fwd.grid(), + weights: x.own(), + }; + BTFN::new_refresh(&fwd.bt, generator) + } +} + +impl<'a, F, S, P, BT, const N: usize> Linear<DVector<F>> + for PreadjointHelper<'a, SensorGrid<F, S, P, BT, N>, RNDM<F, N>> +where + F: Float, + BT: SensorGridBT<F, S, P, N>, + S: Sensor<F, N>, + P: Spread<F, N>, + Convolution<S, P>: Spread<F, N> + LocalAnalysis<F, Bounds<F>, N>, +{ +}
--- /dev/null Thu Jan 01 00:00:00 1970 +0000 +++ b/src/forward_pdps.rs Mon Feb 17 13:54:53 2025 -0500 @@ -0,0 +1,240 @@ +/*! +Solver for the point source localisation problem using a +primal-dual proximal splitting with a forward step. +*/ + +use numeric_literals::replace_float_literals; +use serde::{Serialize, Deserialize}; + +use alg_tools::iterate::AlgIteratorFactory; +use alg_tools::euclidean::Euclidean; +use alg_tools::mapping::{Mapping, DifferentiableRealMapping, Instance}; +use alg_tools::norms::Norm; +use alg_tools::direct_product::Pair; +use alg_tools::nalgebra_support::ToNalgebraRealField; +use alg_tools::linops::{ + BoundedLinear, AXPY, GEMV, Adjointable, IdOp, +}; +use alg_tools::convex::{Conjugable, Prox}; +use alg_tools::norms::{L2, PairNorm}; + +use crate::types::*; +use crate::measures::{DiscreteMeasure, Radon, RNDM}; +use crate::measures::merging::SpikeMerging; +use crate::forward_model::{ + ForwardModel, + AdjointProductPairBoundedBy, +}; +use crate::plot::{ + SeqPlotter, + Plotting, + PlotLookup +}; +use crate::fb::*; +use crate::regularisation::RegTerm; +use crate::dataterm::calculate_residual; + +/// Settings for [`pointsource_forward_pdps_pair`]. +#[derive(Clone, Copy, Eq, PartialEq, Serialize, Deserialize, Debug)] +#[serde(default)] +pub struct ForwardPDPSConfig<F : Float> { + /// Primal step length scaling. + pub τ0 : F, + /// Primal step length scaling. + pub σp0 : F, + /// Dual step length scaling. + pub σd0 : F, + /// Generic parameters + pub insertion : FBGenericConfig<F>, +} + +#[replace_float_literals(F::cast_from(literal))] +impl<F : Float> Default for ForwardPDPSConfig<F> { + fn default() -> Self { + ForwardPDPSConfig { + τ0 : 0.99, + σd0 : 0.05, + σp0 : 0.99, + insertion : Default::default() + } + } +} + +type MeasureZ<F, Z, const N : usize> = Pair<RNDM<F, N>, Z>; + +/// Iteratively solve the pointsource localisation with an additional variable +/// using primal-dual proximal splitting with a forward step. +#[replace_float_literals(F::cast_from(literal))] +pub fn pointsource_forward_pdps_pair< + F, I, A, S, Reg, P, Z, R, Y, /*KOpM, */ KOpZ, H, const N : usize +>( + opA : &A, + b : &A::Observable, + reg : Reg, + prox_penalty : &P, + config : &ForwardPDPSConfig<F>, + iterator : I, + mut plotter : SeqPlotter<F, N>, + //opKμ : KOpM, + opKz : &KOpZ, + fnR : &R, + fnH : &H, + mut z : Z, + mut y : Y, +) -> MeasureZ<F, Z, N> +where + F : Float + ToNalgebraRealField, + I : AlgIteratorFactory<IterInfo<F, N>>, + A : ForwardModel< + MeasureZ<F, Z, N>, + F, + PairNorm<Radon, L2, L2>, + PreadjointCodomain = Pair<S, Z>, + > + + AdjointProductPairBoundedBy<MeasureZ<F, Z, N>, P, IdOp<Z>, FloatType=F>, + S: DifferentiableRealMapping<F, N>, + for<'b> &'b A::Observable : std::ops::Neg<Output=A::Observable> + Instance<A::Observable>, + PlotLookup : Plotting<N>, + RNDM<F, N> : SpikeMerging<F>, + Reg : RegTerm<F, N>, + P : ProxPenalty<F, S, Reg, N>, + KOpZ : BoundedLinear<Z, L2, L2, F, Codomain=Y> + + GEMV<F, Z> + + Adjointable<Z, Y, AdjointCodomain = Z>, + for<'b> KOpZ::Adjoint<'b> : GEMV<F, Y>, + Y : AXPY<F> + Euclidean<F, Output=Y> + Clone + ClosedAdd, + for<'b> &'b Y : Instance<Y>, + Z : AXPY<F, Owned=Z> + Euclidean<F, Output=Z> + Clone + Norm<F, L2>, + for<'b> &'b Z : Instance<Z>, + R : Prox<Z, Codomain=F>, + H : Conjugable<Y, F, Codomain=F>, + for<'b> H::Conjugate<'b> : Prox<Y>, +{ + + // Check parameters + assert!(config.τ0 > 0.0 && + config.τ0 < 1.0 && + config.σp0 > 0.0 && + config.σp0 < 1.0 && + config.σd0 > 0.0 && + config.σp0 * config.σd0 <= 1.0, + "Invalid step length parameters"); + + // Initialise iterates + let mut μ = DiscreteMeasure::new(); + let mut residual = calculate_residual(Pair(&μ, &z), opA, b); + + // Set up parameters + let bigM = 0.0; //opKμ.adjoint_product_bound(prox_penalty).unwrap().sqrt(); + let nKz = opKz.opnorm_bound(L2, L2); + let opIdZ = IdOp::new(); + let (l, l_z) = opA.adjoint_product_pair_bound(prox_penalty, &opIdZ).unwrap(); + // We need to satisfy + // + // τσ_dM(1-σ_p L_z)/(1 - τ L) + [σ_p L_z + σ_pσ_d‖K_z‖^2] < 1 + // ^^^^^^^^^^^^^^^^^^^^^^^^^ + // with 1 > σ_p L_z and 1 > τ L. + // + // To do so, we first solve σ_p and σ_d from standard PDPS step length condition + // ^^^^^ < 1. then we solve τ from the rest. + let σ_d = config.σd0 / nKz; + let σ_p = config.σp0 / (l_z + config.σd0 * nKz); + // Observe that = 1 - ^^^^^^^^^^^^^^^^^^^^^ = 1 - σ_{p,0} + // We get the condition τσ_d M (1-σ_p L_z) < (1-σ_{p,0})*(1-τ L) + // ⟺ τ [ σ_d M (1-σ_p L_z) + (1-σ_{p,0}) L ] < (1-σ_{p,0}) + let φ = 1.0 - config.σp0; + let a = 1.0 - σ_p * l_z; + let τ = config.τ0 * φ / ( σ_d * bigM * a + φ * l ); + // Acceleration is not currently supported + // let γ = dataterm.factor_of_strong_convexity(); + let ω = 1.0; + + // We multiply tolerance by τ for FB since our subproblems depending on tolerances are scaled + // by τ compared to the conditional gradient approach. + let tolerance = config.insertion.tolerance * τ * reg.tolerance_scaling(); + let mut ε = tolerance.initial(); + + let starH = fnH.conjugate(); + + // Statistics + let full_stats = |residual : &A::Observable, μ : &RNDM<F, N>, z : &Z, ε, stats| IterInfo { + value : residual.norm2_squared_div2() + fnR.apply(z) + + reg.apply(μ) + fnH.apply(/* opKμ.apply(μ) + */ opKz.apply(z)), + n_spikes : μ.len(), + ε, + // postprocessing: config.insertion.postprocessing.then(|| μ.clone()), + .. stats + }; + let mut stats = IterInfo::new(); + + // Run the algorithm + for state in iterator.iter_init(|| full_stats(&residual, &μ, &z, ε, stats.clone())) { + // Calculate initial transport + let Pair(mut τv, τz) = opA.preadjoint().apply(residual * τ); + let μ_base = μ.clone(); + + // Construct μ^{k+1} by solving finite-dimensional subproblems and insert new spikes. + let (maybe_d, _within_tolerances) = prox_penalty.insert_and_reweigh( + &mut μ, &mut τv, &μ_base, None, + τ, ε, &config.insertion, + ®, &state, &mut stats, + ); + + // Merge spikes. + // This crucially expects the merge routine to be stable with respect to spike locations, + // and not to performing any pruning. That is be to done below simultaneously for γ. + // Merge spikes. + // This crucially expects the merge routine to be stable with respect to spike locations, + // and not to performing any pruning. That is be to done below simultaneously for γ. + let ins = &config.insertion; + if ins.merge_now(&state) { + stats.merged += prox_penalty.merge_spikes_no_fitness( + &mut μ, &mut τv, &μ_base, None, τ, ε, ins, ®, + //Some(|μ̃ : &RNDM<F, N>| calculate_residual(Pair(μ̃, &z), opA, b).norm2_squared_div2()), + ); + } + + // Prune spikes with zero weight. + stats.pruned += prune_with_stats(&mut μ); + + // Do z variable primal update + let mut z_new = τz; + opKz.adjoint().gemv(&mut z_new, -σ_p, &y, -σ_p/τ); + z_new = fnR.prox(σ_p, z_new + &z); + // Do dual update + // opKμ.gemv(&mut y, σ_d*(1.0 + ω), &μ, 1.0); // y = y + σ_d K[(1+ω)(μ,z)^{k+1}] + opKz.gemv(&mut y, σ_d*(1.0 + ω), &z_new, 1.0); + // opKμ.gemv(&mut y, -σ_d*ω, μ_base, 1.0);// y = y + σ_d K[(1+ω)(μ,z)^{k+1} - ω (μ,z)^k]-b + opKz.gemv(&mut y, -σ_d*ω, z, 1.0);// y = y + σ_d K[(1+ω)(μ,z)^{k+1} - ω (μ,z)^k]-b + y = starH.prox(σ_d, y); + z = z_new; + + // Update residual + residual = calculate_residual(Pair(&μ, &z), opA, b); + + // Update step length parameters + // let ω = pdpsconfig.acceleration.accelerate(&mut τ, &mut σ, γ); + + // Give statistics if requested + let iter = state.iteration(); + stats.this_iters += 1; + + state.if_verbose(|| { + plotter.plot_spikes(iter, maybe_d.as_ref(), Some(&τv), &μ); + full_stats(&residual, &μ, &z, ε, std::mem::replace(&mut stats, IterInfo::new())) + }); + + // Update main tolerance for next iteration + ε = tolerance.update(ε, iter); + } + + let fit = |μ̃ : &RNDM<F, N>| { + (opA.apply(Pair(μ̃, &z))-b).norm2_squared_div2() + //+ fnR.apply(z) + reg.apply(μ) + + fnH.apply(/* opKμ.apply(&μ̃) + */ opKz.apply(&z)) + }; + + μ.merge_spikes_fitness(config.insertion.final_merging_method(), fit, |&v| v); + μ.prune(); + Pair(μ, z) +}
--- a/src/fourier.rs Tue Aug 01 10:25:09 2023 +0300 +++ b/src/fourier.rs Mon Feb 17 13:54:53 2025 -0500 @@ -3,14 +3,14 @@ */ use alg_tools::types::{Num, Float}; -use alg_tools::mapping::{RealMapping, Mapping}; +use alg_tools::mapping::{RealMapping, Mapping, Space}; use alg_tools::bisection_tree::Weighted; use alg_tools::loc::Loc; /// Trait for Fourier transforms. When F is a non-complex number, the transform /// also has to be non-complex, i.e., the function itself symmetric. pub trait Fourier<F : Num> : Mapping<Self::Domain, Codomain=F> { - type Domain; + type Domain : Space; type Transformed : Mapping<Self::Domain, Codomain=F>; fn fourier(&self) -> Self::Transformed;
--- a/src/frank_wolfe.rs Tue Aug 01 10:25:09 2023 +0300 +++ b/src/frank_wolfe.rs Mon Feb 17 13:54:53 2025 -0500 @@ -14,18 +14,18 @@ */ use numeric_literals::replace_float_literals; +use nalgebra::{DMatrix, DVector}; use serde::{Serialize, Deserialize}; //use colored::Colorize; use alg_tools::iterate::{ AlgIteratorFactory, - AlgIteratorState, AlgIteratorOptions, ValueIteratorFactory, }; use alg_tools::euclidean::Euclidean; use alg_tools::norms::Norm; -use alg_tools::linops::Apply; +use alg_tools::linops::Mapping; use alg_tools::sets::Cube; use alg_tools::loc::Loc; use alg_tools::bisection_tree::{ @@ -40,9 +40,11 @@ }; use alg_tools::mapping::RealMapping; use alg_tools::nalgebra_support::ToNalgebraRealField; +use alg_tools::norms::L2; use crate::types::*; use crate::measures::{ + RNDM, DiscreteMeasure, DeltaMeasure, Radon, @@ -68,10 +70,10 @@ use crate::regularisation::{ NonnegRadonRegTerm, RadonRegTerm, + RegTerm }; -use crate::fb::RegTerm; -/// Settings for [`pointsource_fw`]. +/// Settings for [`pointsource_fw_reg`]. #[derive(Clone, Copy, Eq, PartialEq, Serialize, Deserialize, Debug)] #[serde(default)] pub struct FWConfig<F : Float> { @@ -106,15 +108,25 @@ refinement : Default::default(), inner : Default::default(), variant : FWVariant::FullyCorrective, - merging : Default::default(), + merging : SpikeMergingMethod { enabled : true, ..Default::default() }, } } } -/// Helper struct for pre-initialising the finite-dimensional subproblems solver -/// [`prepare_optimise_weights`]. -/// -/// The pre-initialisation is done by [`prepare_optimise_weights`]. +pub trait FindimQuadraticModel<Domain, F> : ForwardModel<DiscreteMeasure<Domain, F>, F> +where + F : Float + ToNalgebraRealField, + Domain : Clone + PartialEq, +{ + /// Return A_*A and A_* b + fn findim_quadratic_model( + &self, + μ : &DiscreteMeasure<Domain, F>, + b : &Self::Observable + ) -> (DMatrix<F::MixedType>, DVector<F::MixedType>); +} + +/// Helper struct for pre-initialising the finite-dimensional subproblem solver. pub struct FindimData<F : Float> { /// ‖A‖^2 opAnorm_squared : F, @@ -125,7 +137,7 @@ /// Trait for finite dimensional weight optimisation. pub trait WeightOptim< F : Float + ToNalgebraRealField, - A : ForwardModel<Loc<F, N>, F>, + A : ForwardModel<RNDM<F, N>, F>, I : AlgIteratorFactory<F>, const N : usize > { @@ -154,7 +166,7 @@ /// Returns the number of iterations taken by the method configured in `inner`. fn optimise_weights<'a>( &self, - μ : &mut DiscreteMeasure<Loc<F, N>, F>, + μ : &mut RNDM<F, N>, opA : &'a A, b : &A::Observable, findim_data : &FindimData<F>, @@ -166,12 +178,12 @@ /// Trait for regularisation terms supported by [`pointsource_fw_reg`]. pub trait RegTermFW< F : Float + ToNalgebraRealField, - A : ForwardModel<Loc<F, N>, F>, + A : ForwardModel<RNDM<F, N>, F>, I : AlgIteratorFactory<F>, const N : usize > : RegTerm<F, N> + WeightOptim<F, A, I, N> - + for<'a> Apply<&'a DiscreteMeasure<Loc<F, N>, F>, Output = F> { + + Mapping<RNDM<F, N>, Codomain = F> { /// With $g = A\_\*(Aμ-b)$, returns $(x, g(x))$ for $x$ a new point to be inserted /// into $μ$, as determined by the regulariser. @@ -188,7 +200,7 @@ /// Insert point `ξ` into `μ` for the relaxed algorithm from Bredies–Pikkarainen. fn relaxed_insert<'a>( &self, - μ : &mut DiscreteMeasure<Loc<F, N>, F>, + μ : &mut RNDM<F, N>, g : &A::PreadjointCodomain, opA : &'a A, ξ : Loc<F, N>, @@ -201,18 +213,18 @@ impl<F : Float + ToNalgebraRealField, A, I, const N : usize> WeightOptim<F, A, I, N> for RadonRegTerm<F> where I : AlgIteratorFactory<F>, - A : ForwardModel<Loc<F, N>, F> { + A : FindimQuadraticModel<Loc<F, N>, F> { fn prepare_optimise_weights(&self, opA : &A, b : &A::Observable) -> FindimData<F> { FindimData{ - opAnorm_squared : opA.opnorm_bound().powi(2), + opAnorm_squared : opA.opnorm_bound(Radon, L2).powi(2), m0 : b.norm2_squared() / (2.0 * self.α()), } } fn optimise_weights<'a>( &self, - μ : &mut DiscreteMeasure<Loc<F, N>, F>, + μ : &mut RNDM<F, N>, opA : &'a A, b : &A::Observable, findim_data : &FindimData<F>, @@ -232,9 +244,9 @@ // = C sup_{‖x‖_1 ≤ 1} ‖Ax‖_2 = C ‖A‖_{1,2}, // where C = √m satisfies ‖x‖_1 ≤ C ‖x‖_2. Since we are intested in ‖A_*A‖, no // square root is needed when we scale: - let inner_τ = inner.τ0 / (findim_data.opAnorm_squared * F::cast_from(μ.len())); - let iters = quadratic_unconstrained(inner.method, &Ã, &g̃, self.α(), - &mut x, inner_τ, iterator); + let normest = findim_data.opAnorm_squared * F::cast_from(μ.len()); + let iters = quadratic_unconstrained(&Ã, &g̃, self.α(), &mut x, + normest, inner, iterator); // Update masses of μ based on solution of finite-dimensional subproblem. μ.set_masses_dvector(&x); @@ -245,12 +257,19 @@ #[replace_float_literals(F::cast_from(literal))] impl<F : Float + ToNalgebraRealField, A, I, S, GA, BTA, const N : usize> RegTermFW<F, A, I, N> for RadonRegTerm<F> -where Cube<F, N> : P2Minimise<Loc<F, N>, F>, - I : AlgIteratorFactory<F>, - S: RealMapping<F, N> + LocalAnalysis<F, Bounds<F>, N>, - GA : SupportGenerator<F, N, SupportType = S, Id = usize> + Clone, - A : ForwardModel<Loc<F, N>, F, PreadjointCodomain = BTFN<F, GA, BTA, N>>, - BTA : BTSearch<F, N, Data=usize, Agg=Bounds<F>> { +where + Cube<F, N> : P2Minimise<Loc<F, N>, F>, + I : AlgIteratorFactory<F>, + S: RealMapping<F, N> + LocalAnalysis<F, Bounds<F>, N>, + GA : SupportGenerator<F, N, SupportType = S, Id = usize> + Clone, + A : FindimQuadraticModel<Loc<F, N>, F, PreadjointCodomain = BTFN<F, GA, BTA, N>>, + BTA : BTSearch<F, N, Data=usize, Agg=Bounds<F>>, + // FIXME: the following *should not* be needed, they are already implied + RNDM<F, N> : Mapping<A::PreadjointCodomain, Codomain = F>, + DeltaMeasure<Loc<F, N>, F> : Mapping<A::PreadjointCodomain, Codomain = F>, + //A : Mapping<RNDM<F, N>, Codomain = A::Observable>, + //A : Mapping<DeltaMeasure<Loc<F, N>, F>, Codomain = A::Observable>, +{ fn find_insertion( &self, @@ -269,7 +288,7 @@ fn relaxed_insert<'a>( &self, - μ : &mut DiscreteMeasure<Loc<F, N>, F>, + μ : &mut RNDM<F, N>, g : &A::PreadjointCodomain, opA : &'a A, ξ : Loc<F, N>, @@ -282,7 +301,7 @@ let v = if v_ξ.abs() <= α { 0.0 } else { m0 / α * v_ξ }; let δ = DeltaMeasure { x : ξ, α : v }; let dp = μ.apply(g) - δ.apply(g); - let d = opA.apply(&*μ) - opA.apply(&δ); + let d = opA.apply(&*μ) - opA.apply(δ); let r = d.norm2_squared(); let s = if r == 0.0 { 1.0 @@ -298,18 +317,18 @@ impl<F : Float + ToNalgebraRealField, A, I, const N : usize> WeightOptim<F, A, I, N> for NonnegRadonRegTerm<F> where I : AlgIteratorFactory<F>, - A : ForwardModel<Loc<F, N>, F> { + A : FindimQuadraticModel<Loc<F, N>, F> { fn prepare_optimise_weights(&self, opA : &A, b : &A::Observable) -> FindimData<F> { FindimData{ - opAnorm_squared : opA.opnorm_bound().powi(2), + opAnorm_squared : opA.opnorm_bound(Radon, L2).powi(2), m0 : b.norm2_squared() / (2.0 * self.α()), } } fn optimise_weights<'a>( &self, - μ : &mut DiscreteMeasure<Loc<F, N>, F>, + μ : &mut RNDM<F, N>, opA : &'a A, b : &A::Observable, findim_data : &FindimData<F>, @@ -329,9 +348,9 @@ // = C sup_{‖x‖_1 ≤ 1} ‖Ax‖_2 = C ‖A‖_{1,2}, // where C = √m satisfies ‖x‖_1 ≤ C ‖x‖_2. Since we are intested in ‖A_*A‖, no // square root is needed when we scale: - let inner_τ = inner.τ0 / (findim_data.opAnorm_squared * F::cast_from(μ.len())); - let iters = quadratic_nonneg(inner.method, &Ã, &g̃, self.α(), - &mut x, inner_τ, iterator); + let normest = findim_data.opAnorm_squared * F::cast_from(μ.len()); + let iters = quadratic_nonneg(&Ã, &g̃, self.α(), &mut x, + normest, inner, iterator); // Update masses of μ based on solution of finite-dimensional subproblem. μ.set_masses_dvector(&x); @@ -342,12 +361,17 @@ #[replace_float_literals(F::cast_from(literal))] impl<F : Float + ToNalgebraRealField, A, I, S, GA, BTA, const N : usize> RegTermFW<F, A, I, N> for NonnegRadonRegTerm<F> -where Cube<F, N> : P2Minimise<Loc<F, N>, F>, - I : AlgIteratorFactory<F>, - S: RealMapping<F, N> + LocalAnalysis<F, Bounds<F>, N>, - GA : SupportGenerator<F, N, SupportType = S, Id = usize> + Clone, - A : ForwardModel<Loc<F, N>, F, PreadjointCodomain = BTFN<F, GA, BTA, N>>, - BTA : BTSearch<F, N, Data=usize, Agg=Bounds<F>> { +where + Cube<F, N> : P2Minimise<Loc<F, N>, F>, + I : AlgIteratorFactory<F>, + S: RealMapping<F, N> + LocalAnalysis<F, Bounds<F>, N>, + GA : SupportGenerator<F, N, SupportType = S, Id = usize> + Clone, + A : FindimQuadraticModel<Loc<F, N>, F, PreadjointCodomain = BTFN<F, GA, BTA, N>>, + BTA : BTSearch<F, N, Data=usize, Agg=Bounds<F>>, + // FIXME: the following *should not* be needed, they are already implied + RNDM<F, N> : Mapping<A::PreadjointCodomain, Codomain = F>, + DeltaMeasure<Loc<F, N>, F> : Mapping<A::PreadjointCodomain, Codomain = F>, +{ fn find_insertion( &self, @@ -361,7 +385,7 @@ fn relaxed_insert<'a>( &self, - μ : &mut DiscreteMeasure<Loc<F, N>, F>, + μ : &mut RNDM<F, N>, g : &A::PreadjointCodomain, opA : &'a A, ξ : Loc<F, N>, @@ -401,28 +425,26 @@ /// `iterator` is used to iterate the steps of the method, and `plotter` may be used to /// save intermediate iteration states as images. #[replace_float_literals(F::cast_from(literal))] -pub fn pointsource_fw_reg<'a, F, I, A, GA, BTA, S, Reg, const N : usize>( - opA : &'a A, +pub fn pointsource_fw_reg<F, I, A, GA, BTA, S, Reg, const N : usize>( + opA : &A, b : &A::Observable, reg : Reg, //domain : Cube<F, N>, config : &FWConfig<F>, iterator : I, mut plotter : SeqPlotter<F, N>, -) -> DiscreteMeasure<Loc<F, N>, F> +) -> RNDM<F, N> where F : Float + ToNalgebraRealField, I : AlgIteratorFactory<IterInfo<F, N>>, for<'b> &'b A::Observable : std::ops::Neg<Output=A::Observable>, - //+ std::ops::Mul<F, Output=A::Observable>, <-- FIXME: compiler overflow - A::Observable : std::ops::MulAssign<F>, GA : SupportGenerator<F, N, SupportType = S, Id = usize> + Clone, - A : ForwardModel<Loc<F, N>, F, PreadjointCodomain = BTFN<F, GA, BTA, N>>, + A : ForwardModel<RNDM<F, N>, F, PreadjointCodomain = BTFN<F, GA, BTA, N>>, BTA : BTSearch<F, N, Data=usize, Agg=Bounds<F>>, S: RealMapping<F, N> + LocalAnalysis<F, Bounds<F>, N>, BTNodeLookup: BTNode<F, usize, Bounds<F>, N>, Cube<F, N>: P2Minimise<Loc<F, N>, F>, PlotLookup : Plotting<N>, - DiscreteMeasure<Loc<F, N>, F> : SpikeMerging<F>, + RNDM<F, N> : SpikeMerging<F>, Reg : RegTermFW<F, A, ValueIteratorFactory<F, AlgIteratorOptions>, N> { // Set up parameters @@ -438,26 +460,24 @@ let mut μ = DiscreteMeasure::new(); let mut residual = -b; - let mut inner_iters = 0; - let mut this_iters = 0; - let mut pruned = 0; - let mut merged = 0; + // Statistics + let full_stats = |residual : &A::Observable, + ν : &RNDM<F, N>, + ε, stats| IterInfo { + value : residual.norm2_squared_div2() + reg.apply(ν), + n_spikes : ν.len(), + ε, + .. stats + }; + let mut stats = IterInfo::new(); // Run the algorithm - iterator.iterate(|state| { - // Update tolerance + for state in iterator.iter_init(|| full_stats(&residual, &μ, ε, stats.clone())) { let inner_tolerance = ε * config.inner.tolerance_mult; let refinement_tolerance = ε * config.refinement.tolerance_mult; - let ε_prev = ε; - ε = tolerance.update(ε, state.iteration()); // Calculate smooth part of surrogate model. - // - // Using `std::mem::replace` here is not ideal, and expects that `empty_observable` - // has no significant overhead. For some reosn Rust doesn't allow us simply moving - // the residual and replacing it below before the end of this closure. - let r = std::mem::replace(&mut residual, opA.empty_observable()); - let mut g = -preadjA.apply(r); + let mut g = preadjA.apply(residual * (-1.0)); // Find absolute value maximising point let (ξ, v_ξ) = reg.find_insertion(&mut g, refinement_tolerance, @@ -467,108 +487,47 @@ FWVariant::FullyCorrective => { // No point in optimising the weight here: the finite-dimensional algorithm is fast. μ += DeltaMeasure { x : ξ, α : 0.0 }; + stats.inserted += 1; config.inner.iterator_options.stop_target(inner_tolerance) }, FWVariant::Relaxed => { // Perform a relaxed initialisation of μ reg.relaxed_insert(&mut μ, &g, opA, ξ, v_ξ, &findim_data); + stats.inserted += 1; // The stop_target is only needed for the type system. AlgIteratorOptions{ max_iter : 1, .. config.inner.iterator_options}.stop_target(0.0) } }; - inner_iters += reg.optimise_weights(&mut μ, opA, b, &findim_data, &config.inner, inner_it); + stats.inner_iters += reg.optimise_weights(&mut μ, opA, b, &findim_data, + &config.inner, inner_it); // Merge spikes and update residual for next step and `if_verbose` below. - let n_before_merge = μ.len(); - residual = μ.merge_spikes_fitness(config.merging, - |μ̃| opA.apply(μ̃) - b, - A::Observable::norm2_squared); - assert!(μ.len() >= n_before_merge); - merged += μ.len() - n_before_merge; - + let (r, count) = μ.merge_spikes_fitness(config.merging, + |μ̃| opA.apply(μ̃) - b, + A::Observable::norm2_squared); + residual = r; + stats.merged += count; // Prune points with zero mass let n_before_prune = μ.len(); μ.prune(); debug_assert!(μ.len() <= n_before_prune); - pruned += n_before_prune - μ.len(); + stats.pruned += n_before_prune - μ.len(); - this_iters +=1; + stats.this_iters += 1; + let iter = state.iteration(); - // Give function value if needed + // Give statistics if needed state.if_verbose(|| { - plotter.plot_spikes( - format!("iter {} start", state.iteration()), &g, - "".to_string(), None::<&A::PreadjointCodomain>, - None, &μ - ); - let res = IterInfo { - value : residual.norm2_squared_div2() + reg.apply(&μ), - n_spikes : μ.len(), - inner_iters, - this_iters, - merged, - pruned, - ε : ε_prev, - postprocessing : None, - }; - inner_iters = 0; - this_iters = 0; - merged = 0; - pruned = 0; - res - }) - }); + plotter.plot_spikes(iter, Some(&g), Option::<&S>::None, &μ); + full_stats(&residual, &μ, ε, std::mem::replace(&mut stats, IterInfo::new())) + }); + + // Update tolerance + ε = tolerance.update(ε, iter); + } // Return final iterate μ } - -// -// Deprecated interface -// - -#[deprecated(note = "Use `pointsource_fw_reg`")] -pub fn pointsource_fw<'a, F, I, A, GA, BTA, S, const N : usize>( - opA : &'a A, - b : &A::Observable, - α : F, - //domain : Cube<F, N>, - config : &FWConfig<F>, - iterator : I, - plotter : SeqPlotter<F, N>, -) -> DiscreteMeasure<Loc<F, N>, F> -where F : Float + ToNalgebraRealField, - I : AlgIteratorFactory<IterInfo<F, N>>, - for<'b> &'b A::Observable : std::ops::Neg<Output=A::Observable>, - //+ std::ops::Mul<F, Output=A::Observable>, <-- FIXME: compiler overflow - A::Observable : std::ops::MulAssign<F>, - GA : SupportGenerator<F, N, SupportType = S, Id = usize> + Clone, - A : ForwardModel<Loc<F, N>, F, PreadjointCodomain = BTFN<F, GA, BTA, N>>, - BTA : BTSearch<F, N, Data=usize, Agg=Bounds<F>>, - S: RealMapping<F, N> + LocalAnalysis<F, Bounds<F>, N>, - BTNodeLookup: BTNode<F, usize, Bounds<F>, N>, - Cube<F, N>: P2Minimise<Loc<F, N>, F>, - PlotLookup : Plotting<N>, - DiscreteMeasure<Loc<F, N>, F> : SpikeMerging<F> { - - pointsource_fw_reg(opA, b, NonnegRadonRegTerm(α), config, iterator, plotter) -} - -#[deprecated(note = "Use `WeightOptim::optimise_weights`")] -pub fn optimise_weights<'a, F, A, I, const N : usize>( - μ : &mut DiscreteMeasure<Loc<F, N>, F>, - opA : &'a A, - b : &A::Observable, - α : F, - findim_data : &FindimData<F>, - inner : &InnerSettings<F>, - iterator : I -) -> usize -where F : Float + ToNalgebraRealField, - I : AlgIteratorFactory<F>, - A : ForwardModel<Loc<F, N>, F> -{ - NonnegRadonRegTerm(α).optimise_weights(μ, opA, b, findim_data, inner, iterator) -}
--- a/src/kernels.rs Tue Aug 01 10:25:09 2023 +0300 +++ b/src/kernels.rs Mon Feb 17 13:54:53 2025 -0500 @@ -24,4 +24,7 @@ pub use ball_indicator::*; mod hat_convolution; pub use hat_convolution::*; +mod linear; +pub use linear::*; +
--- a/src/kernels/ball_indicator.rs Tue Aug 01 10:25:09 2023 +0300 +++ b/src/kernels/ball_indicator.rs Mon Feb 17 13:54:53 2025 -0500 @@ -1,6 +1,6 @@ //! Implementation of the indicator function of a ball with respect to various norms. -use float_extras::f64::{tgamma as gamma}; +use float_extras::f64::tgamma as gamma; use numeric_literals::replace_float_literals; use serde::Serialize; use alg_tools::types::*; @@ -14,10 +14,16 @@ LocalAnalysis, GlobalAnalysis, }; -use alg_tools::mapping::Apply; +use alg_tools::mapping::{ + Mapping, + Differential, + DifferentiableImpl, +}; +use alg_tools::instance::Instance; +use alg_tools::euclidean::StaticEuclidean; use alg_tools::maputil::array_init; use alg_tools::coefficients::factorial; - +use crate::types::*; use super::base::*; /// Representation of the indicator of the ball $𝔹_q = \\{ x ∈ ℝ^N \mid \\|x\\|\_q ≤ r \\}$, @@ -36,14 +42,17 @@ #[replace_float_literals(C::Type::cast_from(literal))] impl<'a, F : Float, C : Constant<Type=F>, Exponent : NormExponent, const N : usize> -Apply<&'a Loc<C::Type, N>> +Mapping<Loc<C::Type, N>> for BallIndicator<C, Exponent, N> -where Loc<F, N> : Norm<F, Exponent> { - type Output = C::Type; +where + Loc<F, N> : Norm<F, Exponent> +{ + type Codomain = C::Type; + #[inline] - fn apply(&self, x : &'a Loc<C::Type, N>) -> Self::Output { + fn apply<I : Instance<Loc<C::Type, N>>>(&self, x : I) -> Self::Codomain { let r = self.r.value(); - let n = x.norm(self.exponent); + let n = x.eval(|x| x.norm(self.exponent)); if n <= r { 1.0 } else { @@ -52,14 +61,79 @@ } } +impl<'a, F : Float, C : Constant<Type=F>, Exponent : NormExponent, const N : usize> +DifferentiableImpl<Loc<C::Type, N>> +for BallIndicator<C, Exponent, N> +where + C : Constant, + Loc<F, N> : Norm<F, Exponent> +{ + type Derivative = Loc<C::Type, N>; + + #[inline] + fn differential_impl<I : Instance<Loc<C::Type, N>>>(&self, _x : I) -> Self::Derivative { + Self::Derivative::origin() + } +} + impl<F : Float, C : Constant<Type=F>, Exponent : NormExponent, const N : usize> -Apply<Loc<C::Type, N>> +Lipschitz<L2> for BallIndicator<C, Exponent, N> -where Loc<F, N> : Norm<F, Exponent> { - type Output = C::Type; - #[inline] - fn apply(&self, x : Loc<C::Type, N>) -> Self::Output { - self.apply(&x) +where C : Constant, + Loc<F, N> : Norm<F, Exponent> { + type FloatType = C::Type; + + fn lipschitz_factor(&self, _l2 : L2) -> Option<C::Type> { + None + } +} + +impl<'b, F : Float, C : Constant<Type=F>, Exponent : NormExponent, const N : usize> +Lipschitz<L2> +for Differential<'b, Loc<F, N>, BallIndicator<C, Exponent, N>> +where C : Constant, + Loc<F, N> : Norm<F, Exponent> { + type FloatType = C::Type; + + fn lipschitz_factor(&self, _l2 : L2) -> Option<C::Type> { + None + } +} + +impl<'a, 'b, F : Float, C : Constant<Type=F>, Exponent : NormExponent, const N : usize> +Lipschitz<L2> +for Differential<'b, Loc<F, N>, &'a BallIndicator<C, Exponent, N>> +where C : Constant, + Loc<F, N> : Norm<F, Exponent> { + type FloatType = C::Type; + + fn lipschitz_factor(&self, _l2 : L2) -> Option<C::Type> { + None + } +} + + +impl<'b, F : Float, C : Constant<Type=F>, Exponent : NormExponent, const N : usize> +NormBounded<L2> +for Differential<'b, Loc<F, N>, BallIndicator<C, Exponent, N>> +where C : Constant, + Loc<F, N> : Norm<F, Exponent> { + type FloatType = C::Type; + + fn norm_bound(&self, _l2 : L2) -> C::Type { + F::INFINITY + } +} + +impl<'a, 'b, F : Float, C : Constant<Type=F>, Exponent : NormExponent, const N : usize> +NormBounded<L2> +for Differential<'b, Loc<F, N>, &'a BallIndicator<C, Exponent, N>> +where C : Constant, + Loc<F, N> : Norm<F, Exponent> { + type FloatType = C::Type; + + fn norm_bound(&self, _l2 : L2) -> C::Type { + F::INFINITY } } @@ -188,32 +262,21 @@ #[replace_float_literals(F::cast_from(literal))] -impl<'a, F : Float, R, const N : usize> Apply<&'a Loc<F, N>> +impl<'a, F : Float, R, const N : usize> Mapping<Loc<F, N>> for AutoConvolution<CubeIndicator<R, N>> where R : Constant<Type=F> { - type Output = F; + type Codomain = F; #[inline] - fn apply(&self, y : &'a Loc<F, N>) -> F { + fn apply<I : Instance<Loc<F, N>>>(&self, y : I) -> F { let two_r = 2.0 * self.0.r.value(); // This is just a product of one-dimensional versions - y.iter().map(|&x| { + y.cow().iter().map(|&x| { 0.0.max(two_r - x.abs()) }).product() } } -impl<F : Float, R, const N : usize> Apply<Loc<F, N>> -for AutoConvolution<CubeIndicator<R, N>> -where R : Constant<Type=F> { - type Output = F; - - #[inline] - fn apply(&self, y : Loc<F, N>) -> F { - self.apply(&y) - } -} - #[replace_float_literals(F::cast_from(literal))] impl<F : Float, R, const N : usize> Support<F, N> for AutoConvolution<CubeIndicator<R, N>>
--- a/src/kernels/base.rs Tue Aug 01 10:25:09 2023 +0300 +++ b/src/kernels/base.rs Mon Feb 17 13:54:53 2025 -0500 @@ -14,15 +14,22 @@ GlobalAnalysis, Bounded, }; -use alg_tools::mapping::Apply; -use alg_tools::maputil::{array_init, map2}; +use alg_tools::mapping::{ + Mapping, + DifferentiableImpl, + DifferentiableMapping, + Differential, +}; +use alg_tools::instance::{Instance, Space}; +use alg_tools::maputil::{array_init, map2, map1_indexed}; use alg_tools::sets::SetOrd; use crate::fourier::Fourier; +use crate::types::*; /// Representation of the product of two kernels. /// -/// The kernels typically implement [`Support`] and [`Mapping`][alg_tools::mapping::Mapping]. +/// The kernels typically implement [`Support`] and [`Mapping`]. /// /// The implementation [`Support`] only uses the [`Support::support_hint`] of the first parameter! #[derive(Copy,Clone,Serialize,Debug)] @@ -33,32 +40,94 @@ pub B ); -impl<A, B, F : Float, const N : usize> Apply<Loc<F, N>> +impl<A, B, F : Float, const N : usize> Mapping<Loc<F, N>> +for SupportProductFirst<A, B> +where + A : Mapping<Loc<F, N>, Codomain = F>, + B : Mapping<Loc<F, N>, Codomain = F>, +{ + type Codomain = F; + + #[inline] + fn apply<I : Instance<Loc<F, N>>>(&self, x : I) -> Self::Codomain { + self.0.apply(x.ref_instance()) * self.1.apply(x) + } +} + +impl<A, B, F : Float, const N : usize> DifferentiableImpl<Loc<F, N>> for SupportProductFirst<A, B> -where A : for<'a> Apply<&'a Loc<F, N>, Output=F>, - B : for<'a> Apply<&'a Loc<F, N>, Output=F> { - type Output = F; +where + A : DifferentiableMapping< + Loc<F, N>, + DerivativeDomain=Loc<F, N>, + Codomain = F + >, + B : DifferentiableMapping< + Loc<F, N>, + DerivativeDomain=Loc<F, N>, + Codomain = F, + > +{ + type Derivative = Loc<F, N>; + #[inline] - fn apply(&self, x : Loc<F, N>) -> Self::Output { - self.0.apply(&x) * self.1.apply(&x) + fn differential_impl<I : Instance<Loc<F, N>>>(&self, x : I) -> Self::Derivative { + let xr = x.ref_instance(); + self.0.differential(xr) * self.1.apply(xr) + self.1.differential(xr) * self.0.apply(x) } } -impl<'a, A, B, F : Float, const N : usize> Apply<&'a Loc<F, N>> +impl<A, B, M : Copy, F : Float> Lipschitz<M> for SupportProductFirst<A, B> -where A : Apply<&'a Loc<F, N>, Output=F>, - B : Apply<&'a Loc<F, N>, Output=F> { - type Output = F; +where A : Lipschitz<M, FloatType = F> + Bounded<F>, + B : Lipschitz<M, FloatType = F> + Bounded<F> { + type FloatType = F; #[inline] - fn apply(&self, x : &'a Loc<F, N>) -> Self::Output { - self.0.apply(x) * self.1.apply(x) + fn lipschitz_factor(&self, m : M) -> Option<F> { + // f(x)g(x) - f(y)g(y) = f(x)[g(x)-g(y)] - [f(y)-f(x)]g(y) + let &SupportProductFirst(ref f, ref g) = self; + f.lipschitz_factor(m).map(|l| l * g.bounds().uniform()) + .zip(g.lipschitz_factor(m).map(|l| l * f.bounds().uniform())) + .map(|(a, b)| a + b) } } +impl<'a, A, B, M : Copy, Domain, F : Float> Lipschitz<M> +for Differential<'a, Domain, SupportProductFirst<A, B>> +where + Domain : Space, + A : Clone + DifferentiableMapping<Domain> + Lipschitz<M, FloatType = F> + Bounded<F>, + B : Clone + DifferentiableMapping<Domain> + Lipschitz<M, FloatType = F> + Bounded<F>, + SupportProductFirst<A, B> : DifferentiableMapping<Domain>, + for<'b> A::Differential<'b> : Lipschitz<M, FloatType = F> + NormBounded<L2, FloatType=F>, + for<'b> B::Differential<'b> : Lipschitz<M, FloatType = F> + NormBounded<L2, FloatType=F> +{ + type FloatType = F; + #[inline] + fn lipschitz_factor(&self, m : M) -> Option<F> { + // ∇[gf] = f∇g + g∇f + // ⟹ ∇[gf](x) - ∇[gf](y) = f(x)∇g(x) + g(x)∇f(x) - f(y)∇g(y) + g(y)∇f(y) + // = f(x)[∇g(x)-∇g(y)] + g(x)∇f(x) - [f(y)-f(x)]∇g(y) + g(y)∇f(y) + // = f(x)[∇g(x)-∇g(y)] + g(x)[∇f(x)-∇f(y)] + // - [f(y)-f(x)]∇g(y) + [g(y)-g(x)]∇f(y) + let &SupportProductFirst(ref f, ref g) = self.base_fn(); + let (df, dg) = (f.diff_ref(), g.diff_ref()); + [ + df.lipschitz_factor(m).map(|l| l * g.bounds().uniform()), + dg.lipschitz_factor(m).map(|l| l * f.bounds().uniform()), + f.lipschitz_factor(m).map(|l| l * dg.norm_bound(L2)), + g.lipschitz_factor(m).map(|l| l * df.norm_bound(L2)) + ].into_iter().sum() + } +} + + impl<'a, A, B, F : Float, const N : usize> Support<F, N> for SupportProductFirst<A, B> -where A : Support<F, N>, - B : Support<F, N> { +where + A : Support<F, N>, + B : Support<F, N> +{ #[inline] fn support_hint(&self) -> Cube<F, N> { self.0.support_hint() @@ -97,7 +166,7 @@ /// Representation of the sum of two kernels /// -/// The kernels typically implement [`Support`] and [`Mapping`][alg_tools::mapping::Mapping]. +/// The kernels typically implement [`Support`] and [`Mapping`]. /// /// The implementation [`Support`] only uses the [`Support::support_hint`] of the first parameter! #[derive(Copy,Clone,Serialize,Debug)] @@ -108,33 +177,48 @@ pub B ); -impl<'a, A, B, F : Float, const N : usize> Apply<&'a Loc<F, N>> +impl<'a, A, B, F : Float, const N : usize> Mapping<Loc<F, N>> for SupportSum<A, B> -where A : Apply<&'a Loc<F, N>, Output=F>, - B : Apply<&'a Loc<F, N>, Output=F> { - type Output = F; +where + A : Mapping<Loc<F, N>, Codomain = F>, + B : Mapping<Loc<F, N>, Codomain = F>, +{ + type Codomain = F; + #[inline] - fn apply(&self, x : &'a Loc<F, N>) -> Self::Output { - self.0.apply(x) + self.1.apply(x) + fn apply<I : Instance<Loc<F, N>>>(&self, x : I) -> Self::Codomain { + self.0.apply(x.ref_instance()) + self.1.apply(x) } } -impl<A, B, F : Float, const N : usize> Apply<Loc<F, N>> +impl<'a, A, B, F : Float, const N : usize> DifferentiableImpl<Loc<F, N>> for SupportSum<A, B> -where A : for<'a> Apply<&'a Loc<F, N>, Output=F>, - B : for<'a> Apply<&'a Loc<F, N>, Output=F> { - type Output = F; +where + A : DifferentiableMapping< + Loc<F, N>, + DerivativeDomain = Loc<F, N> + >, + B : DifferentiableMapping< + Loc<F, N>, + DerivativeDomain = Loc<F, N>, + > +{ + + type Derivative = Loc<F, N>; + #[inline] - fn apply(&self, x : Loc<F, N>) -> Self::Output { - self.0.apply(&x) + self.1.apply(&x) + fn differential_impl<I : Instance<Loc<F, N>>>(&self, x : I) -> Self::Derivative { + self.0.differential(x.ref_instance()) + self.1.differential(x) } } + impl<'a, A, B, F : Float, const N : usize> Support<F, N> for SupportSum<A, B> where A : Support<F, N>, B : Support<F, N>, Cube<F, N> : SetOrd { + #[inline] fn support_hint(&self) -> Cube<F, N> { self.0.support_hint().common(&self.1.support_hint()) @@ -174,9 +258,42 @@ } } +impl<F : Float, M : Copy, A, B> Lipschitz<M> for SupportSum<A, B> +where A : Lipschitz<M, FloatType = F>, + B : Lipschitz<M, FloatType = F> { + type FloatType = F; + + fn lipschitz_factor(&self, m : M) -> Option<F> { + match (self.0.lipschitz_factor(m), self.1.lipschitz_factor(m)) { + (Some(l0), Some(l1)) => Some(l0 + l1), + _ => None + } + } +} + +impl<'b, F : Float, M : Copy, A, B, Domain> Lipschitz<M> +for Differential<'b, Domain, SupportSum<A, B>> +where + Domain : Space, + A : Clone + DifferentiableMapping<Domain, Codomain=F>, + B : Clone + DifferentiableMapping<Domain, Codomain=F>, + SupportSum<A, B> : DifferentiableMapping<Domain, Codomain=F>, + for<'a> A :: Differential<'a> : Lipschitz<M, FloatType = F>, + for<'a> B :: Differential<'a> : Lipschitz<M, FloatType = F> +{ + type FloatType = F; + + fn lipschitz_factor(&self, m : M) -> Option<F> { + let base = self.base_fn(); + base.0.diff_ref().lipschitz_factor(m) + .zip(base.1.diff_ref().lipschitz_factor(m)) + .map(|(a, b)| a + b) + } +} + /// Representation of the convolution of two kernels. /// -/// The kernels typically implement [`Support`]s and [`Mapping`][alg_tools::mapping::Mapping]. +/// The kernels typically implement [`Support`]s and [`Mapping`]. // /// Trait implementations have to be on a case-by-case basis. #[derive(Copy,Clone,Serialize,Debug,Eq,PartialEq)] @@ -187,9 +304,46 @@ pub B ); +impl<F : Float, M, A, B> Lipschitz<M> for Convolution<A, B> +where A : Norm<F, L1> , + B : Lipschitz<M, FloatType = F> { + type FloatType = F; + + fn lipschitz_factor(&self, m : M) -> Option<F> { + // For [f * g](x) = ∫ f(x-y)g(y) dy we have + // [f * g](x) - [f * g](z) = ∫ [f(x-y)-f(z-y)]g(y) dy. + // Hence |[f * g](x) - [f * g](z)| ≤ ∫ |f(x-y)-f(z-y)|g(y)| dy. + // ≤ L|x-z| ∫ |g(y)| dy, + // where L is the Lipschitz factor of f. + self.1.lipschitz_factor(m).map(|l| l * self.0.norm(L1)) + } +} + +impl<'b, F : Float, M, A, B, Domain> Lipschitz<M> +for Differential<'b, Domain, Convolution<A, B>> +where + Domain : Space, + A : Clone + Norm<F, L1> , + Convolution<A, B> : DifferentiableMapping<Domain, Codomain=F>, + B : Clone + DifferentiableMapping<Domain, Codomain=F>, + for<'a> B :: Differential<'a> : Lipschitz<M, FloatType = F> +{ + type FloatType = F; + + fn lipschitz_factor(&self, m : M) -> Option<F> { + // For [f * g](x) = ∫ f(x-y)g(y) dy we have + // ∇[f * g](x) - ∇[f * g](z) = ∫ [∇f(x-y)-∇f(z-y)]g(y) dy. + // Hence |∇[f * g](x) - ∇[f * g](z)| ≤ ∫ |∇f(x-y)-∇f(z-y)|g(y)| dy. + // ≤ L|x-z| ∫ |g(y)| dy, + // where L is the Lipschitz factor of ∇f. + let base = self.base_fn(); + base.1.diff_ref().lipschitz_factor(m).map(|l| l * base.0.norm(L1)) + } +} + /// Representation of the autoconvolution of a kernel. /// -/// The kernel typically implements [`Support`] and [`Mapping`][alg_tools::mapping::Mapping]. +/// The kernel typically implements [`Support`] and [`Mapping`]. /// /// Trait implementations have to be on a case-by-case basis. #[derive(Copy,Clone,Serialize,Debug,Eq,PartialEq)] @@ -198,34 +352,130 @@ pub A ); +impl<F : Float, M, C> Lipschitz<M> for AutoConvolution<C> +where C : Lipschitz<M, FloatType = F> + Norm<F, L1> { + type FloatType = F; + + fn lipschitz_factor(&self, m : M) -> Option<F> { + self.0.lipschitz_factor(m).map(|l| l * self.0.norm(L1)) + } +} + +impl<'b, F : Float, M, C, Domain> Lipschitz<M> +for Differential<'b, Domain, AutoConvolution<C>> +where + Domain : Space, + C : Clone + Norm<F, L1> + DifferentiableMapping<Domain, Codomain=F>, + AutoConvolution<C> : DifferentiableMapping<Domain, Codomain=F>, + for<'a> C :: Differential<'a> : Lipschitz<M, FloatType = F> +{ + type FloatType = F; + + fn lipschitz_factor(&self, m : M) -> Option<F> { + let base = self.base_fn(); + base.0.diff_ref().lipschitz_factor(m).map(|l| l * base.0.norm(L1)) + } +} + + /// Representation a multi-dimensional product of a one-dimensional kernel. /// /// For $G: ℝ → ℝ$, this is the function $F(x\_1, …, x\_n) := \prod_{i=1}^n G(x\_i)$. -/// The kernel $G$ typically implements [`Support`] and [`Mapping`][alg_tools::mapping::Mapping] +/// The kernel $G$ typically implements [`Support`] and [`Mapping`] /// on [`Loc<F, 1>`]. Then the product implements them on [`Loc<F, N>`]. #[derive(Copy,Clone,Serialize,Debug,Eq,PartialEq)] +#[allow(dead_code)] struct UniformProduct<G, const N : usize>( /// The one-dimensional kernel G ); -impl<'a, G, F : Float, const N : usize> Apply<&'a Loc<F, N>> +impl<'a, G, F : Float, const N : usize> Mapping<Loc<F, N>> for UniformProduct<G, N> -where G : Apply<Loc<F, 1>, Output=F> { - type Output = F; +where + G : Mapping<Loc<F, 1>, Codomain = F> +{ + type Codomain = F; + #[inline] - fn apply(&self, x : &'a Loc<F, N>) -> F { - x.iter().map(|&y| self.0.apply(Loc([y]))).product() + fn apply<I : Instance<Loc<F, N>>>(&self, x : I) -> F { + x.cow().iter().map(|&y| self.0.apply(Loc([y]))).product() + } +} + + + +impl<'a, G, F : Float, const N : usize> DifferentiableImpl<Loc<F, N>> +for UniformProduct<G, N> +where + G : DifferentiableMapping< + Loc<F, 1>, + DerivativeDomain = F, + Codomain = F, + > +{ + type Derivative = Loc<F, N>; + + #[inline] + fn differential_impl<I : Instance<Loc<F, N>>>(&self, x0 : I) -> Loc<F, N> { + x0.eval(|x| { + let vs = x.map(|y| self.0.apply(Loc([y]))); + product_differential(x, &vs, |y| self.0.differential(Loc([y]))) + }) } } -impl<G, F : Float, const N : usize> Apply<Loc<F, N>> -for UniformProduct<G, N> -where G : Apply<Loc<F, 1>, Output=F> { - type Output = F; - #[inline] - fn apply(&self, x : Loc<F, N>) -> F { - x.into_iter().map(|y| self.0.apply(Loc([y]))).product() +/// Helper function to calulate the differential of $f(x)=∏_{i=1}^N g(x_i)$. +/// +/// The vector `x` is the location, `vs` consists of the values `g(x_i)`, and +/// `gd` calculates the derivative `g'`. +#[inline] +pub(crate) fn product_differential<F : Float, G : Fn(F) -> F, const N : usize>( + x : &Loc<F, N>, + vs : &Loc<F, N>, + gd : G +) -> Loc<F, N> { + map1_indexed(x, |i, &y| { + gd(y) * vs.iter() + .zip(0..) + .filter_map(|(v, j)| (j != i).then_some(*v)) + .product() + }).into() +} + +/// Helper function to calulate the Lipschitz factor of $∇f$ for $f(x)=∏_{i=1}^N g(x_i)$. +/// +/// The parameter `bound` is a bound on $|g|_∞$, `lip` is a Lipschitz factor for $g$, +/// `dbound` is a bound on $|∇g|_∞$, and `dlip` a Lipschitz factor for $∇g$. +#[inline] +pub(crate) fn product_differential_lipschitz_factor<F : Float, const N : usize>( + bound : F, + lip : F, + dbound : F, + dlip : F +) -> F { + // For arbitrary ψ(x) = ∏_{i=1}^n ψ_i(x_i), we have + // ψ(x) - ψ(y) = ∑_i [ψ_i(x_i)-ψ_i(y_i)] ∏_{j ≠ i} ψ_j(x_j) + // by a simple recursive argument. In particular, if ψ_i=g for all i, j, we have + // |ψ(x) - ψ(y)| ≤ ∑_i L_g M_g^{n-1}|x-y|, where L_g is the Lipschitz factor of g, and + // M_g a bound on it. + // + // We also have in the general case ∇ψ(x) = ∑_i ∇ψ_i(x_i) ∏_{j ≠ i} ψ_j(x_j), whence + // using the previous formula for each i with f_i=∇ψ_i and f_j=ψ_j for j ≠ i, we get + // ∇ψ(x) - ∇ψ(y) = ∑_i[ ∇ψ_i(x_i)∏_{j ≠ i} ψ_j(x_j) - ∇ψ_i(y_i)∏_{j ≠ i} ψ_j(y_j)] + // = ∑_i[ [∇ψ_i(x_i) - ∇ψ_j(x_j)] ∏_{j ≠ i}ψ_j(x_j) + // + [∑_{k ≠ i} [ψ_k(x_k) - ∇ψ_k(x_k)] ∏_{j ≠ i, k}ψ_j(x_j)]∇ψ_i(x_i)]. + // With $ψ_i=g for all i, j, it follows that + // |∇ψ(x) - ∇ψ(y)| ≤ ∑_i L_{∇g} M_g^{n-1} + ∑_{k ≠ i} L_g M_g^{n-2} M_{∇g} + // = n [L_{∇g} M_g^{n-1} + (n-1) L_g M_g^{n-2} M_{∇g}]. + // = n M_g^{n-2}[L_{∇g} M_g + (n-1) L_g M_{∇g}]. + if N >= 2 { + F::cast_from(N) * bound.powi((N-2) as i32) + * (dlip * bound + F::cast_from(N-1) * lip * dbound) + } else if N==1 { + dlip + } else { + panic!("Invalid dimension") } }
--- a/src/kernels/gaussian.rs Tue Aug 01 10:25:09 2023 +0300 +++ b/src/kernels/gaussian.rs Mon Feb 17 13:54:53 2025 -0500 @@ -17,9 +17,15 @@ Weighted, Bounded, }; -use alg_tools::mapping::Apply; +use alg_tools::mapping::{ + Mapping, + Instance, + Differential, + DifferentiableImpl, +}; use alg_tools::maputil::array_init; +use crate::types::*; use crate::fourier::Fourier; use super::base::*; use super::ball_indicator::CubeIndicator; @@ -58,28 +64,104 @@ #[replace_float_literals(S::Type::cast_from(literal))] -impl<'a, S, const N : usize> Apply<&'a Loc<S::Type, N>> for Gaussian<S, N> -where S : Constant { - type Output = S::Type; +impl<'a, S, const N : usize> Mapping<Loc<S::Type, N>> for Gaussian<S, N> +where + S : Constant +{ + type Codomain = S::Type; + // This is not normalised to neither to have value 1 at zero or integral 1 // (unless the cut-off ε=0). #[inline] - fn apply(&self, x : &'a Loc<S::Type, N>) -> Self::Output { - let d_squared = x.norm2_squared(); + fn apply<I : Instance<Loc<S::Type, N>>>(&self, x : I) -> Self::Codomain { + let d_squared = x.eval(|x| x.norm2_squared()); let σ2 = self.variance.value(); let scale = self.scale(); (-d_squared / (2.0 * σ2)).exp() / scale } } -impl<S, const N : usize> Apply<Loc<S::Type, N>> for Gaussian<S, N> +#[replace_float_literals(S::Type::cast_from(literal))] +impl<'a, S, const N : usize> DifferentiableImpl<Loc<S::Type, N>> for Gaussian<S, N> +where S : Constant { + type Derivative = Loc<S::Type, N>; + + #[inline] + fn differential_impl<I : Instance<Loc<S::Type, N>>>(&self, x0 : I) -> Self::Derivative { + let x = x0.cow(); + let f = -self.apply(&*x) / self.variance.value(); + *x * f + } +} + + +// To calculate the the Lipschitz factors, we consider +// f(t) = e^{-t²/2} +// f'(t) = -t f(t) which has max at t=1 by f''(t)=0 +// f''(t) = (t²-1)f(t) which has max at t=√3 by f'''(t)=0 +// f'''(t) = -(t³-3t) +// So f has the Lipschitz factor L=f'(1), and f' has the Lipschitz factor L'=f''(√3). +// +// Now g(x) = Cf(‖x‖/σ) for a scaling factor C is the Gaussian. +// Thus ‖g(x)-g(y)‖ = C‖f(‖x‖/σ)-f(‖y‖/σ)‖ ≤ (C/σ)L‖x-y‖, +// so g has the Lipschitz factor (C/σ)f'(1) = (C/σ)exp(-0.5). +// +// Also ∇g(x)= Cx/(σ‖x‖)f'(‖x‖/σ) (*) +// = -(C/σ²)xf(‖x‖/σ) +// = -C/σ (x/σ) f(‖x/σ‖) +// ∇²g(x) = -(C/σ)[Id/σ f(‖x‖/σ) + x ⊗ x/(σ²‖x‖) f'(‖x‖/σ)] +// = (C/σ²)[-Id + x ⊗ x/σ²]f(‖x‖/σ). +// Thus ‖∇²g(x)‖ = (C/σ²)‖-Id + x ⊗ x/σ²‖f(‖x‖/σ), where +// ‖-Id + x ⊗ x/σ²‖ = ‖[-Id + x ⊗ x/σ²](x/‖x‖)‖ = |-1 + ‖x²/σ^2‖|. +// This means that ‖∇²g(x)‖ = (C/σ²)|f''(‖x‖/σ)|, which is maximised with ‖x‖/σ=√3. +// Hence the Lipschitz factor of ∇g is (C/σ²)f''(√3) = (C/σ²)2e^{-3/2}. + +#[replace_float_literals(S::Type::cast_from(literal))] +impl<S, const N : usize> Lipschitz<L2> for Gaussian<S, N> where S : Constant { - type Output = S::Type; - // This is not normalised to neither to have value 1 at zero or integral 1 - // (unless the cut-off ε=0). - #[inline] - fn apply(&self, x : Loc<S::Type, N>) -> Self::Output { - self.apply(&x) + type FloatType = S::Type; + fn lipschitz_factor(&self, L2 : L2) -> Option<Self::FloatType> { + Some((-0.5).exp() / (self.scale() * self.variance.value().sqrt())) + } +} + + +#[replace_float_literals(S::Type::cast_from(literal))] +impl<'a, S : Constant, const N : usize> Lipschitz<L2> +for Differential<'a, Loc<S::Type, N>, Gaussian<S, N>> { + type FloatType = S::Type; + + fn lipschitz_factor(&self, _l2 : L2) -> Option<S::Type> { + let g = self.base_fn(); + let σ2 = g.variance.value(); + let scale = g.scale(); + Some(2.0*(-3.0/2.0).exp()/(σ2*scale)) + } +} + +// From above, norm bounds on the differnential can be calculated as achieved +// for f' at t=1, i.e., the bound is |f'(1)|. +// For g then |C/σ f'(1)|. +// It follows that the norm bounds on the differential are just the Lipschitz +// factors of the undifferentiated function, given how the latter is calculed above. + +#[replace_float_literals(S::Type::cast_from(literal))] +impl<'b, S : Constant, const N : usize> NormBounded<L2> +for Differential<'b, Loc<S::Type, N>, Gaussian<S, N>> { + type FloatType = S::Type; + + fn norm_bound(&self, _l2 : L2) -> S::Type { + self.base_fn().lipschitz_factor(L2).unwrap() + } +} + +#[replace_float_literals(S::Type::cast_from(literal))] +impl<'b, 'a, S : Constant, const N : usize> NormBounded<L2> +for Differential<'b, Loc<S::Type, N>, &'a Gaussian<S, N>> { + type FloatType = S::Type; + + fn norm_bound(&self, _l2 : L2) -> S::Type { + self.base_fn().lipschitz_factor(L2).unwrap() } } @@ -169,19 +251,19 @@ Gaussian<S, N>>; -/// This implements $χ\_{[-b, b]^n} \* (f χ\_{[-a, a]^n})$ -/// where $a,b>0$ and $f$ is a gaussian kernel on $ℝ^n$. +/// This implements $g := χ\_{[-b, b]^n} \* (f χ\_{[-a, a]^n})$ where $a,b>0$ and $f$ is +/// a gaussian kernel on $ℝ^n$. For an expression for $g$, see Lemma 3.9 in the manuscript. #[replace_float_literals(F::cast_from(literal))] -impl<'a, F : Float, R, C, S, const N : usize> Apply<&'a Loc<F, N>> +impl<'a, F : Float, R, C, S, const N : usize> Mapping<Loc<F, N>> for Convolution<CubeIndicator<R, N>, BasicCutGaussian<C, S, N>> where R : Constant<Type=F>, C : Constant<Type=F>, S : Constant<Type=F> { - type Output = F; + type Codomain = F; #[inline] - fn apply(&self, y : &'a Loc<F, N>) -> F { + fn apply<I : Instance<Loc<F, N>>>(&self, y : I) -> F { let Convolution(ref ind, SupportProductFirst(ref cut, ref gaussian)) = self; @@ -192,7 +274,7 @@ let c = 0.5; // 1/(σ√(2π) * σ√(π/2) = 1/2 // This is just a product of one-dimensional versions - y.product_map(|x| { + y.cow().product_map(|x| { let c1 = -(a.min(b + x)); //(-a).max(-x-b); let c2 = a.min(b - x); if c1 >= c2 { @@ -207,20 +289,143 @@ } } -impl<F : Float, R, C, S, const N : usize> Apply<Loc<F, N>> +/// This implements the differential of $g := χ\_{[-b, b]^n} \* (f χ\_{[-a, a]^n})$ where $a,b>0$ +/// and $f$ is a gaussian kernel on $ℝ^n$. For an expression for the value of $g$, from which the +/// derivative readily arises (at points of differentiability), see Lemma 3.9 in the manuscript. +#[replace_float_literals(F::cast_from(literal))] +impl<'a, F : Float, R, C, S, const N : usize> DifferentiableImpl<Loc<F, N>> for Convolution<CubeIndicator<R, N>, BasicCutGaussian<C, S, N>> where R : Constant<Type=F>, C : Constant<Type=F>, S : Constant<Type=F> { - type Output = F; + type Derivative = Loc<F, N>; + /// Although implemented, this function is not differentiable. #[inline] - fn apply(&self, y : Loc<F, N>) -> F { - self.apply(&y) + fn differential_impl<I : Instance<Loc<F, N>>>(&self, y0 : I) -> Loc<F, N> { + let Convolution(ref ind, + SupportProductFirst(ref cut, + ref gaussian)) = self; + let y = y0.cow(); + let a = cut.r.value(); + let b = ind.r.value(); + let σ = gaussian.variance.value().sqrt(); + let t = F::SQRT_2 * σ; + let c = 0.5; // 1/(σ√(2π) * σ√(π/2) = 1/2 + let c_mul_erf_scale_div_t = c * F::FRAC_2_SQRT_PI / t; + + // Calculate the values for all component functions of the + // product. This is just the loop from apply above. + let unscaled_vs = y.map(|x| { + let c1 = -(a.min(b + x)); //(-a).max(-x-b); + let c2 = a.min(b - x); + if c1 >= c2 { + 0.0 + } else { + let e1 = F::cast_from(erf((c1 / t).as_())); + let e2 = F::cast_from(erf((c2 / t).as_())); + debug_assert!(e2 >= e1); + c * (e2 - e1) + } + }); + // This computes the gradient for each coordinate + product_differential(&*y, &unscaled_vs, |x| { + let c1 = -(a.min(b + x)); //(-a).max(-x-b); + let c2 = a.min(b - x); + if c1 >= c2 { + 0.0 + } else { + // erf'(z) = (2/√π)*exp(-z^2), and we get extra factor 1/(√2*σ) = -1/t + // from the chain rule (the minus comes from inside c_1 or c_2, and changes the + // order of de2 and de1 in the final calculation). + let de1 = if b + x < a { + (-((b+x)/t).powi(2)).exp() + } else { + 0.0 + }; + let de2 = if b - x < a { + (-((b-x)/t).powi(2)).exp() + } else { + 0.0 + }; + c_mul_erf_scale_div_t * (de1 - de2) + } + }) } } + +#[replace_float_literals(F::cast_from(literal))] +impl<'a, F : Float, R, C, S, const N : usize> Lipschitz<L1> +for Convolution<CubeIndicator<R, N>, BasicCutGaussian<C, S, N>> +where R : Constant<Type=F>, + C : Constant<Type=F>, + S : Constant<Type=F> { + type FloatType = F; + + fn lipschitz_factor(&self, L1 : L1) -> Option<F> { + // To get the product Lipschitz factor, we note that for any ψ_i, we have + // ∏_{i=1}^N φ_i(x_i) - ∏_{i=1}^N φ_i(y_i) + // = [φ_1(x_1)-φ_1(y_1)] ∏_{i=2}^N φ_i(x_i) + // + φ_1(y_1)[ ∏_{i=2}^N φ_i(x_i) - ∏_{i=2}^N φ_i(y_i)] + // = ∑_{j=1}^N [φ_j(x_j)-φ_j(y_j)]∏_{i > j} φ_i(x_i) ∏_{i < j} φ_i(y_i) + // Thus + // |∏_{i=1}^N φ_i(x_i) - ∏_{i=1}^N φ_i(y_i)| + // ≤ ∑_{j=1}^N |φ_j(x_j)-φ_j(y_j)| ∏_{j ≠ i} \max_i |φ_i| + // + // Thus we need 1D Lipschitz factors, and the maximum for φ = θ * ψ. + // + // We have + // θ * ψ(x) = 0 if c_1(x) ≥ c_2(x) + // = (1/2)[erf(c_2(x)/(√2σ)) - erf(c_1(x)/(√2σ))] if c_1(x) < c_2(x), + // where c_1(x) = max{-x-b,-a} = -min{b+x,a} and c_2(x)=min{b-x,a}, C is the Gaussian + // normalisation factor, and erf(s) = (2/√π) ∫_0^s e^{-t^2} dt. + // Thus, if c_1(x) < c_2(x) and c_1(y) < c_2(y), we have + // θ * ψ(x) - θ * ψ(y) = (1/√π)[∫_{c_1(x)/(√2σ)}^{c_1(y)/(√2σ) e^{-t^2} dt + // - ∫_{c_2(x)/(√2σ)}^{c_2(y)/(√2σ)] e^{-t^2} dt] + // Thus + // |θ * ψ(x) - θ * ψ(y)| ≤ (1/√π)/(√2σ)(|c_1(x)-c_1(y)|+|c_2(x)-c_2(y)|) + // ≤ 2(1/√π)/(√2σ)|x-y| + // ≤ √2/(√πσ)|x-y|. + // + // For the product we also need the value θ * ψ(0), which is + // (1/2)[erf(min{a,b}/(√2σ))-erf(max{-b,-a}/(√2σ)] + // = (1/2)[erf(min{a,b}/(√2σ))-erf(-min{a,b}/(√2σ))] + // = erf(min{a,b}/(√2σ)) + // + // If c_1(x) ≥ c_2(x), then x ∉ [-(a+b), a+b]. If also y is outside that range, + // θ * ψ(x) = θ * ψ(y). If only y is in the range [-(a+b), a+b], we can replace + // x by -(a+b) or (a+b), either of which is closer to y and still θ * ψ(x)=0. + // Thus same calculations as above work for the Lipschitz factor. + let Convolution(ref ind, + SupportProductFirst(ref cut, + ref gaussian)) = self; + let a = cut.r.value(); + let b = ind.r.value(); + let σ = gaussian.variance.value().sqrt(); + let π = F::PI; + let t = F::SQRT_2 * σ; + let l1d = F::SQRT_2 / (π.sqrt() * σ); + let e0 = F::cast_from(erf((a.min(b) / t).as_())); + Some(l1d * e0.powi(N as i32-1)) + } +} + +/* +impl<'a, F : Float, R, C, S, const N : usize> Lipschitz<L2> +for Convolution<CubeIndicator<R, N>, BasicCutGaussian<C, S, N>> +where R : Constant<Type=F>, + C : Constant<Type=F>, + S : Constant<Type=F> { + type FloatType = F; + #[inline] + fn lipschitz_factor(&self, L2 : L2) -> Option<Self::FloatType> { + self.lipschitz_factor(L1).map(|l1| l1 * <S::Type>::cast_from(N).sqrt()) + } +} +*/ + impl<F : Float, R, C, S, const N : usize> Convolution<CubeIndicator<R, N>, BasicCutGaussian<C, S, N>> where R : Constant<Type=F>,
--- a/src/kernels/hat.rs Tue Aug 01 10:25:09 2023 +0300 +++ b/src/kernels/hat.rs Mon Feb 17 13:54:53 2025 -0500 @@ -14,8 +14,9 @@ GlobalAnalysis, Bounded, }; -use alg_tools::mapping::Apply; -use alg_tools::maputil::{array_init}; +use alg_tools::mapping::{Mapping, Instance}; +use alg_tools::maputil::array_init; +use crate::types::Lipschitz; /// Representation of the hat function $f(x)=1-\\|x\\|\_1/ε$ of `width` $ε$ on $ℝ^N$. #[derive(Copy,Clone,Serialize,Debug,Eq,PartialEq)] @@ -25,26 +26,17 @@ } #[replace_float_literals(C::Type::cast_from(literal))] -impl<'a, C : Constant, const N : usize> Apply<&'a Loc<C::Type, N>> for Hat<C, N> { - type Output = C::Type; +impl<'a, C : Constant, const N : usize> Mapping<Loc<C::Type, N>> for Hat<C, N> { + type Codomain = C::Type; + #[inline] - fn apply(&self, x : &'a Loc<C::Type, N>) -> Self::Output { + fn apply<I : Instance<Loc<C::Type, N>>>(&self, x : I) -> Self::Codomain { let ε = self.width.value(); - 0.0.max(1.0-x.norm(L1)/ε) + 0.0.max(1.0-x.cow().norm(L1)/ε) } } #[replace_float_literals(C::Type::cast_from(literal))] -impl<C : Constant, const N : usize> Apply<Loc<C::Type, N>> for Hat<C, N> { - type Output = C::Type; - #[inline] - fn apply(&self, x : Loc<C::Type, N>) -> Self::Output { - self.apply(&x) - } -} - - -#[replace_float_literals(C::Type::cast_from(literal))] impl<'a, C : Constant, const N : usize> Support<C::Type, N> for Hat<C, N> { #[inline] fn support_hint(&self) -> Cube<C::Type,N> { @@ -94,6 +86,26 @@ } } +#[replace_float_literals(C::Type::cast_from(literal))] +impl<'a, C : Constant, const N : usize> Lipschitz<L1> for Hat<C, N> { + type FloatType = C::Type; + + fn lipschitz_factor(&self, _l1 : L1) -> Option<C::Type> { + Some(1.0/self.width.value()) + } +} + +#[replace_float_literals(C::Type::cast_from(literal))] +impl<'a, C : Constant, const N : usize> Lipschitz<L2> for Hat<C, N> { + type FloatType = C::Type; + + fn lipschitz_factor(&self, _l2 : L2) -> Option<C::Type> { + self.lipschitz_factor(L1).map(|l1| + <L2 as Dominated<C::Type, L1, Loc<C::Type,N>>>::from_norm(&L2, l1, L1) + ) + } +} + impl<'a, C : Constant, const N : usize> LocalAnalysis<C::Type, Bounds<C::Type>, N> for Hat<C, N> {
--- a/src/kernels/hat_convolution.rs Tue Aug 01 10:25:09 2023 +0300 +++ b/src/kernels/hat_convolution.rs Mon Feb 17 13:54:53 2025 -0500 @@ -14,9 +14,15 @@ GlobalAnalysis, Bounded, }; -use alg_tools::mapping::Apply; +use alg_tools::mapping::{ + Mapping, + Instance, + DifferentiableImpl, + Differential, +}; use alg_tools::maputil::array_init; +use crate::types::Lipschitz; use super::base::*; use super::ball_indicator::CubeIndicator; @@ -38,6 +44,31 @@ /// -\frac{2}{3} (y-1)^3 & \frac{1}{2}\leq y<1. \\\\ /// \end{cases} /// $$ +// Hence +// $$ +// (h\*h)'(y) = +// \begin{cases} +// 2 (y+1)^2 & -1<y\leq -\frac{1}{2}, \\\\ +// -6 y^2-4 y & -\frac{1}{2}<y\leq 0, \\\\ +// 6 y^2-4 y & 0<y<\frac{1}{2}, \\\\ +// -2 (y-1)^2 & \frac{1}{2}\leq y<1. \\\\ +// \end{cases} +// $$ +// as well as +// $$ +// (h\*h)''(y) = +// \begin{cases} +// 4 (y+1) & -1<y\leq -\frac{1}{2}, \\\\ +// -12 y-4 & -\frac{1}{2}<y\leq 0, \\\\ +// 12 y-4 & 0<y<\frac{1}{2}, \\\\ +// -4 (y-1) & \frac{1}{2}\leq y<1. \\\\ +// \end{cases} +// $$ +// This is maximised at y=±1/2 with value 2, and minimised at y=0 with value -4. +// Now observe that +// $$ +// [∇f(x\_1, …, x\_n)]_j = \frac{4}{σ} (h\*h)'(x\_j/σ) \prod\_{j ≠ i} \frac{4}{σ} (h\*h)(x\_i/σ) +// $$ #[derive(Copy,Clone,Debug,Serialize,Eq)] pub struct HatConv<S : Constant, const N : usize> { /// The parameter $σ$ of the kernel. @@ -60,24 +91,85 @@ } } -impl<'a, S, const N : usize> Apply<&'a Loc<S::Type, N>> for HatConv<S, N> +impl<'a, S, const N : usize> Mapping<Loc<S::Type, N>> for HatConv<S, N> where S : Constant { - type Output = S::Type; + type Codomain = S::Type; + #[inline] - fn apply(&self, y : &'a Loc<S::Type, N>) -> Self::Output { + fn apply<I : Instance<Loc<S::Type, N>>>(&self, y : I) -> Self::Codomain { let σ = self.radius(); - y.product_map(|x| { + y.cow().product_map(|x| { self.value_1d_σ1(x / σ) / σ }) } } -impl<'a, S, const N : usize> Apply<Loc<S::Type, N>> for HatConv<S, N> +#[replace_float_literals(S::Type::cast_from(literal))] +impl<S, const N : usize> Lipschitz<L1> for HatConv<S, N> +where S : Constant { + type FloatType = S::Type; + #[inline] + fn lipschitz_factor(&self, L1 : L1) -> Option<Self::FloatType> { + // For any ψ_i, we have + // ∏_{i=1}^N ψ_i(x_i) - ∏_{i=1}^N ψ_i(y_i) + // = [ψ_1(x_1)-ψ_1(y_1)] ∏_{i=2}^N ψ_i(x_i) + // + ψ_1(y_1)[ ∏_{i=2}^N ψ_i(x_i) - ∏_{i=2}^N ψ_i(y_i)] + // = ∑_{j=1}^N [ψ_j(x_j)-ψ_j(y_j)]∏_{i > j} ψ_i(x_i) ∏_{i < j} ψ_i(y_i) + // Thus + // |∏_{i=1}^N ψ_i(x_i) - ∏_{i=1}^N ψ_i(y_i)| + // ≤ ∑_{j=1}^N |ψ_j(x_j)-ψ_j(y_j)| ∏_{j ≠ i} \max_j |ψ_j| + let σ = self.radius(); + let l1d = self.lipschitz_1d_σ1() / (σ*σ); + let m1d = self.value_1d_σ1(0.0) / σ; + Some(l1d * m1d.powi(N as i32 - 1)) + } +} + +impl<S, const N : usize> Lipschitz<L2> for HatConv<S, N> +where S : Constant { + type FloatType = S::Type; + #[inline] + fn lipschitz_factor(&self, L2 : L2) -> Option<Self::FloatType> { + self.lipschitz_factor(L1).map(|l1| l1 * <S::Type>::cast_from(N).sqrt()) + } +} + + +impl<'a, S, const N : usize> DifferentiableImpl<Loc<S::Type, N>> for HatConv<S, N> where S : Constant { - type Output = S::Type; + type Derivative = Loc<S::Type, N>; + #[inline] - fn apply(&self, y : Loc<S::Type, N>) -> Self::Output { - self.apply(&y) + fn differential_impl<I : Instance<Loc<S::Type, N>>>(&self, y0 : I) -> Self::Derivative { + let y = y0.cow(); + let σ = self.radius(); + let σ2 = σ * σ; + let vs = y.map(|x| { + self.value_1d_σ1(x / σ) / σ + }); + product_differential(&*y, &vs, |x| { + self.diff_1d_σ1(x / σ) / σ2 + }) + } +} + + +#[replace_float_literals(S::Type::cast_from(literal))] +impl<'a, F : Float, S, const N : usize> Lipschitz<L2> +for Differential<'a, Loc<F, N>, HatConv<S, N>> +where S : Constant<Type=F> { + type FloatType = F; + + #[inline] + fn lipschitz_factor(&self, _l2 : L2) -> Option<F> { + let h = self.base_fn(); + let σ = h.radius(); + Some(product_differential_lipschitz_factor::<F, N>( + h.value_1d_σ1(0.0) / σ, + h.lipschitz_1d_σ1() / (σ*σ), + h.maxabsdiff_1d_σ1() / (σ*σ), + h.lipschitz_diff_1d_σ1() / (σ*σ), + )) } } @@ -97,6 +189,54 @@ (4.0/3.0) + 8.0 * y * y * (y - 1.0) } } + + /// Computes the differential of the kernel for $n=1$ with $σ=1$. + #[inline] + fn diff_1d_σ1(&self, x : F) -> F { + let y = x.abs(); + if y >= 1.0 { + 0.0 + } else if y > 0.5 { + - 8.0 * (y - 1.0).powi(2) + } else /* 0 ≤ y ≤ 0.5 */ { + (24.0 * y - 16.0) * y + } + } + + /// Computes the Lipschitz factor of the kernel for $n=1$ with $σ=1$. + #[inline] + fn lipschitz_1d_σ1(&self) -> F { + // Maximal absolute differential achieved at ±0.5 by diff_1d_σ1 analysis + 2.0 + } + + /// Computes the maximum absolute differential of the kernel for $n=1$ with $σ=1$. + #[inline] + fn maxabsdiff_1d_σ1(&self) -> F { + // Maximal absolute differential achieved at ±0.5 by diff_1d_σ1 analysis + 2.0 + } + + /// Computes the second differential of the kernel for $n=1$ with $σ=1$. + #[inline] + #[allow(dead_code)] + fn diff2_1d_σ1(&self, x : F) -> F { + let y = x.abs(); + if y >= 1.0 { + 0.0 + } else if y > 0.5 { + - 16.0 * (y - 1.0) + } else /* 0 ≤ y ≤ 0.5 */ { + 48.0 * y - 16.0 + } + } + + /// Computes the differential of the kernel for $n=1$ with $σ=1$. + #[inline] + fn lipschitz_diff_1d_σ1(&self) -> F { + // Maximal absolute second differential achieved at 0 by diff2_1d_σ1 analysis + 16.0 + } } impl<'a, S, const N : usize> Support<S::Type, N> for HatConv<S, N> @@ -159,21 +299,21 @@ } #[replace_float_literals(F::cast_from(literal))] -impl<'a, F : Float, R, C, const N : usize> Apply<&'a Loc<F, N>> +impl<'a, F : Float, R, C, const N : usize> Mapping<Loc<F, N>> for Convolution<CubeIndicator<R, N>, HatConv<C, N>> where R : Constant<Type=F>, C : Constant<Type=F> { - type Output = F; + type Codomain = F; #[inline] - fn apply(&self, y : &'a Loc<F, N>) -> F { + fn apply<I : Instance<Loc<F, N>>>(&self, y : I) -> F { let Convolution(ref ind, ref hatconv) = self; let β = ind.r.value(); let σ = hatconv.radius(); // This is just a product of one-dimensional versions - y.product_map(|x| { + y.cow().product_map(|x| { // With $u_σ(x) = u_1(x/σ)/σ$ the normalised hat convolution // we have // $$ @@ -188,24 +328,66 @@ } } -impl<'a, F : Float, R, C, const N : usize> Apply<Loc<F, N>> +#[replace_float_literals(F::cast_from(literal))] +impl<'a, F : Float, R, C, const N : usize> DifferentiableImpl<Loc<F, N>> for Convolution<CubeIndicator<R, N>, HatConv<C, N>> where R : Constant<Type=F>, C : Constant<Type=F> { - type Output = F; + type Derivative = Loc<F, N>; #[inline] - fn apply(&self, y : Loc<F, N>) -> F { - self.apply(&y) + fn differential_impl<I : Instance<Loc<F, N>>>(&self, y0 : I) -> Loc<F, N> { + let y = y0.cow(); + let Convolution(ref ind, ref hatconv) = self; + let β = ind.r.value(); + let σ = hatconv.radius(); + let σ2 = σ * σ; + + let vs = y.map(|x| { + self.value_1d_σ1(x / σ, β / σ) + }); + product_differential(&*y, &vs, |x| { + self.diff_1d_σ1(x / σ, β / σ) / σ2 + }) } } +/// Integrate $f$, whose support is $[c, d]$, on $[a, b]$. +/// If $b > d$, add $g()$ to the result. +#[inline] +#[replace_float_literals(F::cast_from(literal))] +fn i<F: Float>(a : F, b : F, c : F, d : F, f : impl Fn(F) -> F, + g : impl Fn() -> F) -> F { + if b < c { + 0.0 + } else if b <= d { + if a <= c { + f(b) - f(c) + } else { + f(b) - f(a) + } + } else /* b > d */ { + g() + if a <= c { + f(d) - f(c) + } else if a < d { + f(d) - f(a) + } else { + 0.0 + } + } +} + #[replace_float_literals(F::cast_from(literal))] impl<F : Float, C, R, const N : usize> Convolution<CubeIndicator<R, N>, HatConv<C, N>> where R : Constant<Type=F>, C : Constant<Type=F> { + + /// Calculates the value of the 1D hat convolution further convolved by a interval indicator. + /// As both functions are piecewise polynomials, this is implemented by explicit integral over + /// all subintervals of polynomiality of the cube indicator, using easily formed + /// antiderivatives. #[inline] pub fn value_1d_σ1(&self, x : F, β : F) -> F { // The integration interval @@ -218,34 +400,10 @@ y * y } - /// Integrate $f$, whose support is $[c, d]$, on $[a, b]$. - /// If $b > d$, add $g()$ to the result. - #[inline] - fn i<F: Float>(a : F, b : F, c : F, d : F, f : impl Fn(F) -> F, - g : impl Fn() -> F) -> F { - if b < c { - 0.0 - } else if b <= d { - if a <= c { - f(b) - f(c) - } else { - f(b) - f(a) - } - } else /* b > d */ { - g() + if a <= c { - f(d) - f(c) - } else if a < d { - f(d) - f(a) - } else { - 0.0 - } - } - } - // Observe the factor 1/6 at the front from the antiderivatives below. // The factor 4 is from normalisation of the original function. (4.0/6.0) * i(a, b, -1.0, -0.5, - // (2/3) (y+1)^3 on -1 < y ≤ - 1/2 + // (2/3) (y+1)^3 on -1 < y ≤ -1/2 // The antiderivative is (2/12)(y+1)^4 = (1/6)(y+1)^4 |y| pow4(y+1.0), || i(a, b, -0.5, 0.0, @@ -266,8 +424,53 @@ ) ) } + + /// Calculates the derivative of the 1D hat convolution further convolved by a interval + /// indicator. The implementation is similar to [`Self::value_1d_σ1`], using the fact that + /// $(θ * ψ)' = θ * ψ'$. + #[inline] + pub fn diff_1d_σ1(&self, x : F, β : F) -> F { + // The integration interval + let a = x - β; + let b = x + β; + + // The factor 4 is from normalisation of the original function. + 4.0 * i(a, b, -1.0, -0.5, + // (2/3) (y+1)^3 on -1 < y ≤ -1/2 + |y| (2.0/3.0) * (y + 1.0).powi(3), + || i(a, b, -0.5, 0.0, + // -2 y^3 - 2 y^2 + 1/3 on -1/2 < y ≤ 0 + |y| -2.0*(y + 1.0) * y * y + (1.0/3.0), + || i(a, b, 0.0, 0.5, + // 2 y^3 - 2 y^2 + 1/3 on 0 < y < 1/2 + |y| 2.0*(y - 1.0) * y * y + (1.0/3.0), + || i(a, b, 0.5, 1.0, + // -(2/3) (y-1)^3 on 1/2 < y ≤ 1 + |y| -(2.0/3.0) * (y - 1.0).powi(3), + || 0.0 + ) + ) + ) + ) + } } +/* +impl<'a, F : Float, R, C, const N : usize> Lipschitz<L2> +for Differential<Loc<F, N>, Convolution<CubeIndicator<R, N>, HatConv<C, N>>> +where R : Constant<Type=F>, + C : Constant<Type=F> { + + type FloatType = F; + + #[inline] + fn lipschitz_factor(&self, _l2 : L2) -> Option<F> { + dbg!("unimplemented"); + None + } +} +*/ + impl<F : Float, R, C, const N : usize> Convolution<CubeIndicator<R, N>, HatConv<C, N>> where R : Constant<Type=F>, @@ -409,7 +612,7 @@ #[cfg(test)] mod tests { use alg_tools::lingrid::linspace; - use alg_tools::mapping::Apply; + use alg_tools::mapping::Mapping; use alg_tools::norms::Linfinity; use alg_tools::loc::Loc; use crate::kernels::{BallIndicator, CubeIndicator, Convolution};
--- /dev/null Thu Jan 01 00:00:00 1970 +0000 +++ b/src/kernels/linear.rs Mon Feb 17 13:54:53 2025 -0500 @@ -0,0 +1,94 @@ +//! Implementation of the linear function + +use numeric_literals::replace_float_literals; +use serde::Serialize; +use alg_tools::types::*; +use alg_tools::norms::*; +use alg_tools::loc::Loc; +use alg_tools::sets::Cube; +use alg_tools::bisection_tree::{ + Support, + Bounds, + LocalAnalysis, + GlobalAnalysis, + Bounded, +}; +use alg_tools::mapping::{Mapping, Instance}; +use alg_tools::maputil::array_init; +use alg_tools::euclidean::Euclidean; + +/// Representation of the hat function $f(x)=1-\\|x\\|\_1/ε$ of `width` $ε$ on $ℝ^N$. +#[derive(Copy,Clone,Serialize,Debug,Eq,PartialEq)] +pub struct Linear<F : Float, const N : usize> { + /// The parameter $ε>0$. + pub v : Loc<F, N>, +} + +#[replace_float_literals(F::cast_from(literal))] +impl<F : Float, const N : usize> Mapping<Loc<F, N>> for Linear<F, N> { + type Codomain = F; + + #[inline] + fn apply<I : Instance<Loc<F, N>>>(&self, x : I) -> Self::Codomain { + x.eval(|x| self.v.dot(x)) + } +} + + +#[replace_float_literals(F::cast_from(literal))] +impl<'a, F : Float, const N : usize> Support<F, N> for Linear<F, N> { + #[inline] + fn support_hint(&self) -> Cube<F,N> { + array_init(|| [F::NEG_INFINITY, F::INFINITY]).into() + } + + #[inline] + fn in_support(&self, _x : &Loc<F,N>) -> bool { + true + } + + /*fn fully_in_support(&self, _cube : &Cube<F,N>) -> bool { + todo!("Not implemented, but not used at the moment") + }*/ + + #[inline] + fn bisection_hint(&self, _cube : &Cube<F,N>) -> [Option<F>; N] { + [None; N] + } +} + + +#[replace_float_literals(F::cast_from(literal))] +impl<'a, F : Float, const N : usize> +GlobalAnalysis<F, Bounds<F>> +for Linear<F, N> { + #[inline] + fn global_analysis(&self) -> Bounds<F> { + Bounds(F::NEG_INFINITY, F::INFINITY) + } +} + +impl<'a, F : Float, const N : usize> +LocalAnalysis<F, Bounds<F>, N> +for Linear<F, N> { + #[inline] + fn local_analysis(&self, cube : &Cube<F, N>) -> Bounds<F> { + let (lower, upper) = cube.iter_corners() + .map(|x| self.apply(x)) + .fold((F::INFINITY, F::NEG_INFINITY), |(lower, upper), v| { + (lower.min(v), upper.max(v)) + }); + Bounds(lower, upper) + } +} + +#[replace_float_literals(F::cast_from(literal))] +impl<'a, F : Float, const N : usize> +Norm<F, Linfinity> +for Linear<F, N> { + #[inline] + fn norm(&self, _ : Linfinity) -> F { + self.bounds().upper() + } +} +
--- a/src/kernels/mollifier.rs Tue Aug 01 10:25:09 2023 +0300 +++ b/src/kernels/mollifier.rs Mon Feb 17 13:54:53 2025 -0500 @@ -2,7 +2,7 @@ //! Implementation of the standard mollifier use rgsl::hypergeometric::hyperg_U; -use float_extras::f64::{tgamma as gamma}; +use float_extras::f64::tgamma as gamma; use numeric_literals::replace_float_literals; use serde::Serialize; use alg_tools::types::*; @@ -17,7 +17,7 @@ LocalAnalysis, GlobalAnalysis }; -use alg_tools::mapping::Apply; +use alg_tools::mapping::{Mapping, Instance}; use alg_tools::maputil::array_init; /// Reresentation of the (unnormalised) standard mollifier. @@ -36,13 +36,14 @@ } #[replace_float_literals(C::Type::cast_from(literal))] -impl<'a, C : Constant, const N : usize> Apply<&'a Loc<C::Type, N>> for Mollifier<C, N> { - type Output = C::Type; +impl<C : Constant, const N : usize> Mapping<Loc<C::Type, N>> for Mollifier<C, N> { + type Codomain = C::Type; + #[inline] - fn apply(&self, x : &'a Loc<C::Type, N>) -> Self::Output { + fn apply<I : Instance<Loc<C::Type, N>>>(&self, x : I) -> Self::Codomain { let ε = self.width.value(); let ε2 = ε*ε; - let n2 = x.norm2_squared(); + let n2 = x.eval(|x| x.norm2_squared()); if n2 < ε2 { (n2 / (n2 - ε2)).exp() } else { @@ -51,13 +52,6 @@ } } -impl<C : Constant, const N : usize> Apply<Loc<C::Type, N>> for Mollifier<C, N> { - type Output = C::Type; - #[inline] - fn apply(&self, x : Loc<C::Type, N>) -> Self::Output { - self.apply(&x) - } -} impl<'a, C : Constant, const N : usize> Support<C::Type, N> for Mollifier<C, N> { #[inline]
--- a/src/main.rs Tue Aug 01 10:25:09 2023 +0300 +++ b/src/main.rs Mon Feb 17 13:54:53 2025 -0500 @@ -10,12 +10,13 @@ // Linear operators may be written e.g. as `opA`, to keep the capital letters of mathematical // convention while referring to the type (trait) of the operator as `A`. #![allow(non_snake_case)] -// We need the drain filter for inertial prune. -#![feature(drain_filter)] +// Need to create parse errors +#![feature(dec2flt)] use clap::Parser; use serde::{Serialize, Deserialize}; use serde_json; +use serde_with::skip_serializing_none; use itertools::Itertools; use std::num::NonZeroUsize; @@ -30,11 +31,17 @@ pub mod kernels; pub mod seminorms; pub mod forward_model; +pub mod preadjoint_helper; pub mod plot; pub mod subproblem; pub mod tolerance; pub mod regularisation; +pub mod dataterm; +pub mod prox_penalty; pub mod fb; +pub mod sliding_fb; +pub mod sliding_pdps; +pub mod forward_pdps; pub mod frank_wolfe; pub mod pdps; pub mod run; @@ -49,12 +56,12 @@ AlgorithmConfig, }; use experiments::DefaultExperiment; -use measures::merging::SpikeMergingMethod; use DefaultExperiment::*; use DefaultAlgorithm::*; /// Command line parameters -#[derive(Parser, Debug, Serialize)] +#[skip_serializing_none] +#[derive(Parser, Debug, Serialize, Default, Clone)] #[clap( about = env!("CARGO_PKG_DESCRIPTION"), author = env!("CARGO_PKG_AUTHORS"), @@ -89,7 +96,7 @@ /// Not all algorithms are available for all the experiments. /// In particular, only PDPS is available for the experiments with L¹ data term. #[arg(value_enum, value_name = "ALGORITHM", long, short = 'a', - default_values_t = [FB, FISTA, PDPS, FW, FWRelax])] + default_values_t = [FB, PDPS, SlidingFB, FW, RadonFB])] algorithm : Vec<DefaultAlgorithm>, /// Saved algorithm configration(s) to use on the experiments @@ -112,6 +119,10 @@ /// Number of threads. Overrides the maximum number. num_threads : Option<usize>, + #[arg(long, default_value_t = false)] + /// Load saved value ranges (if exists) to do partial update. + load_valuerange : bool, + #[clap(flatten, next_help_heading = "Experiment overrides")] /// Experiment setup overrides experiment_overrides : ExperimentOverrides<float>, @@ -122,7 +133,8 @@ } /// Command line experiment setup overrides -#[derive(Parser, Debug, Serialize, Deserialize)] +#[skip_serializing_none] +#[derive(Parser, Debug, Serialize, Deserialize, Default, Clone)] pub struct ExperimentOverrides<F : ClapFloat> { #[arg(long)] /// Regularisation parameter override. @@ -145,7 +157,8 @@ } /// Command line algorithm parametrisation overrides -#[derive(Parser, Debug, Serialize, Deserialize)] +#[skip_serializing_none] +#[derive(Parser, Debug, Serialize, Deserialize, Default, Clone)] pub struct AlgorithmOverrides<F : ClapFloat> { #[arg(long, value_names = &["COUNT", "EACH"])] /// Override bootstrap insertion iterations for --algorithm. @@ -162,21 +175,44 @@ tau0 : Option<F>, #[arg(long, requires = "algorithm")] + /// Second primal step length parameter override for SlidingPDPS. + /// + /// Only use if running just a single algorithm, as different algorithms have different + /// regularisation parameters. + sigmap0 : Option<F>, + + #[arg(long, requires = "algorithm")] /// Dual step length parameter override for --algorithm. /// /// Only use if running just a single algorithm, as different algorithms have different /// regularisation parameters. Only affects PDPS. sigma0 : Option<F>, + #[arg(long)] + /// Normalised transport step length for sliding methods. + theta0 : Option<F>, + + #[arg(long)] + /// A posteriori transport tolerance multiplier (C_pos) + transport_tolerance_pos : Option<F>, + + #[arg(long)] + /// Transport adaptation factor. Must be in (0, 1). + transport_adaptation : Option<F>, + + #[arg(long)] + /// Minimal step length parameter for sliding methods. + tau0_min : Option<F>, + #[arg(value_enum, long)] /// PDPS acceleration, when available. acceleration : Option<pdps::Acceleration>, - #[arg(long)] - /// Perform postprocess weight optimisation for saved iterations - /// - /// Only affects FB, FISTA, and PDPS. - postprocessing : Option<bool>, + // #[arg(long)] + // /// Perform postprocess weight optimisation for saved iterations + // /// + // /// Only affects FB, FISTA, and PDPS. + // postprocessing : Option<bool>, #[arg(value_name = "n", long)] /// Merging frequency, if merging enabled (every n iterations) @@ -184,18 +220,26 @@ /// Only affects FB, FISTA, and PDPS. merge_every : Option<usize>, - #[arg(value_enum, long)]//, value_parser = SpikeMergingMethod::<float>::value_parser())] - /// Merging strategy - /// - /// Either the string "none", or a radius value for heuristic merging. - merging : Option<SpikeMergingMethod<F>>, + #[arg(long)] + /// Enable merging (default: determined by algorithm) + merge : Option<bool>, + + #[arg(long)] + /// Merging radius (default: determined by experiment) + merge_radius : Option<F>, - #[arg(value_enum, long)]//, value_parser = SpikeMergingMethod::<float>::value_parser())] - /// Final merging strategy - /// - /// Either the string "none", or a radius value for heuristic merging. - /// Only affects FB, FISTA, and PDPS. - final_merging : Option<SpikeMergingMethod<F>>, + #[arg(long)] + /// Interpolate when merging (default : determined by algorithm) + merge_interp : Option<bool>, + + #[arg(long)] + /// Enable final merging (default: determined by algorithm) + final_merging : Option<bool>, + + #[arg(long)] + /// Enable fitness-based merging for relevant FB-type methods. + /// This has worse convergence guarantees that merging based on optimality conditions. + fitness_merging : Option<bool>, #[arg(long, value_names = &["ε", "θ", "p"])] /// Set the tolerance to ε_k = ε/(1+θk)^p @@ -230,9 +274,15 @@ for experiment_shorthand in cli.experiments.iter().unique() { let experiment = experiment_shorthand.get_experiment(&cli.experiment_overrides).unwrap(); let mut algs : Vec<Named<AlgorithmConfig<float>>> - = cli.algorithm.iter() - .map(|alg| experiment.algorithm_defaults(*alg, &cli.algoritm_overrides)) - .collect(); + = cli.algorithm + .iter() + .map(|alg| { + let cfg = alg.default_config() + .cli_override(&experiment.algorithm_overrides(*alg)) + .cli_override(&cli.algoritm_overrides); + alg.to_named(cfg) + }) + .collect(); for filename in cli.saved_algorithm.iter() { let f = std::fs::File::open(filename).unwrap(); let alg = serde_json::from_reader(f).unwrap();
--- a/src/measures.rs Tue Aug 01 10:25:09 2023 +0300 +++ b/src/measures.rs Mon Feb 17 13:54:53 2025 -0500 @@ -7,3 +7,4 @@ mod discrete; pub use discrete::*; pub mod merging; +
--- a/src/measures/base.rs Tue Aug 01 10:25:09 2023 +0300 +++ b/src/measures/base.rs Mon Feb 17 13:54:53 2025 -0500 @@ -16,3 +16,6 @@ type Domain; } +/// Decomposition of measures +pub struct MeasureDecomp; +
--- a/src/measures/delta.rs Tue Aug 01 10:25:09 2023 +0300 +++ b/src/measures/delta.rs Mon Feb 17 13:54:53 2025 -0500 @@ -7,8 +7,9 @@ use crate::types::*; use std::ops::{Div, Mul, DivAssign, MulAssign, Neg}; use serde::ser::{Serialize, Serializer, SerializeStruct}; -use alg_tools::norms::{Norm, Dist}; -use alg_tools::linops::{Apply, Linear}; +use alg_tools::norms::Norm; +use alg_tools::linops::{Mapping, Linear}; +use alg_tools::instance::{Instance, Space}; /// Representation of a delta measure. /// @@ -50,43 +51,50 @@ } -impl<Domain : PartialEq, F : Float> Measure<F> for DeltaMeasure<Domain, F> { +impl<Domain, F : Float> Measure<F> for DeltaMeasure<Domain, F> { type Domain = Domain; } -impl<Domain : PartialEq, F : Float> Norm<F, Radon> for DeltaMeasure<Domain, F> { +impl<Domain, F : Float> Norm<F, Radon> for DeltaMeasure<Domain, F> { #[inline] fn norm(&self, _ : Radon) -> F { self.α.abs() } } -impl<Domain : PartialEq, F : Float> Dist<F, Radon> for DeltaMeasure<Domain, F> { +// impl<Domain : PartialEq, F : Float> Dist<F, Radon> for DeltaMeasure<Domain, F> { +// #[inline] +// fn dist(&self, other : &Self, _ : Radon) -> F { +// if self.x == other. x { +// (self.α - other.α).abs() +// } else { +// self.α.abs() + other.α.abs() +// } +// } +// } + +impl<Domain, G, F : Num> Mapping<G> for DeltaMeasure<Domain, F> +where + Domain : Space, + G::Codomain : Mul<F, Output=G::Codomain>, + G : Mapping<Domain> + Clone + Space, + for<'b> &'b Domain : Instance<Domain>, +{ + type Codomain = G::Codomain; + #[inline] - fn dist(&self, other : &Self, _ : Radon) -> F { - if self.x == other. x { - (self.α - other.α).abs() - } else { - self.α.abs() + other.α.abs() - } + fn apply<I : Instance<G>>(&self, g : I) -> Self::Codomain { + g.eval(|g̃| g̃.apply(&self.x) * self.α) } } -impl<'b, Domain, G, F : Num, V : Mul<F, Output=V>> Apply<G> for DeltaMeasure<Domain, F> -where G: for<'a> Apply<&'a Domain, Output = V>, - V : Mul<F> { - type Output = V; - - #[inline] - fn apply(&self, g : G) -> Self::Output { - g.apply(&self.x) * self.α - } -} - -impl<Domain, G, F : Num, V : Mul<F, Output=V>> Linear<G> for DeltaMeasure<Domain, F> -where G: for<'a> Apply<&'a Domain, Output = V> { - type Codomain = V; -} +impl<Domain, G, F : Num> Linear<G> for DeltaMeasure<Domain, F> +where + Domain : Space, + G::Codomain : Mul<F, Output=G::Codomain>, + G : Mapping<Domain> + Clone + Space, + for<'b> &'b Domain : Instance<Domain>, +{ } // /// Partial blanket implementation of [`DeltaMeasure`] as a linear functional of [`Mapping`]s. // /// A full blanket implementation is not possible due to annoying Rust limitations: only [`Apply`] @@ -141,12 +149,13 @@ } } -/*impl<F : Num> From<(F, F)> for DeltaMeasure<Loc<F, 1>, F> { +impl<'a, Domain : Clone, F : Num> From<&'a DeltaMeasure<Domain, F>> for DeltaMeasure<Domain, F> { #[inline] - fn from((x, α) : (F, F)) -> Self { - DeltaMeasure{x: Loc([x]), α: α} + fn from(d : &'a DeltaMeasure<Domain, F>) -> Self { + d.clone() } -}*/ +} + impl<Domain, F : Num> DeltaMeasure<Domain, F> { /// Set the mass of the spike. @@ -186,6 +195,26 @@ } } +impl<Domain, F : Num> IntoIterator for DeltaMeasure<Domain, F> { + type Item = Self; + type IntoIter = std::iter::Once<Self>; + + #[inline] + fn into_iter(self) -> Self::IntoIter { + std::iter::once(self) + } +} + +impl<'a, Domain, F : Num> IntoIterator for &'a DeltaMeasure<Domain, F> { + type Item = Self; + type IntoIter = std::iter::Once<Self>; + + #[inline] + fn into_iter(self) -> Self::IntoIter { + std::iter::once(self) + } +} + macro_rules! make_delta_scalarop_rhs { ($trait:ident, $fn:ident, $trait_assign:ident, $fn_assign:ident) => {
--- a/src/measures/discrete.rs Tue Aug 01 10:25:09 2023 +0300 +++ b/src/measures/discrete.rs Mon Feb 17 13:54:53 2025 -0500 @@ -11,9 +11,11 @@ use alg_tools::norms::Norm; use alg_tools::tabledump::TableDump; -use alg_tools::linops::{Apply, Linear}; +use alg_tools::linops::{Mapping, Linear}; use alg_tools::iter::{MapF,Mappable}; use alg_tools::nalgebra_support::ToNalgebraRealField; +use alg_tools::collection::Collection; +use alg_tools::instance::{Instance, Decomposition, MyCow, EitherDecomp, Space}; use crate::types::*; use super::base::*; @@ -29,6 +31,8 @@ pub(super) spikes : Vec<DeltaMeasure<Domain, F>>, } +pub type RNDM<F, const N : usize> = DiscreteMeasure<Loc<F, N>, F>; + /// Iterator over the [`DeltaMeasure`] spikes of a [`DiscreteMeasure`]. pub type SpikeIter<'a, Domain, F> = std::slice::Iter<'a, DeltaMeasure<Domain, F>>; @@ -59,6 +63,20 @@ self.spikes.len() } + /// Replace with the zero measure. + #[inline] + pub fn clear(&mut self) { + self.spikes.clear() + } + + /// Remove `i`:th spike, not maintaining order. + /// + /// Panics if indiex is out of bounds. + #[inline] + pub fn swap_remove(&mut self, i : usize) -> DeltaMeasure<Domain, F>{ + self.spikes.swap_remove(i) + } + /// Iterate over (references to) the [`DeltaMeasure`] spikes in this measure #[inline] pub fn iter_spikes(&self) -> SpikeIter<'_, Domain, F> { @@ -95,6 +113,13 @@ self.spikes.iter_mut().zip(iter).for_each(|(δ, α)| δ.set_mass(α)); } + /// Update the locations of all the spikes to those produced by an iterator. + #[inline] + pub fn set_locations<'a, I : Iterator<Item=&'a Domain>>(&mut self, iter : I) + where Domain : 'static + Clone { + self.spikes.iter_mut().zip(iter.cloned()).for_each(|(δ, α)| δ.set_location(α)); + } + // /// Map the masses of all the spikes using a function and an iterator // #[inline] // pub fn zipmap_masses< @@ -107,50 +132,162 @@ /// Prune all spikes with zero mass. #[inline] pub fn prune(&mut self) { - self.spikes.retain(|δ| δ.α != F::ZERO); + self.prune_by(|δ| δ.α != F::ZERO); + } + + /// Prune spikes by the predicate `g`. + #[inline] + pub fn prune_by<G : FnMut(&DeltaMeasure<Domain, F>) -> bool>(&mut self, g : G) { + self.spikes.retain(g); + } + + /// Add the spikes produced by `iter` to this measure. + #[inline] + pub fn extend<I : Iterator<Item=DeltaMeasure<Domain, F>>>( + &mut self, + iter : I + ) { + self.spikes.extend(iter); + } + + /// Add a spike to the measure + #[inline] + pub fn push(&mut self, δ : DeltaMeasure<Domain, F>) { + self.spikes.push(δ); + } + + /// Iterate over triples of masses and locations of two discrete measures, which are assumed + /// to have equal locations of same spike indices. + pub fn both_matching<'a>(&'a self, other : &'a DiscreteMeasure<Domain, F>) -> + impl Iterator<Item=(F, F, &'a Domain)> { + let m = self.len().max(other.len()); + self.iter_spikes().map(Some).chain(std::iter::repeat(None)) + .zip(other.iter_spikes().map(Some).chain(std::iter::repeat(None))) + .take(m) + .map(|(oδ, orδ)| { + match (oδ, orδ) { + (Some(δ), Some(rδ)) => (δ.α, rδ.α, &δ.x), // Assumed δ.x=rδ.x + (Some(δ), None) => (δ.α, F::ZERO, &δ.x), + (None, Some(rδ)) => (F::ZERO, rδ.α, &rδ.x), + (None, None) => panic!("This cannot happen!"), + } + }) + } + + /// Subtract `other` from `self`, assuming equal locations of same spike indices + pub fn sub_matching(&self, other : &DiscreteMeasure<Domain, F>) -> DiscreteMeasure<Domain, F> + where Domain : Clone { + self.both_matching(other) + .map(|(α, β, x)| (x.clone(), α - β)) + .collect() + } + + /// Add `other` to `self`, assuming equal locations of same spike indices + pub fn add_matching(&self, other : &DiscreteMeasure<Domain, F>) -> DiscreteMeasure<Domain, F> + where Domain : Clone { + self.both_matching(other) + .map(|(α, β, x)| (x.clone(), α + β)) + .collect() + } + + /// Calculate the Radon-norm distance of `self` to `other`, + /// assuming equal locations of same spike indices. + pub fn dist_matching(&self, other : &DiscreteMeasure<Domain, F>) -> F where F : Float { + self.both_matching(other) + .map(|(α, β, _)| (α-β).abs()) + .sum() + } +} + +impl<Domain, F : Num> IntoIterator for DiscreteMeasure<Domain, F> { + type Item = DeltaMeasure<Domain, F>; + type IntoIter = std::vec::IntoIter<DeltaMeasure<Domain, F>>; + + #[inline] + fn into_iter(self) -> Self::IntoIter { + self.spikes.into_iter() + } +} + +impl<'a, Domain, F : Num> IntoIterator for &'a DiscreteMeasure<Domain, F> { + type Item = &'a DeltaMeasure<Domain, F>; + type IntoIter = SpikeIter<'a, Domain, F>; + + #[inline] + fn into_iter(self) -> Self::IntoIter { + self.spikes.iter() + } +} + +impl<Domain, F : Num> Sum<DeltaMeasure<Domain, F>> for DiscreteMeasure<Domain, F> { + // Required method + fn sum<I>(iter: I) -> Self + where + I : Iterator<Item = DeltaMeasure<Domain, F>> + { + Self::from_iter(iter) + } +} + +impl<'a, Domain : Clone, F : Num> Sum<&'a DeltaMeasure<Domain, F>> + for DiscreteMeasure<Domain, F> +{ + // Required method + fn sum<I>(iter: I) -> Self + where + I : Iterator<Item = &'a DeltaMeasure<Domain, F>> + { + Self::from_iter(iter.cloned()) + } +} + +impl<Domain, F : Num> Sum<DiscreteMeasure<Domain, F>> for DiscreteMeasure<Domain, F> { + // Required method + fn sum<I>(iter: I) -> Self + where + I : Iterator<Item = DiscreteMeasure<Domain, F>> + { + Self::from_iter(iter.map(|μ| μ.into_iter()).flatten()) + } +} + +impl<'a, Domain : Clone, F : Num> Sum<&'a DiscreteMeasure<Domain, F>> + for DiscreteMeasure<Domain, F> +{ + // Required method + fn sum<I>(iter: I) -> Self + where + I : Iterator<Item = &'a DiscreteMeasure<Domain, F>> + { + Self::from_iter(iter.map(|μ| μ.iter_spikes()).flatten().cloned()) } } impl<Domain : Clone, F : Float> DiscreteMeasure<Domain, F> { /// Computes `μ1 ← θ * μ1 - ζ * μ2`, pruning entries where both `μ1` (`self`) and `μ2` have - // zero weight. `μ2` will contain copy of pruned original `μ1` without arithmetic performed. - /// **This expects `self` and `μ2` to have matching coordinates in each index**. + // zero weight. `μ2` will contain a pruned copy of pruned original `μ1` without arithmetic + /// performed. **This expects `self` and `μ2` to have matching coordinates in each index**. // `μ2` can be than `self`, but not longer. pub fn pruning_sub(&mut self, θ : F, ζ : F, μ2 : &mut Self) { - let mut μ2_get = 0; - let mut μ2_insert = 0; - self.spikes.drain_filter(|&mut DeltaMeasure{ α : ref mut α_ref, ref x }| { - // Get weight of spike in μ2, zero if out of bounds. - let β = μ2.spikes.get(μ2_get).map_or(F::ZERO, DeltaMeasure::get_mass); - μ2_get += 1; - - if *α_ref == F::ZERO && β == F::ZERO { - // Prune - true + for δ in &self[μ2.len()..] { + μ2.push(DeltaMeasure{ x : δ.x.clone(), α : F::ZERO}); + } + debug_assert_eq!(self.len(), μ2.len()); + let mut dest = 0; + for i in 0..self.len() { + let α = self[i].α; + let α_new = θ * α - ζ * μ2[i].α; + if dest < i { + μ2[dest] = DeltaMeasure{ x : self[i].x.clone(), α }; + self[dest] = DeltaMeasure{ x : self[i].x.clone(), α : α_new }; } else { - // Save self weight - let α = *α_ref; - // Modify self - *α_ref = θ * α - ζ * β; - // Make copy of old self weight in μ2 - let δ = DeltaMeasure{ α, x : x.clone() }; - match μ2.spikes.get_mut(μ2_insert) { - Some(replace) => { - *replace = δ; - }, - None => { - debug_assert_eq!(μ2.len(), μ2_insert); - μ2.spikes.push(δ); - }, - } - μ2_insert += 1; - // Keep - false + μ2[i].α = α; + self[i].α = α_new; } - }); - // Truncate μ2 to same length as self. - μ2.spikes.truncate(μ2_insert); - debug_assert_eq!(μ2.len(), self.len()); + dest += 1; + } + self.spikes.truncate(dest); + μ2.spikes.truncate(dest); } } @@ -174,23 +311,61 @@ pub fn set_masses_dvector(&mut self, x : &DVector<F::MixedType>) { self.set_masses(x.iter().map(|&α| F::from_nalgebra_mixed(α))); } + + // /// Extracts the masses of the spikes as a [`Vec`]. + // pub fn masses_vec(&self) -> Vec<F::MixedType> { + // self.iter_masses() + // .map(|α| α.to_nalgebra_mixed()) + // .collect() + // } + + // /// Sets the masses of the spikes from the values of a [`Vec`]. + // pub fn set_masses_vec(&mut self, x : &Vec<F::MixedType>) { + // self.set_masses(x.iter().map(|&α| F::from_nalgebra_mixed(α))); + // } } -impl<Domain, F :Num> Index<usize> for DiscreteMeasure<Domain, F> { - type Output = DeltaMeasure<Domain, F>; +// impl<Domain, F :Num> Index<usize> for DiscreteMeasure<Domain, F> { +// type Output = DeltaMeasure<Domain, F>; +// #[inline] +// fn index(&self, i : usize) -> &Self::Output { +// self.spikes.index(i) +// } +// } + +// impl<Domain, F :Num> IndexMut<usize> for DiscreteMeasure<Domain, F> { +// #[inline] +// fn index_mut(&mut self, i : usize) -> &mut Self::Output { +// self.spikes.index_mut(i) +// } +// } + +impl< + Domain, + F : Num, + I : std::slice::SliceIndex<[DeltaMeasure<Domain, F>]> +> Index<I> +for DiscreteMeasure<Domain, F> { + type Output = <I as std::slice::SliceIndex<[DeltaMeasure<Domain, F>]>>::Output; #[inline] - fn index(&self, i : usize) -> &Self::Output { + fn index(&self, i : I) -> &Self::Output { self.spikes.index(i) } } -impl<Domain, F :Num> IndexMut<usize> for DiscreteMeasure<Domain, F> { +impl< + Domain, + F : Num, + I : std::slice::SliceIndex<[DeltaMeasure<Domain, F>]> +> IndexMut<I> +for DiscreteMeasure<Domain, F> { #[inline] - fn index_mut(&mut self, i : usize) -> &mut Self::Output { + fn index_mut(&mut self, i : I) -> &mut Self::Output { self.spikes.index_mut(i) } } + impl<Domain, F : Num, D : Into<DeltaMeasure<Domain, F>>, const K : usize> From<[D; K]> for DiscreteMeasure<Domain, F> { #[inline] @@ -199,6 +374,45 @@ } } +impl<Domain, F : Num> From<Vec<DeltaMeasure<Domain, F>>> +for DiscreteMeasure<Domain, F> { + #[inline] + fn from(spikes : Vec<DeltaMeasure<Domain, F>>) -> Self { + DiscreteMeasure{ spikes } + } +} + +impl<'a, Domain, F : Num, D> From<&'a [D]> +for DiscreteMeasure<Domain, F> +where &'a D : Into<DeltaMeasure<Domain, F>> { + #[inline] + fn from(list : &'a [D]) -> Self { + list.into_iter().map(|d| d.into()).collect() + } +} + + +impl<Domain, F : Num> From<DeltaMeasure<Domain, F>> +for DiscreteMeasure<Domain, F> { + #[inline] + fn from(δ : DeltaMeasure<Domain, F>) -> Self { + DiscreteMeasure{ + spikes : vec!(δ) + } + } +} + +impl<'a, Domain : Clone, F : Num> From<&'a DeltaMeasure<Domain, F>> +for DiscreteMeasure<Domain, F> { + #[inline] + fn from(δ : &'a DeltaMeasure<Domain, F>) -> Self { + DiscreteMeasure{ + spikes : vec!(δ.clone()) + } + } +} + + impl<Domain, F : Num, D : Into<DeltaMeasure<Domain, F>>> FromIterator<D> for DiscreteMeasure<Domain, F> { #[inline] @@ -258,19 +472,28 @@ } } -impl<Domain, G, F : Num, Y : Sum + Mul<F, Output=Y>> Apply<G> for DiscreteMeasure<Domain, F> -where G: for<'a> Apply<&'a Domain, Output = Y> { - type Output = Y; +impl<Domain, G, F : Num> Mapping<G> for DiscreteMeasure<Domain, F> +where + Domain : Space, + G::Codomain : Sum + Mul<F, Output=G::Codomain>, + G : Mapping<Domain, Codomain=F> + Clone + Space, + for<'b> &'b Domain : Instance<Domain>, +{ + type Codomain = G::Codomain; + #[inline] - fn apply(&self, g : G) -> Y { - self.spikes.iter().map(|m| g.apply(&m.x) * m.α).sum() + fn apply<I : Instance<G>>(&self, g : I) -> Self::Codomain { + g.eval(|g| self.spikes.iter().map(|m| g.apply(&m.x) * m.α).sum()) } } -impl<Domain, G, F : Num, Y : Sum + Mul<F, Output=Y>> Linear<G> for DiscreteMeasure<Domain, F> -where G : for<'a> Apply<&'a Domain, Output = Y> { - type Codomain = Y; -} +impl<Domain, G, F : Num> Linear<G> for DiscreteMeasure<Domain, F> +where + Domain : Space, + G::Codomain : Sum + Mul<F, Output=G::Codomain>, + G : Mapping<Domain, Codomain=F> + Clone + Space, + for<'b> &'b Domain : Instance<Domain>, +{ } /// Helper trait for constructing arithmetic operations for combinations @@ -278,6 +501,7 @@ trait Lift<F : Num, Domain> { type Producer : Iterator<Item=DeltaMeasure<Domain, F>>; + #[allow(dead_code)] /// Lifts `self` into a [`DiscreteMeasure`]. fn lift(self) -> DiscreteMeasure<Domain, F>; @@ -574,3 +798,217 @@ make_discrete_scalarop_lhs!(Mul, mul; f32 f64 i8 i16 i32 i64 isize u8 u16 u32 u64 usize); make_discrete_scalarop_lhs!(Div, div; f32 f64 i8 i16 i32 i64 isize u8 u16 u32 u64 usize); + +impl<F : Num, Domain> Collection for DiscreteMeasure<Domain, F> { + type Element = DeltaMeasure<Domain, F>; + type RefsIter<'a> = std::slice::Iter<'a, Self::Element> where Self : 'a; + + #[inline] + fn iter_refs(&self) -> Self::RefsIter<'_> { + self.iter_spikes() + } +} + +impl<Domain : Clone, F : Num> Space for DiscreteMeasure<Domain, F> { + type Decomp = MeasureDecomp; +} + +pub type SpikeSlice<'b, Domain, F> = &'b [DeltaMeasure<Domain, F>]; + +pub type EitherSlice<'b, Domain, F> = EitherDecomp< + Vec<DeltaMeasure<Domain, F>>, + SpikeSlice<'b, Domain, F> +>; + +impl<F : Num, Domain : Clone> Decomposition<DiscreteMeasure<Domain, F>> for MeasureDecomp { + type Decomposition<'b> = EitherSlice<'b, Domain, F> where DiscreteMeasure<Domain, F> : 'b; + type Reference<'b> = SpikeSlice<'b, Domain, F> where DiscreteMeasure<Domain, F> : 'b; + + /// Left the lightweight reference type into a full decomposition type. + fn lift<'b>(r : Self::Reference<'b>) -> Self::Decomposition<'b> { + EitherDecomp::Borrowed(r) + } +} + +impl<F : Num, Domain : Clone> Instance<DiscreteMeasure<Domain, F>, MeasureDecomp> +for DiscreteMeasure<Domain, F> +{ + fn decompose<'b>(self) + -> <MeasureDecomp as Decomposition<DiscreteMeasure<Domain, F>>>::Decomposition<'b> + where Self : 'b, DiscreteMeasure<Domain, F> : 'b { + EitherDecomp::Owned(self.spikes) + } + + fn ref_instance(&self) + -> <MeasureDecomp as Decomposition<DiscreteMeasure<Domain, F>>>::Reference<'_> + { + self.spikes.as_slice() + } + + fn cow<'b>(self) -> MyCow<'b, DiscreteMeasure<Domain, F>> where Self : 'b { + MyCow::Owned(self) + } + + fn own(self) -> DiscreteMeasure<Domain, F> { + self + } +} + +impl<'a, F : Num, Domain : Clone> Instance<DiscreteMeasure<Domain, F>, MeasureDecomp> +for &'a DiscreteMeasure<Domain, F> +{ + fn decompose<'b>(self) + -> <MeasureDecomp as Decomposition<DiscreteMeasure<Domain, F>>>::Decomposition<'b> + where Self : 'b, DiscreteMeasure<Domain, F> : 'b { + EitherDecomp::Borrowed(self.spikes.as_slice()) + } + + fn ref_instance(&self) + -> <MeasureDecomp as Decomposition<DiscreteMeasure<Domain, F>>>::Reference<'_> + { + self.spikes.as_slice() + } + + fn cow<'b>(self) -> MyCow<'b, DiscreteMeasure<Domain, F>> where Self : 'b { + MyCow::Borrowed(self) + } + + fn own(self) -> DiscreteMeasure<Domain, F> { + self.clone() + } +} + +impl<'a, F : Num, Domain : Clone> Instance<DiscreteMeasure<Domain, F>, MeasureDecomp> +for EitherSlice<'a, Domain, F> +{ + fn decompose<'b>(self) + -> <MeasureDecomp as Decomposition<DiscreteMeasure<Domain, F>>>::Decomposition<'b> + where Self : 'b, DiscreteMeasure<Domain, F> : 'b { + self + } + + fn ref_instance(&self) + -> <MeasureDecomp as Decomposition<DiscreteMeasure<Domain, F>>>::Reference<'_> + { + match self { + EitherDecomp::Owned(v) => v.as_slice(), + EitherDecomp::Borrowed(s) => s, + } + } + + fn own(self) -> DiscreteMeasure<Domain, F> { + match self { + EitherDecomp::Owned(v) => v.into(), + EitherDecomp::Borrowed(s) => s.into(), + } + } +} + +impl<'a, F : Num, Domain : Clone> Instance<DiscreteMeasure<Domain, F>, MeasureDecomp> +for &'a EitherSlice<'a, Domain, F> +{ + fn decompose<'b>(self) + -> <MeasureDecomp as Decomposition<DiscreteMeasure<Domain, F>>>::Decomposition<'b> + where Self : 'b, DiscreteMeasure<Domain, F> : 'b { + match self { + EitherDecomp::Owned(v) => EitherDecomp::Borrowed(v.as_slice()), + EitherDecomp::Borrowed(s) => EitherDecomp::Borrowed(s), + } + } + + fn ref_instance(&self) + -> <MeasureDecomp as Decomposition<DiscreteMeasure<Domain, F>>>::Reference<'_> + { + match self { + EitherDecomp::Owned(v) => v.as_slice(), + EitherDecomp::Borrowed(s) => s, + } + } + + fn own(self) -> DiscreteMeasure<Domain, F> { + match self { + EitherDecomp::Owned(v) => v.as_slice(), + EitherDecomp::Borrowed(s) => s + }.into() + } +} + +impl<'a, F : Num, Domain : Clone> Instance<DiscreteMeasure<Domain, F>, MeasureDecomp> +for SpikeSlice<'a, Domain, F> +{ + fn decompose<'b>(self) + -> <MeasureDecomp as Decomposition<DiscreteMeasure<Domain, F>>>::Decomposition<'b> + where Self : 'b, DiscreteMeasure<Domain, F> : 'b { + EitherDecomp::Borrowed(self) + } + + fn ref_instance(&self) + -> <MeasureDecomp as Decomposition<DiscreteMeasure<Domain, F>>>::Reference<'_> + { + self + } + + fn own(self) -> DiscreteMeasure<Domain, F> { + self.into() + } +} + +impl<'a, F : Num, Domain : Clone> Instance<DiscreteMeasure<Domain, F>, MeasureDecomp> +for &'a SpikeSlice<'a, Domain, F> +{ + fn decompose<'b>(self) + -> <MeasureDecomp as Decomposition<DiscreteMeasure<Domain, F>>>::Decomposition<'b> + where Self : 'b, DiscreteMeasure<Domain, F> : 'b { + EitherDecomp::Borrowed(*self) + } + + fn ref_instance(&self) + -> <MeasureDecomp as Decomposition<DiscreteMeasure<Domain, F>>>::Reference<'_> + { + *self + } + + fn own(self) -> DiscreteMeasure<Domain, F> { + (*self).into() + } +} + +impl<F : Num, Domain : Clone > Instance<DiscreteMeasure<Domain, F>, MeasureDecomp> +for DeltaMeasure<Domain, F> +{ + fn decompose<'b>(self) + -> <MeasureDecomp as Decomposition<DiscreteMeasure<Domain, F>>>::Decomposition<'b> + where Self : 'b, DiscreteMeasure<Domain, F> : 'b { + EitherDecomp::Owned(vec![self]) + } + + fn ref_instance(&self) + -> <MeasureDecomp as Decomposition<DiscreteMeasure<Domain, F>>>::Reference<'_> + { + std::slice::from_ref(self) + } + + fn own(self) -> DiscreteMeasure<Domain, F> { + self.into() + } +} + +impl<'a, F : Num, Domain : Clone> Instance<DiscreteMeasure<Domain, F>, MeasureDecomp> +for &'a DeltaMeasure<Domain, F> +{ + fn decompose<'b>(self) + -> <MeasureDecomp as Decomposition<DiscreteMeasure<Domain, F>>>::Decomposition<'b> + where Self : 'b, DiscreteMeasure<Domain, F> : 'b { + EitherDecomp::Borrowed(std::slice::from_ref(self)) + } + + fn ref_instance(&self) + -> <MeasureDecomp as Decomposition<DiscreteMeasure<Domain, F>>>::Reference<'_> + { + std::slice::from_ref(*self) + } + + fn own(self) -> DiscreteMeasure<Domain, F> { + self.into() + } +}
--- a/src/measures/merging.rs Tue Aug 01 10:25:09 2023 +0300 +++ b/src/measures/merging.rs Mon Feb 17 13:54:53 2025 -0500 @@ -7,60 +7,35 @@ */ use numeric_literals::replace_float_literals; +use serde::{Deserialize, Serialize}; use std::cmp::Ordering; -use serde::{Serialize, Deserialize}; //use clap::builder::{PossibleValuesParser, PossibleValue}; use alg_tools::nanleast::NaNLeast; -use crate::types::*; use super::delta::*; use super::discrete::*; +use crate::types::*; /// Spike merging heuristic selection #[derive(Clone, Copy, Eq, PartialEq, Serialize, Deserialize, Debug)] #[allow(dead_code)] -pub enum SpikeMergingMethod<F> { - /// Try to merge spikes within a given radius of eachother - HeuristicRadius(F), - /// No merging - None, -} - -// impl<F : Float> SpikeMergingMethod<F> { -// /// This is for [`clap`] to display command line help. -// pub fn value_parser() -> PossibleValuesParser { -// PossibleValuesParser::new([ -// PossibleValue::new("none").help("No merging"), -// PossibleValue::new("<radius>").help("Heuristic merging within indicated radius") -// ]) -// } -// } - -impl<F : ClapFloat> std::fmt::Display for SpikeMergingMethod<F> { - fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> Result<(), std::fmt::Error> { - match self { - Self::None => write!(f, "none"), - Self::HeuristicRadius(r) => std::fmt::Display::fmt(r, f), - } - } -} - -impl<F : ClapFloat> std::str::FromStr for SpikeMergingMethod<F> { - type Err = F::Err; - - fn from_str(s: &str) -> Result<Self, Self::Err> { - if s == "none" { - Ok(Self::None) - } else { - Ok(Self::HeuristicRadius(F::from_str(s)?)) - } - } +pub struct SpikeMergingMethod<F> { + // Merging radius + pub(crate) radius: F, + // Enabled + pub(crate) enabled: bool, + // Interpolate merged points + pub(crate) interp: bool, } #[replace_float_literals(F::cast_from(literal))] -impl<F : Float> Default for SpikeMergingMethod<F> { +impl<F: Float> Default for SpikeMergingMethod<F> { fn default() -> Self { - SpikeMergingMethod::HeuristicRadius(0.02) + SpikeMergingMethod { + radius: 0.01, + enabled: false, + interp: true, + } } } @@ -75,13 +50,18 @@ /// an arbitrary value. This method will return that value for the *last* accepted merge, or /// [`None`] if no merge was accepted. /// - /// This method is stable with respect to spike locations: on merge, the weight of existing - /// spikes is set to zero, and a new one inserted at the end of the spike vector. - fn merge_spikes<G, V>(&mut self, method : SpikeMergingMethod<F>, accept : G) -> Option<V> - where G : Fn(&'_ Self) -> Option<V> { - match method { - SpikeMergingMethod::HeuristicRadius(ρ) => self.do_merge_spikes_radius(ρ, accept), - SpikeMergingMethod::None => None, + /// This method is stable with respect to spike locations: on merge, the weights of existing + /// removed spikes is set to zero, new ones inserted at the end of the spike vector. + /// They merge may also be performed by increasing the weights of the existing spikes, + /// without inserting new spikes. + fn merge_spikes<G>(&mut self, method: SpikeMergingMethod<F>, accept: G) -> usize + where + G: FnMut(&'_ Self) -> bool, + { + if method.enabled { + self.do_merge_spikes_radius(method.radius, method.interp, accept) + } else { + 0 } } @@ -90,35 +70,37 @@ /// Calls [`SpikeMerging::merge_spikes`] with `accept` constructed from the composition of /// `value` and `fitness`, compared to initial fitness. Returns the last return value of `value` // for a merge accepted by `fitness`. If no merge was accepted, `value` applied to the initial - /// `self` is returned. + /// `self` is returned. also the number of merges is returned; fn merge_spikes_fitness<G, H, V, O>( &mut self, - method : SpikeMergingMethod<F>, - value : G, - fitness : H - ) -> V - where G : Fn(&'_ Self) -> V, - H : Fn(&'_ V) -> O, - O : PartialOrd { - let initial_res = value(self); - let initial_fitness = fitness(&initial_res); - self.merge_spikes(method, |μ| { - let res = value(μ); - (fitness(&res) <= initial_fitness).then_some(res) - }).unwrap_or(initial_res) + method: SpikeMergingMethod<F>, + value: G, + fitness: H, + ) -> (V, usize) + where + G: Fn(&'_ Self) -> V, + H: Fn(&'_ V) -> O, + O: PartialOrd, + { + let mut res = value(self); + let initial_fitness = fitness(&res); + let count = self.merge_spikes(method, |μ| { + res = value(μ); + fitness(&res) <= initial_fitness + }); + (res, count) } /// Attempt to merge spikes that are within radius $ρ$ of each other (unspecified norm). /// - /// This method implements [`SpikeMerging::merge_spikes`] for - /// [`SpikeMergingMethod::HeuristicRadius`]. The closure `accept` and the return value are - /// as for that method. - fn do_merge_spikes_radius<G, V>(&mut self, ρ : F, accept : G) -> Option<V> - where G : Fn(&'_ Self) -> Option<V>; + /// This method implements [`SpikeMerging::merge_spikes`]. + fn do_merge_spikes_radius<G>(&mut self, ρ: F, interp: bool, accept: G) -> usize + where + G: FnMut(&'_ Self) -> bool; } #[replace_float_literals(F::cast_from(literal))] -impl<F : Float, const N : usize> DiscreteMeasure<Loc<F, N>, F> { +impl<F: Float, const N: usize> DiscreteMeasure<Loc<F, N>, F> { /// Attempts to merge spikes with indices `i` and `j`. /// /// This assumes that the weights of the two spikes have already been checked not to be zero. @@ -126,78 +108,72 @@ /// The parameter `res` points to the current “result” for [`SpikeMerging::merge_spikes`]. /// If the merge is accepted by `accept` returning a [`Some`], `res` will be replaced by its /// return value. - fn attempt_merge<G, V>( + /// + /// Returns the index of `self.spikes` storing the new spike. + fn attempt_merge<G>( &mut self, - res : &mut Option<V>, - i : usize, - j : usize, - accept : &G - ) -> bool - where G : Fn(&'_ Self) -> Option<V> { - let &DeltaMeasure{ x : xi, α : αi } = &self.spikes[i]; - let &DeltaMeasure{ x : xj, α : αj } = &self.spikes[j]; + i: usize, + j: usize, + interp: bool, + accept: &mut G, + ) -> Option<usize> + where + G: FnMut(&'_ Self) -> bool, + { + let &DeltaMeasure { x: xi, α: αi } = &self.spikes[i]; + let &DeltaMeasure { x: xj, α: αj } = &self.spikes[j]; - // Merge inplace - self.spikes[i].α = 0.0; - self.spikes[j].α = 0.0; - //self.spikes.push(DeltaMeasure{ α : αi + αj, x : (xi + xj)/2.0 }); - self.spikes.push(DeltaMeasure{ α : αi + αj, x : (xi * αi + xj * αj) / (αi + αj) }); - match accept(self) { - some@Some(..) => { - // Merge accepted, update our return value - *res = some; - // On next iteration process the newly merged spike. - //indices[k+1] = self.spikes.len() - 1; - true - }, - None => { + if interp { + // Merge inplace + self.spikes[i].α = 0.0; + self.spikes[j].α = 0.0; + let αia = αi.abs(); + let αja = αj.abs(); + self.spikes.push(DeltaMeasure { + α: αi + αj, + x: (xi * αia + xj * αja) / (αia + αja), + }); + if accept(self) { + Some(self.spikes.len() - 1) + } else { // Merge not accepted, restore modification self.spikes[i].α = αi; self.spikes[j].α = αj; self.spikes.pop(); - false + None + } + } else { + // Attempt merge inplace, first combination + self.spikes[i].α = αi + αj; + self.spikes[j].α = 0.0; + if accept(self) { + // Merge accepted + Some(i) + } else { + // Attempt merge inplace, second combination + self.spikes[i].α = 0.0; + self.spikes[j].α = αi + αj; + if accept(self) { + // Merge accepted + Some(j) + } else { + // Merge not accepted, restore modification + self.spikes[i].α = αi; + self.spikes[j].α = αj; + None + } } } } - - /* - /// Attempts to merge spikes with indices i and j, acceptance through a delta. - fn attempt_merge_change<G, V>( - &mut self, - res : &mut Option<V>, - i : usize, - j : usize, - accept_change : &G - ) -> bool - where G : Fn(&'_ Self) -> Option<V> { - let &DeltaMeasure{ x : xi, α : αi } = &self.spikes[i]; - let &DeltaMeasure{ x : xj, α : αj } = &self.spikes[j]; - let δ = DeltaMeasure{ α : αi + αj, x : (xi + xj)/2.0 }; - let λ = [-self.spikes[i], -self.spikes[j], δ.clone()].into(); - - match accept_change(&λ) { - some@Some(..) => { - // Merge accepted, update our return value - *res = some; - self.spikes[i].α = 0.0; - self.spikes[j].α = 0.0; - self.spikes.push(δ); - true - }, - None => { - false - } - } - }*/ - } /// Sorts a vector of indices into `slice` by `compare`. /// /// The closure `compare` operators on references to elements of `slice`. /// Returns the sorted vector of indices into `slice`. -pub fn sort_indices_by<V, F>(slice : &[V], mut compare : F) -> Vec<usize> -where F : FnMut(&V, &V) -> Ordering +pub fn sort_indices_by<V, F>(slice: &[V], mut compare: F) -> Vec<usize> +where + F: FnMut(&V, &V) -> Ordering, { let mut indices = Vec::from_iter(0..slice.len()); indices.sort_by(|&i, &j| compare(&slice[i], &slice[j])); @@ -205,14 +181,11 @@ } #[replace_float_literals(F::cast_from(literal))] -impl<F : Float> SpikeMerging<F> for DiscreteMeasure<Loc<F, 1>, F> { - - fn do_merge_spikes_radius<G, V>( - &mut self, - ρ : F, - accept : G - ) -> Option<V> - where G : Fn(&'_ Self) -> Option<V> { +impl<F: Float> SpikeMerging<F> for DiscreteMeasure<Loc<F, 1>, F> { + fn do_merge_spikes_radius<G>(&mut self, ρ: F, interp: bool, mut accept: G) -> usize + where + G: FnMut(&'_ Self) -> bool, + { // Sort by coordinate into an indexing array. let mut indices = sort_indices_by(&self.spikes, |&δ1, &δ2| { let &Loc([x1]) = &δ1.x; @@ -222,34 +195,43 @@ }); // Initialise result - let mut res = None; + let mut count = 0; // Scan consecutive pairs and merge if close enough and accepted by `accept`. if indices.len() == 0 { - return res + return count; } - for k in 0..(indices.len()-1) { + for k in 0..(indices.len() - 1) { let i = indices[k]; - let j = indices[k+1]; - let &DeltaMeasure{ x : Loc([xi]), α : αi } = &self.spikes[i]; - let &DeltaMeasure{ x : Loc([xj]), α : αj } = &self.spikes[j]; + let j = indices[k + 1]; + let &DeltaMeasure { + x: Loc([xi]), + α: αi, + } = &self.spikes[i]; + let &DeltaMeasure { + x: Loc([xj]), + α: αj, + } = &self.spikes[j]; debug_assert!(xi <= xj); // If close enough, attempt merging if αi != 0.0 && αj != 0.0 && xj <= xi + ρ { - if self.attempt_merge(&mut res, i, j, &accept) { - indices[k+1] = self.spikes.len() - 1; + if let Some(l) = self.attempt_merge(i, j, interp, &mut accept) { + // For this to work (the debug_assert! to not trigger above), the new + // coordinate produced by attempt_merge has to be at most xj. + indices[k + 1] = l; + count += 1 } } } - res + count } } /// Orders `δ1` and `δ1` according to the first coordinate. -fn compare_first_coordinate<F : Float>( - δ1 : &DeltaMeasure<Loc<F, 2>, F>, - δ2 : &DeltaMeasure<Loc<F, 2>, F> +fn compare_first_coordinate<F: Float>( + δ1: &DeltaMeasure<Loc<F, 2>, F>, + δ2: &DeltaMeasure<Loc<F, 2>, F>, ) -> Ordering { let &Loc([x11, ..]) = &δ1.x; let &Loc([x21, ..]) = &δ2.x; @@ -258,28 +240,32 @@ } #[replace_float_literals(F::cast_from(literal))] -impl<F : Float> SpikeMerging<F> for DiscreteMeasure<Loc<F, 2>, F> { - - fn do_merge_spikes_radius<G, V>(&mut self, ρ : F, accept : G) -> Option<V> - where G : Fn(&'_ Self) -> Option<V> { +impl<F: Float> SpikeMerging<F> for DiscreteMeasure<Loc<F, 2>, F> { + fn do_merge_spikes_radius<G>(&mut self, ρ: F, interp: bool, mut accept: G) -> usize + where + G: FnMut(&'_ Self) -> bool, + { // Sort by first coordinate into an indexing array. let mut indices = sort_indices_by(&self.spikes, compare_first_coordinate); // Initialise result - let mut res = None; + let mut count = 0; let mut start_scan_2nd = 0; // Scan in order if indices.len() == 0 { - return res + return count; } - for k in 0..indices.len()-1 { + for k in 0..indices.len() - 1 { let i = indices[k]; - let &DeltaMeasure{ x : Loc([xi1, xi2]), α : αi } = &self[i]; + let &DeltaMeasure { + x: Loc([xi1, xi2]), + α: αi, + } = &self[i]; if αi == 0.0 { // Nothin to be done if the weight is already zero - continue + continue; } let mut closest = None; @@ -289,57 +275,59 @@ // the _closest_ mergeable spike might have index less than `k` in `indices`, and a // merge with it might have not been attempted with this spike if a different closer // spike was discovered based on the second coordinate. - 'scan_2nd: for l in (start_scan_2nd+1)..indices.len() { + 'scan_2nd: for l in (start_scan_2nd + 1)..indices.len() { if l == k { // Do not attempt to merge a spike with itself - continue + continue; } let j = indices[l]; - let &DeltaMeasure{ x : Loc([xj1, xj2]), α : αj } = &self[j]; + let &DeltaMeasure { + x: Loc([xj1, xj2]), + α: αj, + } = &self[j]; if xj1 < xi1 - ρ { // Spike `j = indices[l]` has too low first coordinate. Update starting index // for next iteration, and continue scanning. start_scan_2nd = l; - continue 'scan_2nd + continue 'scan_2nd; } else if xj1 > xi1 + ρ { // Break out: spike `j = indices[l]` has already too high first coordinate, no // more close enough spikes can be found due to the sorting of `indices`. - break 'scan_2nd + break 'scan_2nd; } // If also second coordinate is close enough, attempt merging if closer than // previously discovered mergeable spikes. - let d2 = (xi2-xj2).abs(); + let d2 = (xi2 - xj2).abs(); if αj != 0.0 && d2 <= ρ { - let r1 = xi1-xj1; - let d = (d2*d2 + r1*r1).sqrt(); + let r1 = xi1 - xj1; + let d = (d2 * d2 + r1 * r1).sqrt(); match closest { None => closest = Some((l, j, d)), Some((_, _, r)) if r > d => closest = Some((l, j, d)), - _ => {}, + _ => {} } } } // Attempt merging closest close-enough spike if let Some((l, j, _)) = closest { - if self.attempt_merge(&mut res, i, j, &accept) { + if let Some(n) = self.attempt_merge(i, j, interp, &mut accept) { // If merge was succesfull, make new spike candidate for merging. - indices[l] = self.spikes.len() - 1; - let compare = |i, j| compare_first_coordinate(&self.spikes[i], - &self.spikes[j]); + indices[l] = n; + count += 1; + let compare = |i, j| compare_first_coordinate(&self.spikes[i], &self.spikes[j]); // Re-sort relevant range of indices if l < k { indices[l..k].sort_by(|&i, &j| compare(i, j)); } else { - indices[k+1..=l].sort_by(|&i, &j| compare(i, j)); + indices[k + 1..=l].sort_by(|&i, &j| compare(i, j)); } } } } - res + count } } -
--- a/src/pdps.rs Tue Aug 01 10:25:09 2023 +0300 +++ b/src/pdps.rs Mon Feb 17 13:54:53 2025 -0500 @@ -6,8 +6,7 @@ * Valkonen T. - _Proximal methods for point source localisation_, [arXiv:2212.02991](https://arxiv.org/abs/2212.02991). -The main routine is [`pointsource_pdps`]. It is based on specilisatinn of -[`generic_pointsource_fb_reg`] through relevant [`FBSpecialisation`] implementations. +The main routine is [`pointsource_pdps_reg`]. Both norm-2-squared and norm-1 data terms are supported. That is, implemented are solvers for <div> $$ @@ -37,10 +36,6 @@ For $F_0(y)=\frac{1}{2}\|y\|_2^2$ the second part reads $y = Aμ -b$. For $F_0(y)=\|y\|_1$ the second part reads $y ∈ ∂\|·\|_1(Aμ - b)$. </p> - -Based on zero initialisation for $μ$, we use the [`Subdifferentiable`] trait to make an -initialisation corresponding to the second part of the optimality conditions. -In the algorithm itself, standard proximal steps are taking with respect to $F\_0^* + ⟨b, ·⟩$. */ use numeric_literals::replace_float_literals; @@ -48,37 +43,23 @@ use nalgebra::DVector; use clap::ValueEnum; -use alg_tools::iterate:: AlgIteratorFactory; -use alg_tools::sets::Cube; -use alg_tools::loc::Loc; +use alg_tools::iterate::AlgIteratorFactory; use alg_tools::euclidean::Euclidean; +use alg_tools::linops::Mapping; use alg_tools::norms::{ - L1, Linfinity, - Projection, Norm, + Linfinity, + Projection, }; -use alg_tools::bisection_tree::{ - BTFN, - PreBTFN, - Bounds, - BTNodeLookup, - BTNode, - BTSearch, - P2Minimise, - SupportGenerator, - LocalAnalysis, -}; -use alg_tools::mapping::RealMapping; +use alg_tools::mapping::{RealMapping, Instance}; use alg_tools::nalgebra_support::ToNalgebraRealField; use alg_tools::linops::AXPY; use crate::types::*; -use crate::measures::DiscreteMeasure; -use crate::measures::merging::{ - SpikeMerging, -}; -use crate::forward_model::ForwardModel; -use crate::seminorms::{ - DiscreteMeasureOp, Lipschitz +use crate::measures::{DiscreteMeasure, RNDM}; +use crate::measures::merging::SpikeMerging; +use crate::forward_model::{ + ForwardModel, + AdjointProductBoundedBy, }; use crate::plot::{ SeqPlotter, @@ -86,12 +67,21 @@ PlotLookup }; use crate::fb::{ + postprocess, + prune_with_stats +}; +pub use crate::prox_penalty::{ FBGenericConfig, - FBSpecialisation, - generic_pointsource_fb_reg, - RegTerm, + ProxPenalty }; -use crate::regularisation::NonnegRadonRegTerm; +use crate::regularisation::RegTerm; +use crate::dataterm::{ + DataTerm, + L2Squared, + L1 +}; +use crate::measures::merging::SpikeMergingMethod; + /// Acceleration #[derive(Clone, Copy, Eq, PartialEq, Serialize, Deserialize, ValueEnum, Debug)] @@ -107,7 +97,30 @@ Full } -/// Settings for [`pointsource_pdps`]. +#[replace_float_literals(F::cast_from(literal))] +impl Acceleration { + /// PDPS parameter acceleration. Updates τ and σ and returns ω. + /// This uses dual strong convexity, not primal. + fn accelerate<F : Float>(self, τ : &mut F, σ : &mut F, γ : F) -> F { + match self { + Acceleration::None => 1.0, + Acceleration::Partial => { + let ω = 1.0 / (1.0 + γ * (*σ)).sqrt(); + *σ *= ω; + *τ /= ω; + ω + }, + Acceleration::Full => { + let ω = 1.0 / (1.0 + 2.0 * γ * (*σ)).sqrt(); + *σ *= ω; + *τ /= ω; + ω + }, + } + } +} + +/// Settings for [`pointsource_pdps_reg`]. #[derive(Clone, Copy, Eq, PartialEq, Serialize, Deserialize, Debug)] #[serde(default)] pub struct PDPSConfig<F : Float> { @@ -118,176 +131,77 @@ /// Accelerate if available pub acceleration : Acceleration, /// Generic parameters - pub insertion : FBGenericConfig<F>, + pub generic : FBGenericConfig<F>, } #[replace_float_literals(F::cast_from(literal))] impl<F : Float> Default for PDPSConfig<F> { fn default() -> Self { - let τ0 = 0.5; + let τ0 = 5.0; PDPSConfig { τ0, σ0 : 0.99/τ0, acceleration : Acceleration::Partial, - insertion : Default::default() + generic : FBGenericConfig { + merging : SpikeMergingMethod { enabled : true, ..Default::default() }, + .. Default::default() + }, } } } -/// Trait for subdifferentiable objects -pub trait Subdifferentiable<F : Float, V, U=V> { - /// Calculate some subdifferential at `x` - fn some_subdifferential(&self, x : V) -> U; +/// Trait for data terms for the PDPS +#[replace_float_literals(F::cast_from(literal))] +pub trait PDPSDataTerm<F : Float, V, const N : usize> : DataTerm<F, V, N> { + /// Calculate some subdifferential at `x` for the conjugate + fn some_subdifferential(&self, x : V) -> V; + + /// Factor of strong convexity of the conjugate + #[inline] + fn factor_of_strong_convexity(&self) -> F { + 0.0 + } + + /// Perform dual update + fn dual_update(&self, _y : &mut V, _y_prev : &V, _σ : F); } -/// Type for indicating norm-2-squared data fidelity. -pub struct L2Squared; -impl<F : Float, V : Euclidean<F>> Subdifferentiable<F, V> for L2Squared { +#[replace_float_literals(F::cast_from(literal))] +impl<F, V, const N : usize> PDPSDataTerm<F, V, N> +for L2Squared +where + F : Float, + V : Euclidean<F> + AXPY<F>, + for<'b> &'b V : Instance<V>, +{ fn some_subdifferential(&self, x : V) -> V { x } + + fn factor_of_strong_convexity(&self) -> F { + 1.0 + } + + #[inline] + fn dual_update(&self, y : &mut V, y_prev : &V, σ : F) { + y.axpy(1.0 / (1.0 + σ), y_prev, σ / (1.0 + σ)); + } } -impl<F : Float + nalgebra::RealField> Subdifferentiable<F, DVector<F>> for L1 { +#[replace_float_literals(F::cast_from(literal))] +impl<F : Float + nalgebra::RealField, const N : usize> +PDPSDataTerm<F, DVector<F>, N> +for L1 { fn some_subdifferential(&self, mut x : DVector<F>) -> DVector<F> { // nalgebra sucks for providing second copies of the same stuff that's elsewhere as well. x.iter_mut() .for_each(|v| if *v != F::ZERO { *v = *v/<F as NumTraitsFloat>::abs(*v) }); x } -} -/// Specialisation of [`generic_pointsource_fb_reg`] to PDPS. -pub struct PDPS< - 'a, - F : Float + ToNalgebraRealField, - A : ForwardModel<Loc<F, N>, F>, - D, - const N : usize -> { - /// The data - b : &'a A::Observable, - /// The forward operator - opA : &'a A, - /// Primal step length - τ : F, - // Dual step length - σ : F, - /// Whether acceleration should be applied (if data term supports) - acceleration : Acceleration, - /// The dataterm. Only used by the type system. - _dataterm : D, - /// Previous dual iterate. - y_prev : A::Observable, -} - -/// Implementation of [`FBSpecialisation`] for μPDPS with norm-2-squared data fidelity. -#[replace_float_literals(F::cast_from(literal))] -impl< - 'a, - F : Float + ToNalgebraRealField, - A : ForwardModel<Loc<F, N>, F>, - const N : usize -> FBSpecialisation<F, A::Observable, N> for PDPS<'a, F, A, L2Squared, N> -where for<'b> &'b A::Observable : std::ops::Add<A::Observable, Output=A::Observable> { - - fn update( - &mut self, - μ : &mut DiscreteMeasure<Loc<F, N>, F>, - μ_base : &DiscreteMeasure<Loc<F, N>, F> - ) -> (A::Observable, Option<F>) { - let σ = self.σ; - let τ = self.τ; - let ω = match self.acceleration { - Acceleration::None => 1.0, - Acceleration::Partial => { - let ω = 1.0 / (1.0 + σ).sqrt(); - self.σ = σ * ω; - self.τ = τ / ω; - ω - }, - Acceleration::Full => { - let ω = 1.0 / (1.0 + 2.0 * σ).sqrt(); - self.σ = σ * ω; - self.τ = τ / ω; - ω - }, - }; - - μ.prune(); - - let mut y = self.b.clone(); - self.opA.gemv(&mut y, 1.0 + ω, μ, -1.0); - self.opA.gemv(&mut y, -ω, μ_base, 1.0); - y.axpy(1.0 / (1.0 + σ), &self.y_prev, σ / (1.0 + σ)); - self.y_prev.copy_from(&y); - - (y, Some(self.τ)) - } - - fn calculate_fit( - &self, - μ : &DiscreteMeasure<Loc<F, N>, F>, - _y : &A::Observable - ) -> F { - self.calculate_fit_simple(μ) - } - - fn calculate_fit_simple( - &self, - μ : &DiscreteMeasure<Loc<F, N>, F>, - ) -> F { - let mut residual = self.b.clone(); - self.opA.gemv(&mut residual, 1.0, μ, -1.0); - residual.norm2_squared_div2() - } -} - -/// Implementation of [`FBSpecialisation`] for μPDPS with norm-1 data fidelity. -#[replace_float_literals(F::cast_from(literal))] -impl< - 'a, - F : Float + ToNalgebraRealField, - A : ForwardModel<Loc<F, N>, F>, - const N : usize -> FBSpecialisation<F, A::Observable, N> for PDPS<'a, F, A, L1, N> -where A::Observable : Projection<F, Linfinity> + Norm<F, L1>, - for<'b> &'b A::Observable : std::ops::Add<A::Observable, Output=A::Observable> { - fn update( - &mut self, - μ : &mut DiscreteMeasure<Loc<F, N>, F>, - μ_base : &DiscreteMeasure<Loc<F, N>, F> - ) -> (A::Observable, Option<F>) { - let σ = self.σ; - - μ.prune(); - - //let ȳ = self.opA.apply(μ) * 2.0 - self.opA.apply(μ_base); - //*y = proj_{[-1,1]}(&self.y_prev + (ȳ - self.b) * σ) - let mut y = self.y_prev.clone(); - self.opA.gemv(&mut y, 2.0 * σ, μ, 1.0); - self.opA.gemv(&mut y, -σ, μ_base, 1.0); - y.axpy(-σ, self.b, 1.0); + #[inline] + fn dual_update(&self, y : &mut DVector<F>, y_prev : &DVector<F>, σ : F) { + y.axpy(1.0, y_prev, σ); y.proj_ball_mut(1.0, Linfinity); - self.y_prev.copy_from(&y); - - (y, None) - } - - fn calculate_fit( - &self, - μ : &DiscreteMeasure<Loc<F, N>, F>, - _y : &A::Observable - ) -> F { - self.calculate_fit_simple(μ) - } - - fn calculate_fit_simple( - &self, - μ : &DiscreteMeasure<Loc<F, N>, F>, - ) -> F { - let mut residual = self.b.clone(); - self.opA.gemv(&mut residual, 1.0, μ, -1.0); - residual.norm(L1) } } @@ -304,93 +218,106 @@ /// /// Returns the final iterate. #[replace_float_literals(F::cast_from(literal))] -pub fn pointsource_pdps_reg<'a, F, I, A, GA, 𝒟, BTA, G𝒟, S, K, D, Reg, const N : usize>( - opA : &'a A, - b : &'a A::Observable, +pub fn pointsource_pdps_reg<F, I, A, D, Reg, P, const N : usize>( + opA : &A, + b : &A::Observable, reg : Reg, - op𝒟 : &'a 𝒟, - config : &PDPSConfig<F>, + prox_penalty : &P, + pdpsconfig : &PDPSConfig<F>, iterator : I, - plotter : SeqPlotter<F, N>, + mut plotter : SeqPlotter<F, N>, dataterm : D, -) -> DiscreteMeasure<Loc<F, N>, F> -where F : Float + ToNalgebraRealField, - I : AlgIteratorFactory<IterInfo<F, N>>, - for<'b> &'b A::Observable : std::ops::Neg<Output=A::Observable> - + std::ops::Add<A::Observable, Output=A::Observable>, - //+ std::ops::Mul<F, Output=A::Observable>, // <-- FIXME: compiler overflow - A::Observable : std::ops::MulAssign<F>, - GA : SupportGenerator<F, N, SupportType = S, Id = usize> + Clone, - A : ForwardModel<Loc<F, N>, F, PreadjointCodomain = BTFN<F, GA, BTA, N>> - + Lipschitz<𝒟, FloatType=F>, - BTA : BTSearch<F, N, Data=usize, Agg=Bounds<F>>, - G𝒟 : SupportGenerator<F, N, SupportType = K, Id = usize> + Clone, - 𝒟 : DiscreteMeasureOp<Loc<F, N>, F, PreCodomain = PreBTFN<F, G𝒟, N>>, - 𝒟::Codomain : RealMapping<F, N>, - S: RealMapping<F, N> + LocalAnalysis<F, Bounds<F>, N>, - K: RealMapping<F, N> + LocalAnalysis<F, Bounds<F>, N>, - BTNodeLookup: BTNode<F, usize, Bounds<F>, N>, - PlotLookup : Plotting<N>, - DiscreteMeasure<Loc<F, N>, F> : SpikeMerging<F>, - PDPS<'a, F, A, D, N> : FBSpecialisation<F, A::Observable, N>, - D : Subdifferentiable<F, A::Observable>, - Reg : RegTerm<F, N> { +) -> RNDM<F, N> +where + F : Float + ToNalgebraRealField, + I : AlgIteratorFactory<IterInfo<F, N>>, + A : ForwardModel<RNDM<F, N>, F> + + AdjointProductBoundedBy<RNDM<F, N>, P, FloatType=F>, + A::PreadjointCodomain : RealMapping<F, N>, + for<'b> &'b A::Observable : std::ops::Neg<Output=A::Observable> + Instance<A::Observable>, + PlotLookup : Plotting<N>, + RNDM<F, N> : SpikeMerging<F>, + D : PDPSDataTerm<F, A::Observable, N>, + Reg : RegTerm<F, N>, + P : ProxPenalty<F, A::PreadjointCodomain, Reg, N>, +{ + + // Check parameters + assert!(pdpsconfig.τ0 > 0.0 && + pdpsconfig.σ0 > 0.0 && + pdpsconfig.τ0 * pdpsconfig.σ0 <= 1.0, + "Invalid step length parameters"); + + // Set up parameters + let config = &pdpsconfig.generic; + let l = opA.adjoint_product_bound(prox_penalty).unwrap().sqrt(); + let mut τ = pdpsconfig.τ0 / l; + let mut σ = pdpsconfig.σ0 / l; + let γ = dataterm.factor_of_strong_convexity(); + + // We multiply tolerance by τ for FB since our subproblems depending on tolerances are scaled + // by τ compared to the conditional gradient approach. + let tolerance = config.tolerance * τ * reg.tolerance_scaling(); + let mut ε = tolerance.initial(); + + // Initialise iterates + let mut μ = DiscreteMeasure::new(); + let mut y = dataterm.some_subdifferential(-b); + let mut y_prev = y.clone(); + let full_stats = |μ : &RNDM<F, N>, ε, stats| IterInfo { + value : dataterm.calculate_fit_op(μ, opA, b) + reg.apply(μ), + n_spikes : μ.len(), + ε, + // postprocessing: config.postprocessing.then(|| μ.clone()), + .. stats + }; + let mut stats = IterInfo::new(); - let y = dataterm.some_subdifferential(-b); - let l = opA.lipschitz_factor(&op𝒟).unwrap().sqrt(); - let τ = config.τ0 / l; - let σ = config.σ0 / l; + // Run the algorithm + for state in iterator.iter_init(|| full_stats(&μ, ε, stats.clone())) { + // Calculate smooth part of surrogate model. + let mut τv = opA.preadjoint().apply(y * τ); + + // Save current base point + let μ_base = μ.clone(); + + // Insert and reweigh + let (maybe_d, _within_tolerances) = prox_penalty.insert_and_reweigh( + &mut μ, &mut τv, &μ_base, None, + τ, ε, + config, ®, &state, &mut stats + ); + + // Prune and possibly merge spikes + if config.merge_now(&state) { + stats.merged += prox_penalty.merge_spikes_no_fitness( + &mut μ, &mut τv, &μ_base, None, τ, ε, config, ®, + ); + } + stats.pruned += prune_with_stats(&mut μ); - let pdps = PDPS { - b, - opA, - τ, - σ, - acceleration : config.acceleration, - _dataterm : dataterm, - y_prev : y.clone(), - }; + // Update step length parameters + let ω = pdpsconfig.acceleration.accelerate(&mut τ, &mut σ, γ); + + // Do dual update + y = b.clone(); // y = b + opA.gemv(&mut y, 1.0 + ω, &μ, -1.0); // y = A[(1+ω)μ^{k+1}]-b + opA.gemv(&mut y, -ω, &μ_base, 1.0); // y = A[(1+ω)μ^{k+1} - ω μ^k]-b + dataterm.dual_update(&mut y, &y_prev, σ); + y_prev.copy_from(&y); - generic_pointsource_fb_reg( - opA, reg, op𝒟, τ, &config.insertion, iterator, plotter, y, pdps - ) + // Give statistics if requested + let iter = state.iteration(); + stats.this_iters += 1; + + state.if_verbose(|| { + plotter.plot_spikes(iter, maybe_d.as_ref(), Some(&τv), &μ); + full_stats(&μ, ε, std::mem::replace(&mut stats, IterInfo::new())) + }); + + ε = tolerance.update(ε, iter); + } + + postprocess(μ, config, dataterm, opA, b) } -// -// Deprecated interfaces -// - -#[deprecated(note = "Use `pointsource_pdps_reg`")] -pub fn pointsource_pdps<'a, F, I, A, GA, 𝒟, BTA, G𝒟, S, K, D, const N : usize>( - opA : &'a A, - b : &'a A::Observable, - α : F, - op𝒟 : &'a 𝒟, - config : &PDPSConfig<F>, - iterator : I, - plotter : SeqPlotter<F, N>, - dataterm : D, -) -> DiscreteMeasure<Loc<F, N>, F> -where F : Float + ToNalgebraRealField, - I : AlgIteratorFactory<IterInfo<F, N>>, - for<'b> &'b A::Observable : std::ops::Neg<Output=A::Observable> - + std::ops::Add<A::Observable, Output=A::Observable>, - A::Observable : std::ops::MulAssign<F>, - GA : SupportGenerator<F, N, SupportType = S, Id = usize> + Clone, - A : ForwardModel<Loc<F, N>, F, PreadjointCodomain = BTFN<F, GA, BTA, N>> - + Lipschitz<𝒟, FloatType=F>, - BTA : BTSearch<F, N, Data=usize, Agg=Bounds<F>>, - G𝒟 : SupportGenerator<F, N, SupportType = K, Id = usize> + Clone, - 𝒟 : DiscreteMeasureOp<Loc<F, N>, F, PreCodomain = PreBTFN<F, G𝒟, N>>, - 𝒟::Codomain : RealMapping<F, N>, - S: RealMapping<F, N> + LocalAnalysis<F, Bounds<F>, N>, - K: RealMapping<F, N> + LocalAnalysis<F, Bounds<F>, N>, - BTNodeLookup: BTNode<F, usize, Bounds<F>, N>, - Cube<F, N>: P2Minimise<Loc<F, N>, F>, - PlotLookup : Plotting<N>, - DiscreteMeasure<Loc<F, N>, F> : SpikeMerging<F>, - PDPS<'a, F, A, D, N> : FBSpecialisation<F, A::Observable, N>, - D : Subdifferentiable<F, A::Observable> { - - pointsource_pdps_reg(opA, b, NonnegRadonRegTerm(α), op𝒟, config, iterator, plotter, dataterm) -}
--- a/src/plot.rs Tue Aug 01 10:25:09 2023 +0300 +++ b/src/plot.rs Mon Feb 17 13:54:53 2025 -0500 @@ -1,29 +1,14 @@ //! Plotting helper utilities use numeric_literals::replace_float_literals; -use std::io::Write; -use image::{ - ImageFormat, - ImageBuffer, - Rgb -}; -use itertools::izip; -use colorbrewer::Palette as CbPalette; - +use serde::Serialize; use alg_tools::types::*; use alg_tools::lingrid::LinGrid; use alg_tools::mapping::RealMapping; use alg_tools::loc::Loc; -use alg_tools::bisection_tree::Bounds; -use alg_tools::maputil::map4; use alg_tools::tabledump::write_csv; use crate::measures::*; -/// Default RGB ramp from [`colorbrewer`]. -/// -/// This is a tuple of parameters to [`colorbrewer::get_color_ramp`]. -const RAMP : (CbPalette, u32) = (CbPalette::RdBu, 11); - /// Helper trait for implementing dimension-dependent plotting routines. pub trait Plotting<const N : usize> { /// Plot several mappings and a discrete measure into a file. @@ -32,13 +17,10 @@ T1 : RealMapping<F, N>, T2 : RealMapping<F, N> > ( - g_explanation : String, - g : &T1, - ω_explanation : String, + g : Option<&T1>, ω : Option<&T2>, grid : LinGrid<F, N>, - bnd : Option<Bounds<F>>, - μ : &DiscreteMeasure<Loc<F, N>, F>, + μ : &RNDM<F, N>, filename : String, ); @@ -50,78 +32,57 @@ g : &T1, grid : LinGrid<F, N>, filename : String, - explanation : String ); } /// Helper type for looking up a [`Plotting`] based on dimension. pub struct PlotLookup; +#[derive(Serialize)] +struct CSVHelper1<F : Float> { + x : F, + f : F, +} + +#[derive(Serialize)] +struct CSVHelper1_2<F : Float>{ + x : F, + g : Option<F>, + omega : Option<F> +} + +#[derive(Serialize)] +struct CSVSpike1<F : Float> { + x : F, + alpha : F, +} + impl Plotting<1> for PlotLookup { fn plot_into_file_spikes< F : Float, T1 : RealMapping<F, 1>, T2 : RealMapping<F, 1> > ( - g_explanation : String, - g : &T1, - ω_explanation : String, + g0 : Option<&T1>, ω0 : Option<&T2>, grid : LinGrid<F, 1>, - bnd0 : Option<Bounds<F>>, μ : &DiscreteMeasure<Loc<F, 1>, F>, filename : String, ) { - let start = grid.start[0].as_(); - let end = grid.end[0].as_(); - let m = μ.iter_masses().fold(F::ZERO, |m, α| m.max(α)); - let s = μ.iter_masses().fold(F::ZERO, |m, α| m.add(α)); - let mut spike_scale = F::ONE; - - let mut plotter = poloto::plot( - "f", "x", - format!("f(x); spike max={:.4}, n={}, ∑={:.4}", m, μ.len(), s) - ).move_into(); - - if let Some(ω) = ω0 { - let graph_ω = grid.into_iter().map(|x@Loc([x0]) : Loc<F, 1>| { - [x0.as_(), ω.apply(&x).as_()] - }); - plotter.line(ω_explanation.as_str(), graph_ω.clone()); - // let csv_f = format!("{}.txt", filename); - // write_csv(graph_ω, csv_f).expect("CSV save error"); - } - - let graph_g = grid.into_iter().map(|x@Loc([x0]) : Loc<F, 1>| { - [x0.as_(), g.apply(&x).as_()] + let data = grid.into_iter().map(|p@Loc([x]) : Loc<F, 1>| CSVHelper1_2 { + x, + g : g0.map(|g| g.apply(&p)), + omega : ω0.map(|ω| ω.apply(&p)) }); - plotter.line(g_explanation.as_str(), graph_g.clone()); - // let csv_f = format!("{}.txt", filename); - // write_csv(graph_g, csv_f).expect("CSV save error"); - - bnd0.map(|bnd| { - let upperb = bnd.upper().as_(); - let lowerb = bnd.lower().as_(); - let upper : [[f64; 2]; 2] = [[start, upperb], [end, upperb]]; - let lower = [[start, lowerb], [end, lowerb]]; - spike_scale *= bnd.upper(); + let csv_f = format!("{}_functions.csv", filename); + write_csv(data, csv_f).expect("CSV save error"); - plotter.line("upper bound", upper) - .line("lower bound", lower) - .ymarker(lowerb) - .ymarker(upperb); + let spikes = μ.iter_spikes().map(|δ| { + let Loc([x]) = δ.x; + CSVSpike1 { x, alpha : δ.α } }); - - for &DeltaMeasure{ α, x : Loc([x]) } in μ.iter_spikes() { - let spike = [[x.as_(), 0.0], [x.as_(), (α/m * spike_scale).as_()]]; - plotter.line("", spike); - } - - let svg = format!("{}", poloto::disp(|a| poloto::simple_theme(a, plotter))); - - std::fs::File::create(filename + ".svg").and_then(|mut file| - file.write_all(svg.as_bytes()) - ).expect("SVG save error"); + let csv_f = format!("{}_spikes.csv", filename); + write_csv(spikes, csv_f).expect("CSV save error"); } fn plot_into_file< @@ -131,150 +92,37 @@ g : &T1, grid : LinGrid<F, 1>, filename : String, - explanation : String ) { - let graph_g = grid.into_iter().map(|x@Loc([x0]) : Loc<F, 1>| { - [x0.as_(), g.apply(&x).as_()] + let data = grid.into_iter().map(|p@Loc([x]) : Loc<F, 1>| CSVHelper1 { + x, + f : g.apply(&p), }); - - let plotter: poloto::Plotter<'_, float, float> = poloto::plot("f", "x", "f(x)") - .line(explanation.as_str(), graph_g.clone()) - .move_into(); - - let svg = format!("{}", poloto::disp(|a| poloto::simple_theme(a, plotter))); - - let svg_f = format!("{}.svg", filename); - std::fs::File::create(svg_f).and_then(|mut file| - file.write_all(svg.as_bytes()) - ).expect("SVG save error"); - let csv_f = format!("{}.txt", filename); - write_csv(graph_g, csv_f).expect("CSV save error"); + write_csv(data, csv_f).expect("CSV save error"); } } -/// Convert $[0, 1] ∈ F$ to $\\\{0, …, M\\\} ∈ F$ where $M=$`F::RANGE_MAX`. -#[inline] -fn scale_uint<F, U>(v : F) -> U -where F : Float + CastFrom<U> + num_traits::cast::AsPrimitive<U>, - U : Unsigned { - (v*F::cast_from(U::RANGE_MAX)).as_() -} - -/// Convert $[a, b] ∈ F$ to $\\\{0, …, M\\\} ∈ F$ where $M=$`F::RANGE_MAX`. -#[replace_float_literals(F::cast_from(literal))] -#[inline] -fn scale_range_uint<F, U>(v : F, &Bounds(a, b) : &Bounds<F>) -> U -where F : Float + CastFrom<U> + num_traits::cast::AsPrimitive<U>, - U : Unsigned { - debug_assert!(a < b); - scale_uint(((v - a)/(b - a)).max(0.0).min(1.0)) -} - - -/// Sample a mapping on a grid. -/// -/// Returns a vector of values as well as upper and lower bounds of the values. -fn rawdata_and_range<F, T>(grid : &LinGrid<F, 2>, g :&T) -> (Vec<F>, Bounds<F>) -where F : Float, - T : RealMapping<F, 2> { - let rawdata : Vec<F> = grid.into_iter().map(|x| g.apply(&x)).collect(); - let range = rawdata.iter() - .map(|&v| Bounds(v, v)) - .reduce(|b1, b2| b1.common(&b2)) - .unwrap(); - (rawdata, range) +#[derive(Serialize)] +struct CSVHelper2<F : Float> { + x : F, + y : F, + f : F, } -/*fn to_range<'a, F, U>(rawdata : &'a Vec<F>, range : &'a Bounds<F>) --> std::iter::Map<std::slice::Iter<'a, F>, impl FnMut(&'a F) -> U> -where F : Float + CastFrom<U> + num_traits::cast::AsPrimitive<U>, - U : Unsigned { - rawdata.iter().map(move |&v| scale_range_uint(v, range)) -}*/ - -/// Convert a scalar value to an RGB triplet. -/// -/// Converts the value `v` supposed to be within the range `[a, b]` to an rgb value according -/// to the given `ramp` of equally-spaced rgb interpolation points. -#[replace_float_literals(F::cast_from(literal))] -fn one_to_ramp<F, U>( - &Bounds(a, b) : &Bounds<F>, - ramp : &Vec<Loc<F, 3>>, - v : F, -) -> Rgb<U> -where F : Float + CastFrom<U> + num_traits::cast::AsPrimitive<U>, - U : Unsigned { - - let n = ramp.len() - 1; - let m = F::cast_from(U::RANGE_MAX); - let ramprange = move |v : F| {let m : usize = v.as_(); m.min(n).max(0) }; - - let w = F::cast_from(n) * (v - a) / (b - a); // convert [0, 1] to [0, n] - let (l, u) = (w.floor(), w.ceil()); // Find closest integers - let (rl, ru) = (ramprange(l), ramprange(u)); - let (cl, cu) = (ramp[rl], ramp[ru]); // Get corresponding colours - let λ = match rl==ru { // Interpolation factor - true => 0.0, - false => (u - w) / (u - l), - }; - let Loc(rgb) = cl * λ + cu * (1.0 - λ); // Interpolate - - Rgb(rgb.map(|v| (v * m).round().min(m).max(0.0).as_())) +#[derive(Serialize)] +struct CSVHelper2_2<F : Float>{ + x : F, + y : F, + g : Option<F>, + omega : Option<F> } -/// Convert a an iterator over scalar values to an iterator over RGB triplets. -/// -/// The conversion is that performed by [`one_to_ramp`]. -#[replace_float_literals(F::cast_from(literal))] -fn to_ramp<'a, F, U, I>( - bounds : &'a Bounds<F>, - ramp : &'a Vec<Loc<F, 3>>, - iter : I, -) -> std::iter::Map<I, impl FnMut(F) -> Rgb<U> + 'a> -where F : Float + CastFrom<U> + num_traits::cast::AsPrimitive<U>, - U : Unsigned, - I : Iterator<Item = F> + 'a { - iter.map(move |v| one_to_ramp(bounds, ramp, v)) -} - -/// Convert a [`colorbrewer`] sepcification to a ramp of rgb triplets. -fn get_ramp<F : Float>((palette, nb) : (CbPalette, u32)) -> Vec<Loc<F, 3>> { - let m = F::cast_from(u8::MAX); - colorbrewer::get_color_ramp(palette, nb) - .expect("Invalid colorbrewer ramp") - .into_iter() - .map(|rgb::RGB{r, g, b}| { - [r, g, b].map(|c| F::cast_from(c) / m).into() - }).collect() -} - -/// Perform hue shifting of an RGB value. -/// -// The hue `ω` is in radians. -#[replace_float_literals(F::cast_from(literal))] -fn hueshift<F, U>(ω : F, Rgb([r_in, g_in, b_in]) : Rgb<U>) -> Rgb<U> -where F : Float + CastFrom<U>, - U : Unsigned { - let m = F::cast_from(U::RANGE_MAX); - let r = F::cast_from(r_in) / m; - let g = F::cast_from(g_in) / m; - let b = F::cast_from(b_in) / m; - let u = ω.cos(); - let w = ω.sin(); - - let nr = (0.299 + 0.701*u + 0.168*w) * r - + (0.587 - 0.587*u + 0.330*w) * g - + (0.114 - 0.114*u - 0.497*w) * b; - let ng = (0.299 - 0.299*u - 0.328*w) * r - + (0.587 + 0.413*u + 0.035*w) * g - + (0.114 - 0.114*u + 0.292*w) *b; - let nb = (0.299 - 0.3*u + 1.25*w) * r - + (0.587 - 0.588*u - 1.05*w) * g - + (0.114 + 0.886*u - 0.203*w) * b; - - Rgb([nr, ng, nb].map(scale_uint)) +#[derive(Serialize)] +struct CSVSpike2<F : Float> { + x : F, + y : F, + alpha : F, } @@ -285,55 +133,27 @@ T1 : RealMapping<F, 2>, T2 : RealMapping<F, 2> > ( - _g_explanation : String, - g : &T1, - _ω_explanation : String, + g0 : Option<&T1>, ω0 : Option<&T2>, grid : LinGrid<F, 2>, - _bnd0 : Option<Bounds<F>>, μ : &DiscreteMeasure<Loc<F, 2>, F>, filename : String, ) { - let [w, h] = grid.count; - let (rawdata_g, range_g) = rawdata_and_range(&grid, g); - let (rawdata_ω, range) = match ω0 { - Some(ω) => { - let (rawdata_ω, range_ω) = rawdata_and_range(&grid, ω); - (rawdata_ω, range_g.common(&range_ω)) - }, - None => { - let mut zeros = Vec::new(); - zeros.resize(rawdata_g.len(), 0.0); - (zeros, range_g) - } - }; - let ramp = get_ramp(RAMP); - let base_im_iter = to_ramp::<F, u16, _>(&range_g, &ramp, rawdata_g.iter().cloned()); - let im_iter = izip!(base_im_iter, rawdata_g.iter(), rawdata_ω.iter()) - .map(|(rgb, &v, &w)| { - hueshift(2.0 * F::PI * (v - w).abs() / range.upper(), rgb) - }); - let mut img = ImageBuffer::new(w as u32, h as u32); - img.pixels_mut() - .zip(im_iter) - .for_each(|(p, v)| *p = v); + let data = grid.into_iter().map(|p@Loc([x, y]) : Loc<F, 2>| CSVHelper2_2 { + x, + y, + g : g0.map(|g| g.apply(&p)), + omega : ω0.map(|ω| ω.apply(&p)) + }); + let csv_f = format!("{}_functions.csv", filename); + write_csv(data, csv_f).expect("CSV save error"); - // Add spikes - let m = μ.iter_masses().fold(F::ZERO, |m, α| m.max(α)); - let μ_range = Bounds(F::ZERO, m); - for &DeltaMeasure{ ref x, α } in μ.iter_spikes() { - let [a, b] = map4(x, &grid.start, &grid.end, &grid.count, |&ξ, &a, &b, &n| { - ((ξ-a)/(b-a)*F::cast_from(n)).as_() - }); - if a < w.as_() && b < h.as_() { - let sc : u16 = scale_range_uint(α, &μ_range); - // TODO: use max of points that map to this pixel. - img[(a, b)] = Rgb([u16::MAX, u16::MAX, sc/2]); - } - } - - img.save_with_format(filename + ".png", ImageFormat::Png) - .expect("Image save error"); + let spikes = μ.iter_spikes().map(|δ| { + let Loc([x, y]) = δ.x; + CSVSpike2 { x, y, alpha : δ.α } + }); + let csv_f = format!("{}_spikes.csv", filename); + write_csv(spikes, csv_f).expect("CSV save error"); } fn plot_into_file< @@ -343,22 +163,14 @@ g : &T1, grid : LinGrid<F, 2>, filename : String, - _explanation : String ) { - let [w, h] = grid.count; - let (rawdata, range) = rawdata_and_range(&grid, g); - let ramp = get_ramp(RAMP); - let im_iter = to_ramp::<F, u16, _>(&range, &ramp, rawdata.iter().cloned()); - let mut img = ImageBuffer::new(w as u32, h as u32); - img.pixels_mut() - .zip(im_iter) - .for_each(|(p, v)| *p = v); - img.save_with_format(filename.clone() + ".png", ImageFormat::Png) - .expect("Image save error"); - - let csv_iter = grid.into_iter().zip(rawdata.iter()).map(|(Loc(x), &v)| (x, v)); - let csv_f = filename + ".txt"; - write_csv(csv_iter, csv_f).expect("CSV save error"); + let data = grid.into_iter().map(|p@Loc([x, y]) : Loc<F, 2>| CSVHelper2 { + x, + y, + f : g.apply(&p), + }); + let csv_f = format!("{}.txt", filename); + write_csv(data, csv_f).expect("CSV save error"); } } @@ -386,12 +198,10 @@ /// This calls [`PlotLookup::plot_into_file_spikes`] with a sequentially numbered file name. pub fn plot_spikes<T1, T2>( &mut self, - g_explanation : String, - g : &T1, - ω_explanation : String, + iter : usize, + g : Option<&T1>, ω : Option<&T2>, - tol : Option<Bounds<F>>, - μ : &DiscreteMeasure<Loc<F, N>, F>, + μ : &RNDM<F, N>, ) where T1 : RealMapping<F, N>, T2 : RealMapping<F, N> { @@ -400,12 +210,11 @@ } if self.plot_count < self.max_plots { PlotLookup::plot_into_file_spikes( - g_explanation, g, - ω_explanation, ω, + g, + ω, self.grid, - tol, μ, - format!("{}out{:03}", self.prefix, self.plot_count) + format!("{}out{:03}", self.prefix, iter) ); self.plot_count += 1; }
--- /dev/null Thu Jan 01 00:00:00 1970 +0000 +++ b/src/preadjoint_helper.rs Mon Feb 17 13:54:53 2025 -0500 @@ -0,0 +1,55 @@ +/*! +Preadjoint construction helper +*/ + +use std::marker::PhantomData; +use alg_tools::types::*; +pub use alg_tools::linops::*; +use alg_tools::norms::{Norm, HasDualExponent}; + +/// Helper structure for constructing preadjoints of `S` where `S : Linear<X>`. +/// [`Linear`] needs to be implemented for each instance, but [`Adjointable`] +/// and [`BoundedLinear`] have blanket implementations. +#[derive(Clone,Debug)] +pub struct PreadjointHelper<'a, S : 'a, X> { + pub forward_op : &'a S, + _domain : PhantomData<X> +} + +impl<'a, S : 'a, X> PreadjointHelper<'a, S, X> { + pub fn new(forward_op : &'a S) -> Self { + PreadjointHelper { forward_op, _domain: PhantomData } + } +} + +impl<'a, X, Ypre, S> Adjointable<Ypre, X> +for PreadjointHelper<'a, S, X> +where + X : Space, + Ypre : Space, + Self : Linear<Ypre>, + S : Clone + Linear<X> +{ + type AdjointCodomain = S::Codomain; + type Adjoint<'b> = S where Self : 'b; + + fn adjoint(&self) -> Self::Adjoint<'_> { + self.forward_op.clone() + } +} + +impl<'a, F, X, Ypre, ExpXpre, ExpYpre, S> BoundedLinear<Ypre, ExpYpre, ExpXpre, F> +for PreadjointHelper<'a, S, X> +where + ExpXpre : HasDualExponent, + ExpYpre : HasDualExponent, + F : Float, + X : Space + Norm<F, ExpXpre::DualExp>, + Ypre : Space + Norm<F, ExpYpre>, + Self : Linear<Ypre>, + S : 'a + Clone + BoundedLinear<X, ExpXpre::DualExp, ExpYpre::DualExp, F> +{ + fn opnorm_bound(&self, expy : ExpYpre, expx : ExpXpre) -> F { + self.forward_op.opnorm_bound(expx.dual_exponent(), expy.dual_exponent()) + } +}
--- /dev/null Thu Jan 01 00:00:00 1970 +0000 +++ b/src/prox_penalty.rs Mon Feb 17 13:54:53 2025 -0500 @@ -0,0 +1,192 @@ +/*! +Proximal penalty abstraction +*/ + +use alg_tools::types::*; +use numeric_literals::replace_float_literals; +use serde::{Deserialize, Serialize}; + +use crate::measures::merging::SpikeMergingMethod; +use crate::measures::RNDM; +use crate::regularisation::RegTerm; +use crate::subproblem::InnerSettings; +use crate::tolerance::Tolerance; +use crate::types::{IterInfo, RefinementSettings}; +use alg_tools::iterate::{AlgIterator, AlgIteratorIteration}; +use alg_tools::mapping::RealMapping; +use alg_tools::nalgebra_support::ToNalgebraRealField; + +pub mod radon_squared; +pub mod wave; +pub use radon_squared::RadonSquared; + +/// Settings for the solution of the stepwise optimality condition. +#[derive(Clone, Copy, Eq, PartialEq, Serialize, Deserialize, Debug)] +#[serde(default)] +pub struct FBGenericConfig<F: Float> { + /// Tolerance for point insertion. + pub tolerance: Tolerance<F>, + + /// Stop looking for predual maximum (where to isert a new point) below + /// `tolerance` multiplied by this factor. + /// + /// Not used by [`crate::prox_penalty::radon_squared`]. + pub insertion_cutoff_factor: F, + + /// Settings for branch and bound refinement when looking for predual maxima + pub refinement: RefinementSettings<F>, + + /// Maximum insertions within each outer iteration + /// + /// Not used by [`crate::prox_penalty::radon_squared`]. + pub max_insertions: usize, + + /// Pair `(n, m)` for maximum insertions `m` on first `n` iterations. + /// + /// Not used by [`crate::prox_penalty::radon_squared`]. + pub bootstrap_insertions: Option<(usize, usize)>, + + /// Inner method settings + pub inner: InnerSettings<F>, + + /// Spike merging method + pub merging: SpikeMergingMethod<F>, + + /// Tolerance multiplier for merges + pub merge_tolerance_mult: F, + + /// Merge spikes after last step (even if merging not generally enabled) + pub final_merging: bool, + + /// Use fitness as merging criterion. Implies worse convergence guarantees. + pub fitness_merging: bool, + + /// Iterations between merging heuristic tries + pub merge_every: usize, + // /// Save $μ$ for postprocessing optimisation + // pub postprocessing : bool +} + +#[replace_float_literals(F::cast_from(literal))] +impl<F: Float> Default for FBGenericConfig<F> { + fn default() -> Self { + FBGenericConfig { + tolerance: Default::default(), + insertion_cutoff_factor: 1.0, + refinement: Default::default(), + max_insertions: 100, + //bootstrap_insertions : None, + bootstrap_insertions: Some((10, 1)), + inner: Default::default(), + merging: Default::default(), + final_merging: true, + fitness_merging: false, + merge_every: 10, + merge_tolerance_mult: 2.0, + // postprocessing : false, + } + } +} + +impl<F: Float> FBGenericConfig<F> { + /// Check if merging should be attempted this iteration + pub fn merge_now<I: AlgIterator>(&self, state: &AlgIteratorIteration<I>) -> bool { + self.merging.enabled && state.iteration() % self.merge_every == 0 + } + + /// Returns the final merging method + pub fn final_merging_method(&self) -> SpikeMergingMethod<F> { + SpikeMergingMethod { + enabled: self.final_merging, + ..self.merging + } + } +} + +/// Trait for proximal penalties +pub trait ProxPenalty<F, PreadjointCodomain, Reg, const N: usize> +where + F: Float + ToNalgebraRealField, + Reg: RegTerm<F, N>, +{ + type ReturnMapping: RealMapping<F, N>; + + /// Insert new spikes into `μ` to approximately satisfy optimality conditions + /// with the forward step term fixed to `τv`. + /// + /// May return `τv + w` for `w` a subdifferential of the regularisation term `reg`, + /// as well as an indication of whether the tolerance bounds `ε` are satisfied. + /// + /// `μ_base + ν_delta` is the base point, where `μ` and `μ_base` are assumed to have the same + /// spike locations, while `ν_delta` may have different locations. + /// + /// `τv` is mutable to allow [`alg_tools::bisection_tree::BTFN`] refinement. + /// Actual values of `τv` are not supposed to be mutated. + fn insert_and_reweigh<I>( + &self, + μ: &mut RNDM<F, N>, + τv: &mut PreadjointCodomain, + μ_base: &RNDM<F, N>, + ν_delta: Option<&RNDM<F, N>>, + τ: F, + ε: F, + config: &FBGenericConfig<F>, + reg: &Reg, + state: &AlgIteratorIteration<I>, + stats: &mut IterInfo<F, N>, + ) -> (Option<Self::ReturnMapping>, bool) + where + I: AlgIterator; + + /// Merge spikes, if possible. + /// + /// Either optimality condition merging or objective value (fitness) merging + /// may be used, the latter only if `fitness` is provided and `config.fitness_merging` + /// is set. + fn merge_spikes( + &self, + μ: &mut RNDM<F, N>, + τv: &mut PreadjointCodomain, + μ_base: &RNDM<F, N>, + ν_delta: Option<&RNDM<F, N>>, + τ: F, + ε: F, + config: &FBGenericConfig<F>, + reg: &Reg, + fitness: Option<impl Fn(&RNDM<F, N>) -> F>, + ) -> usize; + + /// Merge spikes, if possible. + /// + /// Unlike [`Self::merge_spikes`], this variant only supports optimality condition based merging + #[inline] + fn merge_spikes_no_fitness( + &self, + μ: &mut RNDM<F, N>, + τv: &mut PreadjointCodomain, + μ_base: &RNDM<F, N>, + ν_delta: Option<&RNDM<F, N>>, + τ: F, + ε: F, + config: &FBGenericConfig<F>, + reg: &Reg, + ) -> usize { + /// This is a hack to create a `None` of same type as a `Some` + // for the `impl Fn` parameter of `merge_spikes`. + #[inline] + fn into_none<T>(_: Option<T>) -> Option<T> { + None + } + self.merge_spikes( + μ, + τv, + μ_base, + ν_delta, + τ, + ε, + config, + reg, + into_none(Some(|_: &RNDM<F, N>| F::ZERO)), + ) + } +}
--- /dev/null Thu Jan 01 00:00:00 1970 +0000 +++ b/src/prox_penalty/radon_squared.rs Mon Feb 17 13:54:53 2025 -0500 @@ -0,0 +1,182 @@ +/*! +Solver for the point source localisation problem using a simplified forward-backward splitting method. + +Instead of the $𝒟$-norm of `fb.rs`, this uses a standard Radon norm for the proximal map. +*/ + +use numeric_literals::replace_float_literals; +use serde::{Serialize, Deserialize}; +use nalgebra::DVector; + +use alg_tools::iterate::{ + AlgIteratorIteration, + AlgIterator +}; +use alg_tools::norms::{L2, Norm}; +use alg_tools::linops::Mapping; +use alg_tools::bisection_tree::{ + BTFN, + Bounds, + BTSearch, + SupportGenerator, + LocalAnalysis, +}; +use alg_tools::mapping::RealMapping; +use alg_tools::nalgebra_support::ToNalgebraRealField; + +use crate::types::*; +use crate::measures::{ + RNDM, + DeltaMeasure, + Radon, +}; +use crate::measures::merging::SpikeMerging; +use crate::regularisation::RegTerm; +use crate::forward_model::{ + ForwardModel, + AdjointProductBoundedBy +}; +use super::{ + FBGenericConfig, + ProxPenalty, +}; + +/// Radon-norm squared proximal penalty + +#[derive(Copy,Clone,Serialize,Deserialize)] +pub struct RadonSquared; + +#[replace_float_literals(F::cast_from(literal))] +impl<F, GA, BTA, S, Reg, const N : usize> +ProxPenalty<F, BTFN<F, GA, BTA, N>, Reg, N> for RadonSquared +where + F : Float + ToNalgebraRealField, + GA : SupportGenerator<F, N, SupportType = S, Id = usize> + Clone, + BTA : BTSearch<F, N, Data=usize, Agg=Bounds<F>>, + S: RealMapping<F, N> + LocalAnalysis<F, Bounds<F>, N>, + Reg : RegTerm<F, N>, + RNDM<F, N> : SpikeMerging<F>, +{ + type ReturnMapping = BTFN<F, GA, BTA, N>; + + fn insert_and_reweigh<I>( + &self, + μ : &mut RNDM<F, N>, + τv : &mut BTFN<F, GA, BTA, N>, + μ_base : &RNDM<F, N>, + ν_delta: Option<&RNDM<F, N>>, + τ : F, + ε : F, + config : &FBGenericConfig<F>, + reg : &Reg, + _state : &AlgIteratorIteration<I>, + stats : &mut IterInfo<F, N>, + ) -> (Option<Self::ReturnMapping>, bool) + where + I : AlgIterator + { + let mut y = μ_base.masses_dvector(); + + assert!(μ_base.len() <= μ.len()); + + 'i_and_w: for i in 0..=1 { + // Optimise weights + if μ.len() > 0 { + // Form finite-dimensional subproblem. The subproblem references to the original μ^k + // from the beginning of the iteration are all contained in the immutable c and g. + // TODO: observe negation of -τv after switch from minus_τv: finite-dimensional + // problems have not yet been updated to sign change. + let g̃ = DVector::from_iterator(μ.len(), + μ.iter_locations() + .map(|ζ| - F::to_nalgebra_mixed(τv.apply(ζ)))); + let mut x = μ.masses_dvector(); + y.extend(std::iter::repeat(0.0.to_nalgebra_mixed()).take(0.max(x.len()-y.len()))); + assert_eq!(y.len(), x.len()); + // Solve finite-dimensional subproblem. + // TODO: This assumes that ν_delta has no common locations with μ-μ_base, to + // ignore it. + stats.inner_iters += reg.solve_findim_l1squared(&y, &g̃, τ, &mut x, ε, config); + + // Update masses of μ based on solution of finite-dimensional subproblem. + μ.set_masses_dvector(&x); + } + + if i>0 { + // Simple debugging test to see if more inserts would be needed. Doesn't seem so. + //let n = μ.dist_matching(μ_base); + //println!("{:?}", reg.find_tolerance_violation_slack(τv, τ, ε, false, config, n)); + break 'i_and_w + } + + // Calculate ‖μ - μ_base‖_ℳ + // TODO: This assumes that ν_delta has no common locations with μ-μ_base. + let n = μ.dist_matching(μ_base) + ν_delta.map_or(0.0, |ν| ν.norm(Radon)); + + // Find a spike to insert, if needed. + // This only check the overall tolerances, not tolerances on support of μ-μ_base or μ, + // which are supposed to have been guaranteed by the finite-dimensional weight optimisation. + match reg.find_tolerance_violation_slack(τv, τ, ε, false, config, n) { + None => { break 'i_and_w }, + Some((ξ, _v_ξ, _in_bounds)) => { + // Weight is found out by running the finite-dimensional optimisation algorithm + // above + *μ += DeltaMeasure { x : ξ, α : 0.0 }; + stats.inserted += 1; + } + }; + } + + (None, true) + } + + fn merge_spikes( + &self, + μ : &mut RNDM<F, N>, + τv : &mut BTFN<F, GA, BTA, N>, + μ_base : &RNDM<F, N>, + ν_delta: Option<&RNDM<F, N>>, + τ : F, + ε : F, + config : &FBGenericConfig<F>, + reg : &Reg, + fitness : Option<impl Fn(&RNDM<F, N>) -> F>, + ) -> usize + { + if config.fitness_merging { + if let Some(f) = fitness { + return μ.merge_spikes_fitness(config.merging, f, |&v| v) + .1 + } + } + μ.merge_spikes(config.merging, |μ_candidate| { + // Important: μ_candidate's new points are afterwards, + // and do not conflict with μ_base. + // TODO: could simplify to requiring μ_base instead of μ_radon. + // but may complicate with sliding base's exgtra points that need to be + // after μ_candidate's extra points. + // TODO: doesn't seem to work, maybe need to merge μ_base as well? + // Although that doesn't seem to make sense. + let μ_radon = match ν_delta { + None => μ_candidate.sub_matching(μ_base), + Some(ν) => μ_candidate.sub_matching(μ_base) - ν, + }; + reg.verify_merge_candidate_radonsq(τv, μ_candidate, τ, ε, &config, &μ_radon) + //let n = μ_candidate.dist_matching(μ_base); + //reg.find_tolerance_violation_slack(τv, τ, ε, false, config, n).is_none() + }) + } +} + + +impl<F, A, const N : usize> AdjointProductBoundedBy<RNDM<F, N>, RadonSquared> +for A +where + F : Float, + A : ForwardModel<RNDM<F, N>, F> +{ + type FloatType = F; + + fn adjoint_product_bound(&self, _ : &RadonSquared) -> Option<Self::FloatType> { + self.opnorm_bound(Radon, L2).powi(2).into() + } +}
--- /dev/null Thu Jan 01 00:00:00 1970 +0000 +++ b/src/prox_penalty/wave.rs Mon Feb 17 13:54:53 2025 -0500 @@ -0,0 +1,191 @@ +/*! +Basic proximal penalty based on convolution operators $𝒟$. + */ + +use numeric_literals::replace_float_literals; +use nalgebra::DVector; +use colored::Colorize; + +use alg_tools::types::*; +use alg_tools::loc::Loc; +use alg_tools::mapping::{Mapping, RealMapping}; +use alg_tools::nalgebra_support::ToNalgebraRealField; +use alg_tools::norms::Linfinity; +use alg_tools::iterate::{ + AlgIteratorIteration, + AlgIterator, +}; +use alg_tools::bisection_tree::{ + BTFN, + PreBTFN, + Bounds, + BTSearch, + SupportGenerator, + LocalAnalysis, + BothGenerators, +}; +use crate::measures::{ + RNDM, + DeltaMeasure, + Radon, +}; +use crate::measures::merging::{ + SpikeMerging, +}; +use crate::seminorms::DiscreteMeasureOp; +use crate::types::{ + IterInfo, +}; +use crate::regularisation::RegTerm; +use super::{ProxPenalty, FBGenericConfig}; + +#[replace_float_literals(F::cast_from(literal))] +impl<F, GA, BTA, S, Reg, 𝒟, G𝒟, K, const N : usize> +ProxPenalty<F, BTFN<F, GA, BTA, N>, Reg, N> for 𝒟 +where + F : Float + ToNalgebraRealField, + GA : SupportGenerator<F, N, SupportType = S, Id = usize> + Clone, + BTA : BTSearch<F, N, Data=usize, Agg=Bounds<F>>, + S: RealMapping<F, N> + LocalAnalysis<F, Bounds<F>, N>, + G𝒟 : SupportGenerator<F, N, SupportType = K, Id = usize> + Clone, + 𝒟 : DiscreteMeasureOp<Loc<F, N>, F, PreCodomain = PreBTFN<F, G𝒟, N>>, + 𝒟::Codomain : RealMapping<F, N>, + K : RealMapping<F, N> + LocalAnalysis<F, Bounds<F>, N>, + Reg : RegTerm<F, N>, + RNDM<F, N> : SpikeMerging<F>, +{ + type ReturnMapping = BTFN<F, BothGenerators<GA, G𝒟>, BTA, N>; + + fn insert_and_reweigh<I>( + &self, + μ : &mut RNDM<F, N>, + τv : &mut BTFN<F, GA, BTA, N>, + μ_base : &RNDM<F, N>, + ν_delta: Option<&RNDM<F, N>>, + τ : F, + ε : F, + config : &FBGenericConfig<F>, + reg : &Reg, + state : &AlgIteratorIteration<I>, + stats : &mut IterInfo<F, N>, + ) -> (Option<BTFN<F, BothGenerators<GA, G𝒟>, BTA, N>>, bool) + where + I : AlgIterator + { + + let op𝒟norm = self.opnorm_bound(Radon, Linfinity); + + // Maximum insertion count and measure difference calculation depend on insertion style. + let (max_insertions, warn_insertions) = match (state.iteration(), config.bootstrap_insertions) { + (i, Some((l, k))) if i <= l => (k, false), + _ => (config.max_insertions, !state.is_quiet()), + }; + + let ω0 = match ν_delta { + None => self.apply(μ_base), + Some(ν) => self.apply(μ_base + ν), + }; + + // Add points to support until within error tolerance or maximum insertion count reached. + let mut count = 0; + let (within_tolerances, d) = 'insertion: loop { + if μ.len() > 0 { + // Form finite-dimensional subproblem. The subproblem references to the original μ^k + // from the beginning of the iteration are all contained in the immutable c and g. + // TODO: observe negation of -τv after switch from minus_τv: finite-dimensional + // problems have not yet been updated to sign change. + let à = self.findim_matrix(μ.iter_locations()); + let g̃ = DVector::from_iterator(μ.len(), + μ.iter_locations() + .map(|ζ| ω0.apply(ζ) - τv.apply(ζ)) + .map(F::to_nalgebra_mixed)); + let mut x = μ.masses_dvector(); + + // The gradient of the forward component of the inner objective is C^*𝒟Cx - g̃. + // We have |C^*𝒟Cx|_2 = sup_{|z|_2 ≤ 1} ⟨z, C^*𝒟Cx⟩ = sup_{|z|_2 ≤ 1} ⟨Cz|𝒟Cx⟩ + // ≤ sup_{|z|_2 ≤ 1} |Cz|_ℳ |𝒟Cx|_∞ ≤ sup_{|z|_2 ≤ 1} |Cz|_ℳ |𝒟| |Cx|_ℳ + // ≤ sup_{|z|_2 ≤ 1} |z|_1 |𝒟| |x|_1 ≤ sup_{|z|_2 ≤ 1} n |z|_2 |𝒟| |x|_2 + // = n |𝒟| |x|_2, where n is the number of points. Therefore + let Ã_normest = op𝒟norm * F::cast_from(μ.len()); + + // Solve finite-dimensional subproblem. + stats.inner_iters += reg.solve_findim(&Ã, &g̃, τ, &mut x, Ã_normest, ε, config); + + // Update masses of μ based on solution of finite-dimensional subproblem. + μ.set_masses_dvector(&x); + } + + // Form d = τv + 𝒟μ - ω0 = τv + 𝒟(μ - μ^k) for checking the proximate optimality + // conditions in the predual space, and finding new points for insertion, if necessary. + let mut d = &*τv + match ν_delta { + None => self.preapply(μ.sub_matching(μ_base)), + Some(ν) => self.preapply(μ.sub_matching(μ_base) - ν) + }; + + // If no merging heuristic is used, let's be more conservative about spike insertion, + // and skip it after first round. If merging is done, being more greedy about spike + // insertion also seems to improve performance. + let skip_by_rough_check = if config.merging.enabled { + false + } else { + count > 0 + }; + + // Find a spike to insert, if needed + let (ξ, _v_ξ, in_bounds) = match reg.find_tolerance_violation( + &mut d, τ, ε, skip_by_rough_check, config + ) { + None => break 'insertion (true, d), + Some(res) => res, + }; + + // Break if maximum insertion count reached + if count >= max_insertions { + break 'insertion (in_bounds, d) + } + + // No point in optimising the weight here; the finite-dimensional algorithm is fast. + *μ += DeltaMeasure { x : ξ, α : 0.0 }; + count += 1; + stats.inserted += 1; + }; + + if !within_tolerances && warn_insertions { + // Complain (but continue) if we failed to get within tolerances + // by inserting more points. + let err = format!("Maximum insertions reached without achieving \ + subproblem solution tolerance"); + println!("{}", err.red()); + } + + (Some(d), within_tolerances) + } + + fn merge_spikes( + &self, + μ : &mut RNDM<F, N>, + τv : &mut BTFN<F, GA, BTA, N>, + μ_base : &RNDM<F, N>, + ν_delta: Option<&RNDM<F, N>>, + τ : F, + ε : F, + config : &FBGenericConfig<F>, + reg : &Reg, + fitness : Option<impl Fn(&RNDM<F, N>) -> F>, + ) -> usize + { + if config.fitness_merging { + if let Some(f) = fitness { + return μ.merge_spikes_fitness(config.merging, f, |&v| v) + .1 + } + } + μ.merge_spikes(config.merging, |μ_candidate| { + let mut d = &*τv + self.preapply(match ν_delta { + None => μ_candidate.sub_matching(μ_base), + Some(ν) => μ_candidate.sub_matching(μ_base) - ν, + }); + reg.verify_merge_candidate(&mut d, μ_candidate, τ, ε, config) + }) + } +}
--- a/src/regularisation.rs Tue Aug 01 10:25:09 2023 +0300 +++ b/src/regularisation.rs Mon Feb 17 13:54:53 2025 -0500 @@ -2,25 +2,41 @@ Regularisation terms */ -use serde::{Serialize, Deserialize}; -use alg_tools::norms::Norm; -use alg_tools::linops::Apply; -use alg_tools::loc::Loc; +#[allow(unused_imports)] // Used by documentation. +use crate::fb::pointsource_fb_reg; +use crate::fb::FBGenericConfig; +use crate::measures::{DeltaMeasure, Radon, RNDM}; +#[allow(unused_imports)] // Used by documentation. +use crate::sliding_fb::pointsource_sliding_fb_reg; use crate::types::*; -use crate::measures::{ - DiscreteMeasure, - Radon +use alg_tools::instance::Instance; +use alg_tools::linops::Mapping; +use alg_tools::loc::Loc; +use alg_tools::norms::Norm; +use numeric_literals::replace_float_literals; +use serde::{Deserialize, Serialize}; + +use crate::subproblem::{ + l1squared_nonneg::l1squared_nonneg, l1squared_unconstrained::l1squared_unconstrained, + nonneg::quadratic_nonneg, unconstrained::quadratic_unconstrained, }; -#[allow(unused_imports)] // Used by documentation. -use crate::fb::generic_pointsource_fb_reg; +use alg_tools::bisection_tree::{ + BTSearch, Bounded, Bounds, LocalAnalysis, P2Minimise, SupportGenerator, BTFN, +}; +use alg_tools::iterate::AlgIteratorFactory; +use alg_tools::nalgebra_support::ToNalgebraRealField; +use nalgebra::{DMatrix, DVector}; -/// The regularisation term $α\\|μ\\|\_{ℳ(Ω)} + δ_{≥ 0}(μ)$ for [`generic_pointsource_fb_reg`]. +use std::cmp::Ordering::{Equal, Greater, Less}; + +/// The regularisation term $α\\|μ\\|\_{ℳ(Ω)} + δ_{≥ 0}(μ)$ for [`pointsource_fb_reg`] and other +/// algorithms. /// /// The only member of the struct is the regularisation parameter α. #[derive(Copy, Clone, Debug, Serialize, Deserialize)] -pub struct NonnegRadonRegTerm<F : Float>(pub F /* α */); +pub struct NonnegRadonRegTerm<F: Float>(pub F /* α */); -impl<'a, F : Float> NonnegRadonRegTerm<F> { +impl<'a, F: Float> NonnegRadonRegTerm<F> { /// Returns the regularisation parameter pub fn α(&self) -> F { let &NonnegRadonRegTerm(α) = self; @@ -28,24 +44,24 @@ } } -impl<'a, F : Float, const N : usize> Apply<&'a DiscreteMeasure<Loc<F, N>, F>> -for NonnegRadonRegTerm<F> { - type Output = F; - - fn apply(&self, μ : &'a DiscreteMeasure<Loc<F, N>, F>) -> F { - self.α() * μ.norm(Radon) +impl<'a, F: Float, const N: usize> Mapping<RNDM<F, N>> for NonnegRadonRegTerm<F> { + type Codomain = F; + + fn apply<I>(&self, μ: I) -> F + where + I: Instance<RNDM<F, N>>, + { + self.α() * μ.eval(|x| x.norm(Radon)) } } - -/// The regularisation term $α\|μ\|_{ℳ(Ω)}$ for [`generic_pointsource_fb_reg`]. +/// The regularisation term $α\|μ\|_{ℳ(Ω)}$ for [`pointsource_fb_reg`]. /// /// The only member of the struct is the regularisation parameter α. #[derive(Copy, Clone, Debug, Serialize, Deserialize)] -pub struct RadonRegTerm<F : Float>(pub F /* α */); +pub struct RadonRegTerm<F: Float>(pub F /* α */); - -impl<'a, F : Float> RadonRegTerm<F> { +impl<'a, F: Float> RadonRegTerm<F> { /// Returns the regularisation parameter pub fn α(&self) -> F { let &RadonRegTerm(α) = self; @@ -53,32 +69,615 @@ } } -impl<'a, F : Float, const N : usize> Apply<&'a DiscreteMeasure<Loc<F, N>, F>> -for RadonRegTerm<F> { - type Output = F; - - fn apply(&self, μ : &'a DiscreteMeasure<Loc<F, N>, F>) -> F { - self.α() * μ.norm(Radon) +impl<'a, F: Float, const N: usize> Mapping<RNDM<F, N>> for RadonRegTerm<F> { + type Codomain = F; + + fn apply<I>(&self, μ: I) -> F + where + I: Instance<RNDM<F, N>>, + { + self.α() * μ.eval(|x| x.norm(Radon)) } } /// Regularisation term configuration #[derive(Clone, Copy, Eq, PartialEq, Serialize, Deserialize, Debug)] -pub enum Regularisation<F : Float> { +pub enum Regularisation<F: Float> { /// $α \\|μ\\|\_{ℳ(Ω)}$ Radon(F), /// $α\\|μ\\|\_{ℳ(Ω)} + δ_{≥ 0}(μ)$ NonnegRadon(F), } -impl<'a, F : Float, const N : usize> Apply<&'a DiscreteMeasure<Loc<F, N>, F>> -for Regularisation<F> { - type Output = F; - - fn apply(&self, μ : &'a DiscreteMeasure<Loc<F, N>, F>) -> F { +impl<'a, F: Float, const N: usize> Mapping<RNDM<F, N>> for Regularisation<F> { + type Codomain = F; + + fn apply<I>(&self, μ: I) -> F + where + I: Instance<RNDM<F, N>>, + { match *self { Self::Radon(α) => RadonRegTerm(α).apply(μ), Self::NonnegRadon(α) => NonnegRadonRegTerm(α).apply(μ), } } } + +/// Abstraction of regularisation terms. +pub trait RegTerm<F: Float + ToNalgebraRealField, const N: usize>: + Mapping<RNDM<F, N>, Codomain = F> +{ + /// Approximately solve the problem + /// <div>$$ + /// \min_{x ∈ ℝ^n} \frac{1}{2} x^⊤Ax - g^⊤ x + τ G(x) + /// $$</div> + /// for $G$ depending on the trait implementation. + /// + /// The parameter `mA` is $A$. An estimate for its opeator norm should be provided in + /// `mA_normest`. The initial iterate and output is `x`. The current main tolerance is `ε`. + /// + /// Returns the number of iterations taken. + fn solve_findim( + &self, + mA: &DMatrix<F::MixedType>, + g: &DVector<F::MixedType>, + τ: F, + x: &mut DVector<F::MixedType>, + mA_normest: F, + ε: F, + config: &FBGenericConfig<F>, + ) -> usize; + + /// Approximately solve the problem + /// <div>$$ + /// \min_{x ∈ ℝ^n} \frac{1}{2} |x-y|_1^2 - g^⊤ x + τ G(x) + /// $$</div> + /// for $G$ depending on the trait implementation. + /// + /// Returns the number of iterations taken. + fn solve_findim_l1squared( + &self, + y: &DVector<F::MixedType>, + g: &DVector<F::MixedType>, + τ: F, + x: &mut DVector<F::MixedType>, + ε: F, + config: &FBGenericConfig<F>, + ) -> usize; + + /// Find a point where `d` may violate the tolerance `ε`. + /// + /// If `skip_by_rough_check` is set, do not find the point if a rough check indicates that we + /// are in bounds. `ε` is the current main tolerance and `τ` a scaling factor for the + /// regulariser. + /// + /// Returns `None` if `d` is in bounds either based on the rough check, or a more precise check + /// terminating early. Otherwise returns a possibly violating point, the value of `d` there, + /// and a boolean indicating whether the found point is in bounds. + fn find_tolerance_violation<G, BT>( + &self, + d: &mut BTFN<F, G, BT, N>, + τ: F, + ε: F, + skip_by_rough_check: bool, + config: &FBGenericConfig<F>, + ) -> Option<(Loc<F, N>, F, bool)> + where + BT: BTSearch<F, N, Agg = Bounds<F>>, + G: SupportGenerator<F, N, Id = BT::Data>, + G::SupportType: Mapping<Loc<F, N>, Codomain = F> + LocalAnalysis<F, Bounds<F>, N>, + { + self.find_tolerance_violation_slack(d, τ, ε, skip_by_rough_check, config, F::ZERO) + } + + /// Find a point where `d` may violate the tolerance `ε`. + /// + /// This version includes a `slack` parameter to expand the tolerances. + /// It is used for Radon-norm squared proximal term in [`crate::prox_penalty::radon_squared`]. + /// + /// If `skip_by_rough_check` is set, do not find the point if a rough check indicates that we + /// are in bounds. `ε` is the current main tolerance and `τ` a scaling factor for the + /// regulariser. + /// + /// Returns `None` if `d` is in bounds either based on the rough check, or a more precise check + /// terminating early. Otherwise returns a possibly violating point, the value of `d` there, + /// and a boolean indicating whether the found point is in bounds. + fn find_tolerance_violation_slack<G, BT>( + &self, + d: &mut BTFN<F, G, BT, N>, + τ: F, + ε: F, + skip_by_rough_check: bool, + config: &FBGenericConfig<F>, + slack: F, + ) -> Option<(Loc<F, N>, F, bool)> + where + BT: BTSearch<F, N, Agg = Bounds<F>>, + G: SupportGenerator<F, N, Id = BT::Data>, + G::SupportType: Mapping<Loc<F, N>, Codomain = F> + LocalAnalysis<F, Bounds<F>, N>; + + /// Verify that `d` is in bounds `ε` for a merge candidate `μ` + /// + /// `ε` is the current main tolerance and `τ` a scaling factor for the regulariser. + fn verify_merge_candidate<G, BT>( + &self, + d: &mut BTFN<F, G, BT, N>, + μ: &RNDM<F, N>, + τ: F, + ε: F, + config: &FBGenericConfig<F>, + ) -> bool + where + BT: BTSearch<F, N, Agg = Bounds<F>>, + G: SupportGenerator<F, N, Id = BT::Data>, + G::SupportType: Mapping<Loc<F, N>, Codomain = F> + LocalAnalysis<F, Bounds<F>, N>; + + /// Verify that `d` is in bounds `ε` for a merge candidate `μ` + /// + /// This version is s used for Radon-norm squared proximal term in + /// [`crate::prox_penalty::radon_squared`]. + /// The [measures][crate::measures::DiscreteMeasure] `μ` and `radon_μ` are supposed to have + /// same coordinates at same agreeing indices. + /// + /// `ε` is the current main tolerance and `τ` a scaling factor for the regulariser. + fn verify_merge_candidate_radonsq<G, BT>( + &self, + d: &mut BTFN<F, G, BT, N>, + μ: &RNDM<F, N>, + τ: F, + ε: F, + config: &FBGenericConfig<F>, + radon_μ: &RNDM<F, N>, + ) -> bool + where + BT: BTSearch<F, N, Agg = Bounds<F>>, + G: SupportGenerator<F, N, Id = BT::Data>, + G::SupportType: Mapping<Loc<F, N>, Codomain = F> + LocalAnalysis<F, Bounds<F>, N>; + + /// TODO: document this + fn target_bounds(&self, τ: F, ε: F) -> Option<Bounds<F>>; + + /// Returns a scaling factor for the tolerance sequence. + /// + /// Typically this is the regularisation parameter. + fn tolerance_scaling(&self) -> F; +} + +/// Abstraction of regularisation terms for [`pointsource_sliding_fb_reg`]. +pub trait SlidingRegTerm<F: Float + ToNalgebraRealField, const N: usize>: RegTerm<F, N> { + /// Calculate $τ[w(z) - w(y)]$ for some w in the subdifferential of the regularisation + /// term, such that $-ε ≤ τw - d ≤ ε$. + fn goodness<G, BT>( + &self, + d: &mut BTFN<F, G, BT, N>, + μ: &RNDM<F, N>, + y: &Loc<F, N>, + z: &Loc<F, N>, + τ: F, + ε: F, + config: &FBGenericConfig<F>, + ) -> F + where + BT: BTSearch<F, N, Agg = Bounds<F>>, + G: SupportGenerator<F, N, Id = BT::Data>, + G::SupportType: Mapping<Loc<F, N>, Codomain = F> + LocalAnalysis<F, Bounds<F>, N>; + + /// Convert bound on the regulariser to a bond on the Radon norm + fn radon_norm_bound(&self, b: F) -> F; +} + +#[replace_float_literals(F::cast_from(literal))] +impl<F: Float + ToNalgebraRealField, const N: usize> RegTerm<F, N> for NonnegRadonRegTerm<F> +where + Cube<F, N>: P2Minimise<Loc<F, N>, F>, +{ + fn solve_findim( + &self, + mA: &DMatrix<F::MixedType>, + g: &DVector<F::MixedType>, + τ: F, + x: &mut DVector<F::MixedType>, + mA_normest: F, + ε: F, + config: &FBGenericConfig<F>, + ) -> usize { + let inner_tolerance = ε * config.inner.tolerance_mult; + let inner_it = config.inner.iterator_options.stop_target(inner_tolerance); + quadratic_nonneg(mA, g, τ * self.α(), x, mA_normest, &config.inner, inner_it) + } + + fn solve_findim_l1squared( + &self, + y: &DVector<F::MixedType>, + g: &DVector<F::MixedType>, + τ: F, + x: &mut DVector<F::MixedType>, + ε: F, + config: &FBGenericConfig<F>, + ) -> usize { + let inner_tolerance = ε * config.inner.tolerance_mult; + let inner_it = config.inner.iterator_options.stop_target(inner_tolerance); + l1squared_nonneg(y, g, τ * self.α(), 1.0, x, &config.inner, inner_it) + } + + #[inline] + fn find_tolerance_violation_slack<G, BT>( + &self, + d: &mut BTFN<F, G, BT, N>, + τ: F, + ε: F, + skip_by_rough_check: bool, + config: &FBGenericConfig<F>, + slack: F, + ) -> Option<(Loc<F, N>, F, bool)> + where + BT: BTSearch<F, N, Agg = Bounds<F>>, + G: SupportGenerator<F, N, Id = BT::Data>, + G::SupportType: Mapping<Loc<F, N>, Codomain = F> + LocalAnalysis<F, Bounds<F>, N>, + { + let τα = τ * self.α(); + let keep_above = -τα - slack - ε; + let minimise_below = -τα - slack - ε * config.insertion_cutoff_factor; + let refinement_tolerance = ε * config.refinement.tolerance_mult; + + // If preliminary check indicates that we are in bounds, and if it otherwise matches + // the insertion strategy, skip insertion. + if skip_by_rough_check && d.bounds().lower() >= keep_above { + None + } else { + // If the rough check didn't indicate no insertion needed, find minimising point. + d.minimise_below( + minimise_below, + refinement_tolerance, + config.refinement.max_steps, + ) + .map(|(ξ, v_ξ)| (ξ, v_ξ, v_ξ >= keep_above)) + } + } + + fn verify_merge_candidate<G, BT>( + &self, + d: &mut BTFN<F, G, BT, N>, + μ: &RNDM<F, N>, + τ: F, + ε: F, + config: &FBGenericConfig<F>, + ) -> bool + where + BT: BTSearch<F, N, Agg = Bounds<F>>, + G: SupportGenerator<F, N, Id = BT::Data>, + G::SupportType: Mapping<Loc<F, N>, Codomain = F> + LocalAnalysis<F, Bounds<F>, N>, + { + let τα = τ * self.α(); + let refinement_tolerance = ε * config.refinement.tolerance_mult; + let merge_tolerance = config.merge_tolerance_mult * ε; + let keep_above = -τα - merge_tolerance; + let keep_supp_below = -τα + merge_tolerance; + let bnd = d.bounds(); + + return (bnd.upper() <= keep_supp_below + || μ + .iter_spikes() + .all(|&DeltaMeasure { α, ref x }| (α == 0.0) || d.apply(x) <= keep_supp_below)) + && (bnd.lower() >= keep_above + || d.has_lower_bound( + keep_above, + refinement_tolerance, + config.refinement.max_steps, + )); + } + + fn verify_merge_candidate_radonsq<G, BT>( + &self, + d: &mut BTFN<F, G, BT, N>, + μ: &RNDM<F, N>, + τ: F, + ε: F, + config: &FBGenericConfig<F>, + radon_μ: &RNDM<F, N>, + ) -> bool + where + BT: BTSearch<F, N, Agg = Bounds<F>>, + G: SupportGenerator<F, N, Id = BT::Data>, + G::SupportType: Mapping<Loc<F, N>, Codomain = F> + LocalAnalysis<F, Bounds<F>, N>, + { + let τα = τ * self.α(); + let refinement_tolerance = ε * config.refinement.tolerance_mult; + let merge_tolerance = config.merge_tolerance_mult * ε; + let slack = radon_μ.norm(Radon); + let bnd = d.bounds(); + + return { + μ.both_matching(radon_μ).all(|(α, rα, x)| { + let v = -d.apply(x); // TODO: observe ad hoc negation here, after minus_τv + // switch to τv. + let (l1, u1) = match α.partial_cmp(&0.0).unwrap_or(Equal) { + Greater => (τα, τα), + _ => (F::NEG_INFINITY, τα), + // Less should not happen; treated as Equal + }; + let (l2, u2) = match rα.partial_cmp(&0.0).unwrap_or(Equal) { + Greater => (slack, slack), + Equal => (-slack, slack), + Less => (-slack, -slack), + }; + // TODO: both fail. + (l1 + l2 - merge_tolerance <= v) && (v <= u1 + u2 + merge_tolerance) + }) + } && { + let keep_above = -τα - slack - merge_tolerance; + bnd.lower() <= keep_above + || d.has_lower_bound( + keep_above, + refinement_tolerance, + config.refinement.max_steps, + ) + }; + } + + fn target_bounds(&self, τ: F, ε: F) -> Option<Bounds<F>> { + let τα = τ * self.α(); + Some(Bounds(τα - ε, τα + ε)) + } + + fn tolerance_scaling(&self) -> F { + self.α() + } +} + +#[replace_float_literals(F::cast_from(literal))] +impl<F: Float + ToNalgebraRealField, const N: usize> SlidingRegTerm<F, N> for NonnegRadonRegTerm<F> +where + Cube<F, N>: P2Minimise<Loc<F, N>, F>, +{ + fn goodness<G, BT>( + &self, + d: &mut BTFN<F, G, BT, N>, + _μ: &RNDM<F, N>, + y: &Loc<F, N>, + z: &Loc<F, N>, + τ: F, + ε: F, + _config: &FBGenericConfig<F>, + ) -> F + where + BT: BTSearch<F, N, Agg = Bounds<F>>, + G: SupportGenerator<F, N, Id = BT::Data>, + G::SupportType: Mapping<Loc<F, N>, Codomain = F> + LocalAnalysis<F, Bounds<F>, N>, + { + let w = |x| 1.0.min((ε + d.apply(x)) / (τ * self.α())); + w(z) - w(y) + } + + fn radon_norm_bound(&self, b: F) -> F { + b / self.α() + } +} + +#[replace_float_literals(F::cast_from(literal))] +impl<F: Float + ToNalgebraRealField, const N: usize> RegTerm<F, N> for RadonRegTerm<F> +where + Cube<F, N>: P2Minimise<Loc<F, N>, F>, +{ + fn solve_findim( + &self, + mA: &DMatrix<F::MixedType>, + g: &DVector<F::MixedType>, + τ: F, + x: &mut DVector<F::MixedType>, + mA_normest: F, + ε: F, + config: &FBGenericConfig<F>, + ) -> usize { + let inner_tolerance = ε * config.inner.tolerance_mult; + let inner_it = config.inner.iterator_options.stop_target(inner_tolerance); + quadratic_unconstrained(mA, g, τ * self.α(), x, mA_normest, &config.inner, inner_it) + } + + fn solve_findim_l1squared( + &self, + y: &DVector<F::MixedType>, + g: &DVector<F::MixedType>, + τ: F, + x: &mut DVector<F::MixedType>, + ε: F, + config: &FBGenericConfig<F>, + ) -> usize { + let inner_tolerance = ε * config.inner.tolerance_mult; + let inner_it = config.inner.iterator_options.stop_target(inner_tolerance); + l1squared_unconstrained(y, g, τ * self.α(), 1.0, x, &config.inner, inner_it) + } + + fn find_tolerance_violation_slack<G, BT>( + &self, + d: &mut BTFN<F, G, BT, N>, + τ: F, + ε: F, + skip_by_rough_check: bool, + config: &FBGenericConfig<F>, + slack: F, + ) -> Option<(Loc<F, N>, F, bool)> + where + BT: BTSearch<F, N, Agg = Bounds<F>>, + G: SupportGenerator<F, N, Id = BT::Data>, + G::SupportType: Mapping<Loc<F, N>, Codomain = F> + LocalAnalysis<F, Bounds<F>, N>, + { + let τα = τ * self.α(); + let keep_below = τα + slack + ε; + let keep_above = -(τα + slack) - ε; + let maximise_above = τα + slack + ε * config.insertion_cutoff_factor; + let minimise_below = -(τα + slack) - ε * config.insertion_cutoff_factor; + let refinement_tolerance = ε * config.refinement.tolerance_mult; + + // If preliminary check indicates that we are in bonds, and if it otherwise matches + // the insertion strategy, skip insertion. + if skip_by_rough_check && Bounds(keep_above, keep_below).superset(&d.bounds()) { + None + } else { + // If the rough check didn't indicate no insertion needed, find maximising point. + let mx = d.maximise_above( + maximise_above, + refinement_tolerance, + config.refinement.max_steps, + ); + let mi = d.minimise_below( + minimise_below, + refinement_tolerance, + config.refinement.max_steps, + ); + + match (mx, mi) { + (None, None) => None, + (Some((ξ, v_ξ)), None) => Some((ξ, v_ξ, keep_below >= v_ξ)), + (None, Some((ζ, v_ζ))) => Some((ζ, v_ζ, keep_above <= v_ζ)), + (Some((ξ, v_ξ)), Some((ζ, v_ζ))) => { + if v_ξ - τα > τα - v_ζ { + Some((ξ, v_ξ, keep_below >= v_ξ)) + } else { + Some((ζ, v_ζ, keep_above <= v_ζ)) + } + } + } + } + } + + fn verify_merge_candidate<G, BT>( + &self, + d: &mut BTFN<F, G, BT, N>, + μ: &RNDM<F, N>, + τ: F, + ε: F, + config: &FBGenericConfig<F>, + ) -> bool + where + BT: BTSearch<F, N, Agg = Bounds<F>>, + G: SupportGenerator<F, N, Id = BT::Data>, + G::SupportType: Mapping<Loc<F, N>, Codomain = F> + LocalAnalysis<F, Bounds<F>, N>, + { + let τα = τ * self.α(); + let refinement_tolerance = ε * config.refinement.tolerance_mult; + let merge_tolerance = config.merge_tolerance_mult * ε; + let keep_below = τα + merge_tolerance; + let keep_above = -τα - merge_tolerance; + let keep_supp_pos_above = τα - merge_tolerance; + let keep_supp_neg_below = -τα + merge_tolerance; + let bnd = d.bounds(); + + return ((bnd.lower() >= keep_supp_pos_above && bnd.upper() <= keep_supp_neg_below) + || μ + .iter_spikes() + .all(|&DeltaMeasure { α: β, ref x }| match β.partial_cmp(&0.0) { + Some(Greater) => d.apply(x) >= keep_supp_pos_above, + Some(Less) => d.apply(x) <= keep_supp_neg_below, + _ => true, + })) + && (bnd.upper() <= keep_below + || d.has_upper_bound( + keep_below, + refinement_tolerance, + config.refinement.max_steps, + )) + && (bnd.lower() >= keep_above + || d.has_lower_bound( + keep_above, + refinement_tolerance, + config.refinement.max_steps, + )); + } + + fn verify_merge_candidate_radonsq<G, BT>( + &self, + d: &mut BTFN<F, G, BT, N>, + μ: &RNDM<F, N>, + τ: F, + ε: F, + config: &FBGenericConfig<F>, + radon_μ: &RNDM<F, N>, + ) -> bool + where + BT: BTSearch<F, N, Agg = Bounds<F>>, + G: SupportGenerator<F, N, Id = BT::Data>, + G::SupportType: Mapping<Loc<F, N>, Codomain = F> + LocalAnalysis<F, Bounds<F>, N>, + { + let τα = τ * self.α(); + let refinement_tolerance = ε * config.refinement.tolerance_mult; + let merge_tolerance = config.merge_tolerance_mult * ε; + let slack = radon_μ.norm(Radon); + let bnd = d.bounds(); + + return { + μ.both_matching(radon_μ).all(|(α, rα, x)| { + let v = d.apply(x); + let (l1, u1) = match α.partial_cmp(&0.0).unwrap_or(Equal) { + Greater => (τα, τα), + Equal => (-τα, τα), + Less => (-τα, -τα), + }; + let (l2, u2) = match rα.partial_cmp(&0.0).unwrap_or(Equal) { + Greater => (slack, slack), + Equal => (-slack, slack), + Less => (-slack, -slack), + }; + (l1 + l2 - merge_tolerance <= v) && (v <= u1 + u2 + merge_tolerance) + }) + } && { + let keep_below = τα + slack + merge_tolerance; + bnd.upper() <= keep_below + || d.has_upper_bound( + keep_below, + refinement_tolerance, + config.refinement.max_steps, + ) + } && { + let keep_above = -τα - slack - merge_tolerance; + bnd.lower() >= keep_above + || d.has_lower_bound( + keep_above, + refinement_tolerance, + config.refinement.max_steps, + ) + }; + } + + fn target_bounds(&self, τ: F, ε: F) -> Option<Bounds<F>> { + let τα = τ * self.α(); + Some(Bounds(-τα - ε, τα + ε)) + } + + fn tolerance_scaling(&self) -> F { + self.α() + } +} + +#[replace_float_literals(F::cast_from(literal))] +impl<F: Float + ToNalgebraRealField, const N: usize> SlidingRegTerm<F, N> for RadonRegTerm<F> +where + Cube<F, N>: P2Minimise<Loc<F, N>, F>, +{ + fn goodness<G, BT>( + &self, + d: &mut BTFN<F, G, BT, N>, + _μ: &RNDM<F, N>, + y: &Loc<F, N>, + z: &Loc<F, N>, + τ: F, + ε: F, + _config: &FBGenericConfig<F>, + ) -> F + where + BT: BTSearch<F, N, Agg = Bounds<F>>, + G: SupportGenerator<F, N, Id = BT::Data>, + G::SupportType: Mapping<Loc<F, N>, Codomain = F> + LocalAnalysis<F, Bounds<F>, N>, + { + let α = self.α(); + let w = |x| { + let dx = d.apply(x); + ((-ε + dx) / (τ * α)).max(1.0.min(ε + dx) / (τ * α)) + }; + w(z) - w(y) + } + + fn radon_norm_bound(&self, b: F) -> F { + b / self.α() + } +}
--- a/src/run.rs Tue Aug 01 10:25:09 2023 +0300 +++ b/src/run.rs Mon Feb 17 13:54:53 2025 -0500 @@ -26,53 +26,116 @@ AlgIteratorOptions, Verbose, AlgIteratorFactory, + LoggingIteratorFactory, + TimingIteratorFactory, + BasicAlgIteratorFactory, }; use alg_tools::logger::Logger; -use alg_tools::error::DynError; +use alg_tools::error::{ + DynError, + DynResult, +}; use alg_tools::tabledump::TableDump; use alg_tools::sets::Cube; -use alg_tools::mapping::RealMapping; +use alg_tools::mapping::{ + RealMapping, + DifferentiableMapping, + DifferentiableRealMapping, + Instance +}; use alg_tools::nalgebra_support::ToNalgebraRealField; use alg_tools::euclidean::Euclidean; -use alg_tools::norms::L1; -use alg_tools::lingrid::lingrid; +use alg_tools::lingrid::{lingrid, LinSpace}; use alg_tools::sets::SetOrd; +use alg_tools::linops::{RowOp, IdOp /*, ZeroOp*/}; +use alg_tools::discrete_gradient::{Grad, ForwardNeumann}; +use alg_tools::convex::Zero; +use alg_tools::maputil::map3; +use alg_tools::direct_product::Pair; use crate::kernels::*; use crate::types::*; use crate::measures::*; -use crate::measures::merging::SpikeMerging; +use crate::measures::merging::{SpikeMerging,SpikeMergingMethod}; use crate::forward_model::*; +use crate::forward_model::sensor_grid::{ + SensorGrid, + SensorGridBT, + //SensorGridBTFN, + Sensor, + Spread, +}; + use crate::fb::{ FBConfig, + FBGenericConfig, pointsource_fb_reg, - FBMetaAlgorithm, - FBGenericConfig, + pointsource_fista_reg, +}; +use crate::sliding_fb::{ + SlidingFBConfig, + TransportConfig, + pointsource_sliding_fb_reg +}; +use crate::sliding_pdps::{ + SlidingPDPSConfig, + pointsource_sliding_pdps_pair +}; +use crate::forward_pdps::{ + ForwardPDPSConfig, + pointsource_forward_pdps_pair }; use crate::pdps::{ PDPSConfig, - L2Squared, pointsource_pdps_reg, }; use crate::frank_wolfe::{ FWConfig, FWVariant, pointsource_fw_reg, - WeightOptim, + //WeightOptim, }; -use crate::subproblem::InnerSettings; +use crate::subproblem::{InnerSettings, InnerMethod}; use crate::seminorms::*; use crate::plot::*; use crate::{AlgorithmOverrides, CommandLineArgs}; use crate::tolerance::Tolerance; -use crate::regularisation::{Regularisation, RadonRegTerm, NonnegRadonRegTerm}; +use crate::regularisation::{ + Regularisation, + RadonRegTerm, + NonnegRadonRegTerm +}; +use crate::dataterm::{ + L1, + L2Squared, +}; +use crate::prox_penalty::{ + RadonSquared, + //ProxPenalty, +}; +use alg_tools::norms::{L2, NormExponent}; +use alg_tools::operator_arithmetic::Weighted; +use anyhow::anyhow; + +/// Available proximal terms +#[derive(Copy, Clone, Debug, Serialize, Deserialize)] +pub enum ProxTerm { + /// Partial-to-wave operator 𝒟. + Wave, + /// Radon-norm squared + RadonSquared +} /// Available algorithms and their configurations #[derive(Copy, Clone, Debug, Serialize, Deserialize)] pub enum AlgorithmConfig<F : Float> { - FB(FBConfig<F>), + FB(FBConfig<F>, ProxTerm), + FISTA(FBConfig<F>, ProxTerm), FW(FWConfig<F>), - PDPS(PDPSConfig<F>), + PDPS(PDPSConfig<F>, ProxTerm), + SlidingFB(SlidingFBConfig<F>, ProxTerm), + ForwardPDPS(ForwardPDPSConfig<F>, ProxTerm), + SlidingPDPS(SlidingPDPSConfig<F>, ProxTerm), } fn unpack_tolerance<F : Float>(v : &Vec<F>) -> Tolerance<F> { @@ -83,6 +146,13 @@ impl<F : ClapFloat> AlgorithmConfig<F> { /// Override supported parameters based on the command line. pub fn cli_override(self, cli : &AlgorithmOverrides<F>) -> Self { + let override_merging = |g : SpikeMergingMethod<F>| { + SpikeMergingMethod { + enabled : cli.merge.unwrap_or(g.enabled), + radius : cli.merge_radius.unwrap_or(g.radius), + interp : cli.merge_interp.unwrap_or(g.interp), + } + }; let override_fb_generic = |g : FBGenericConfig<F>| { FBGenericConfig { bootstrap_insertions : cli.bootstrap_insertions @@ -90,37 +160,74 @@ .map_or(g.bootstrap_insertions, |n| Some((n[0], n[1]))), merge_every : cli.merge_every.unwrap_or(g.merge_every), - merging : cli.merging.clone().unwrap_or(g.merging), - final_merging : cli.final_merging.clone().unwrap_or(g.final_merging), + merging : override_merging(g.merging), + final_merging : cli.final_merging.unwrap_or(g.final_merging), + fitness_merging : cli.fitness_merging.unwrap_or(g.fitness_merging), tolerance: cli.tolerance.as_ref().map(unpack_tolerance).unwrap_or(g.tolerance), .. g } }; + let override_transport = |g : TransportConfig<F>| { + TransportConfig { + θ0 : cli.theta0.unwrap_or(g.θ0), + tolerance_mult_con: cli.transport_tolerance_pos.unwrap_or(g.tolerance_mult_con), + adaptation: cli.transport_adaptation.unwrap_or(g.adaptation), + .. g + } + }; use AlgorithmConfig::*; match self { - FB(fb) => FB(FBConfig { + FB(fb, prox) => FB(FBConfig { τ0 : cli.tau0.unwrap_or(fb.τ0), - insertion : override_fb_generic(fb.insertion), + generic : override_fb_generic(fb.generic), .. fb - }), - PDPS(pdps) => PDPS(PDPSConfig { + }, prox), + FISTA(fb, prox) => FISTA(FBConfig { + τ0 : cli.tau0.unwrap_or(fb.τ0), + generic : override_fb_generic(fb.generic), + .. fb + }, prox), + PDPS(pdps, prox) => PDPS(PDPSConfig { τ0 : cli.tau0.unwrap_or(pdps.τ0), σ0 : cli.sigma0.unwrap_or(pdps.σ0), acceleration : cli.acceleration.unwrap_or(pdps.acceleration), - insertion : override_fb_generic(pdps.insertion), + generic : override_fb_generic(pdps.generic), .. pdps - }), + }, prox), FW(fw) => FW(FWConfig { - merging : cli.merging.clone().unwrap_or(fw.merging), + merging : override_merging(fw.merging), tolerance : cli.tolerance.as_ref().map(unpack_tolerance).unwrap_or(fw.tolerance), .. fw - }) + }), + SlidingFB(sfb, prox) => SlidingFB(SlidingFBConfig { + τ0 : cli.tau0.unwrap_or(sfb.τ0), + transport : override_transport(sfb.transport), + insertion : override_fb_generic(sfb.insertion), + .. sfb + }, prox), + SlidingPDPS(spdps, prox) => SlidingPDPS(SlidingPDPSConfig { + τ0 : cli.tau0.unwrap_or(spdps.τ0), + σp0 : cli.sigmap0.unwrap_or(spdps.σp0), + σd0 : cli.sigma0.unwrap_or(spdps.σd0), + //acceleration : cli.acceleration.unwrap_or(pdps.acceleration), + transport : override_transport(spdps.transport), + insertion : override_fb_generic(spdps.insertion), + .. spdps + }, prox), + ForwardPDPS(fpdps, prox) => ForwardPDPS(ForwardPDPSConfig { + τ0 : cli.tau0.unwrap_or(fpdps.τ0), + σp0 : cli.sigmap0.unwrap_or(fpdps.σp0), + σd0 : cli.sigma0.unwrap_or(fpdps.σd0), + //acceleration : cli.acceleration.unwrap_or(pdps.acceleration), + insertion : override_fb_generic(fpdps.insertion), + .. fpdps + }, prox), } } } -/// Helper struct for tagging and [`AlgorithmConfig`] or [`Experiment`] with a name. +/// Helper struct for tagging and [`AlgorithmConfig`] or [`ExperimentV2`] with a name. #[derive(Clone, Debug, Serialize, Deserialize)] pub struct Named<Data> { pub name : String, @@ -146,24 +253,89 @@ /// The μPDPS primal-dual proximal splitting method #[clap(name = "pdps")] PDPS, + /// The sliding FB method + #[clap(name = "sliding_fb", alias = "sfb")] + SlidingFB, + /// The sliding PDPS method + #[clap(name = "sliding_pdps", alias = "spdps")] + SlidingPDPS, + /// The PDPS method with a forward step for the smooth function + #[clap(name = "forward_pdps", alias = "fpdps")] + ForwardPDPS, + + // Radon variants + + /// The μFB forward-backward method with radon-norm squared proximal term + #[clap(name = "radon_fb")] + RadonFB, + /// The μFISTA inertial forward-backward method with radon-norm squared proximal term + #[clap(name = "radon_fista")] + RadonFISTA, + /// The μPDPS primal-dual proximal splitting method with radon-norm squared proximal term + #[clap(name = "radon_pdps")] + RadonPDPS, + /// The sliding FB method with radon-norm squared proximal term + #[clap(name = "radon_sliding_fb", alias = "radon_sfb")] + RadonSlidingFB, + /// The sliding PDPS method with radon-norm squared proximal term + #[clap(name = "radon_sliding_pdps", alias = "radon_spdps")] + RadonSlidingPDPS, + /// The PDPS method with a forward step for the smooth function with radon-norm squared proximal term + #[clap(name = "radon_forward_pdps", alias = "radon_fpdps")] + RadonForwardPDPS, } impl DefaultAlgorithm { /// Returns the algorithm configuration corresponding to the algorithm shorthand pub fn default_config<F : Float>(&self) -> AlgorithmConfig<F> { use DefaultAlgorithm::*; + let radon_insertion = FBGenericConfig { + merging : SpikeMergingMethod{ interp : false, .. Default::default() }, + inner : InnerSettings { + method : InnerMethod::PDPS, // SSN not implemented + .. Default::default() + }, + .. Default::default() + }; match *self { - FB => AlgorithmConfig::FB(Default::default()), - FISTA => AlgorithmConfig::FB(FBConfig{ - meta : FBMetaAlgorithm::InertiaFISTA, - .. Default::default() - }), + FB => AlgorithmConfig::FB(Default::default(), ProxTerm::Wave), + FISTA => AlgorithmConfig::FISTA(Default::default(), ProxTerm::Wave), FW => AlgorithmConfig::FW(Default::default()), FWRelax => AlgorithmConfig::FW(FWConfig{ variant : FWVariant::Relaxed, .. Default::default() }), - PDPS => AlgorithmConfig::PDPS(Default::default()), + PDPS => AlgorithmConfig::PDPS(Default::default(), ProxTerm::Wave), + SlidingFB => AlgorithmConfig::SlidingFB(Default::default(), ProxTerm::Wave), + SlidingPDPS => AlgorithmConfig::SlidingPDPS(Default::default(), ProxTerm::Wave), + ForwardPDPS => AlgorithmConfig::ForwardPDPS(Default::default(), ProxTerm::Wave), + + // Radon variants + + RadonFB => AlgorithmConfig::FB( + FBConfig{ generic : radon_insertion, ..Default::default() }, + ProxTerm::RadonSquared + ), + RadonFISTA => AlgorithmConfig::FISTA( + FBConfig{ generic : radon_insertion, ..Default::default() }, + ProxTerm::RadonSquared + ), + RadonPDPS => AlgorithmConfig::PDPS( + PDPSConfig{ generic : radon_insertion, ..Default::default() }, + ProxTerm::RadonSquared + ), + RadonSlidingFB => AlgorithmConfig::SlidingFB( + SlidingFBConfig{ insertion : radon_insertion, ..Default::default() }, + ProxTerm::RadonSquared + ), + RadonSlidingPDPS => AlgorithmConfig::SlidingPDPS( + SlidingPDPSConfig{ insertion : radon_insertion, ..Default::default() }, + ProxTerm::RadonSquared + ), + RadonForwardPDPS => AlgorithmConfig::ForwardPDPS( + ForwardPDPSConfig{ insertion : radon_insertion, ..Default::default() }, + ProxTerm::RadonSquared + ), } } @@ -201,6 +373,12 @@ Iter, } +impl Default for PlotLevel { + fn default() -> Self { + Self::Data + } +} + type DefaultBT<F, const N : usize> = BT< DynamicDepth, F, @@ -223,7 +401,8 @@ iter : usize, cpu_time : f64, value : F, - post_value : F, + relative_value : F, + //post_value : F, n_spikes : usize, inner_iters : usize, merged : usize, @@ -278,7 +457,7 @@ /// Struct for experiment configurations #[derive(Debug, Clone, Serialize)] pub struct ExperimentV2<F, NoiseDistr, S, K, P, const N : usize> -where F : Float, +where F : Float + ClapFloat, [usize; N] : Serialize, NoiseDistr : Distribution<F>, S : Sensor<F, N>, @@ -300,7 +479,7 @@ /// Kernel $ρ$ of $𝒟$. pub kernel : K, /// True point sources - pub μ_hat : DiscreteMeasure<Loc<F, N>, F>, + pub μ_hat : RNDM<F, N>, /// Regularisation term and parameter pub regularisation : Regularisation<F>, /// For plotting : how wide should the kernels be plotted @@ -308,8 +487,27 @@ /// Data term pub dataterm : DataTerm, /// A map of default configurations for algorithms - #[serde(skip)] - pub algorithm_defaults : HashMap<DefaultAlgorithm, AlgorithmConfig<F>>, + pub algorithm_overrides : HashMap<DefaultAlgorithm, AlgorithmOverrides<F>>, + /// Default merge radius + pub default_merge_radius : F, +} + +#[derive(Debug, Clone, Serialize)] +pub struct ExperimentBiased<F, NoiseDistr, S, K, P, B, const N : usize> +where F : Float + ClapFloat, + [usize; N] : Serialize, + NoiseDistr : Distribution<F>, + S : Sensor<F, N>, + P : Spread<F, N>, + K : SimpleConvolutionKernel<F, N>, + B : Mapping<Loc<F, N>, Codomain = F> + Serialize + std::fmt::Debug, +{ + /// Basic setup + pub base : ExperimentV2<F, NoiseDistr, S, K, P, N>, + /// Weight of TV term + pub λ : F, + /// Bias function + pub bias : B, } /// Trait for runnable experiments @@ -319,41 +517,190 @@ algs : Option<Vec<Named<AlgorithmConfig<F>>>>) -> DynError; /// Return algorithm default config - fn algorithm_defaults(&self, alg : DefaultAlgorithm, cli : &AlgorithmOverrides<F>) - -> Named<AlgorithmConfig<F>>; + fn algorithm_overrides(&self, alg : DefaultAlgorithm) -> AlgorithmOverrides<F>; +} + +/// Helper function to print experiment start message and save setup. +/// Returns saving prefix. +fn start_experiment<E, S>( + experiment : &Named<E>, + cli : &CommandLineArgs, + stats : S, +) -> DynResult<String> +where + E : Serialize + std::fmt::Debug, + S : Serialize, +{ + let Named { name : experiment_name, data } = experiment; + + println!("{}\n{}", + format!("Performing experiment {}…", experiment_name).cyan(), + format!("Experiment settings: {}", serde_json::to_string(&data)?).bright_black()); + + // Set up output directory + let prefix = format!("{}/{}/", cli.outdir, experiment_name); + + // Save experiment configuration and statistics + let mkname_e = |t| format!("{prefix}{t}.json", prefix = prefix, t = t); + std::fs::create_dir_all(&prefix)?; + write_json(mkname_e("experiment"), experiment)?; + write_json(mkname_e("config"), cli)?; + write_json(mkname_e("stats"), &stats)?; + + Ok(prefix) +} + +/// Error codes for running an algorithm on an experiment. +enum RunError { + /// Algorithm not implemented for this experiment + NotImplemented, } -// *** macro boilerplate *** -macro_rules! impl_experiment { -($type:ident, $reg_field:ident, $reg_convert:path) => { -// *** macro *** -impl<F, NoiseDistr, S, K, P, const N : usize> RunnableExperiment<F> for -Named<$type<F, NoiseDistr, S, K, P, N>> -where F : ClapFloat + nalgebra::RealField + ToNalgebraRealField<MixedType=F>, - [usize; N] : Serialize, - S : Sensor<F, N> + Copy + Serialize + std::fmt::Debug, - P : Spread<F, N> + Copy + Serialize + std::fmt::Debug, - Convolution<S, P>: Spread<F, N> + Bounded<F> + LocalAnalysis<F, Bounds<F>, N> + Copy, - AutoConvolution<P> : BoundedBy<F, K>, - K : SimpleConvolutionKernel<F, N> + LocalAnalysis<F, Bounds<F>, N> - + Copy + Serialize + std::fmt::Debug, - Cube<F, N>: P2Minimise<Loc<F, N>, F> + SetOrd, - PlotLookup : Plotting<N>, - DefaultBT<F, N> : SensorGridBT<F, S, P, N, Depth=DynamicDepth> + BTSearch<F, N>, - BTNodeLookup: BTNode<F, usize, Bounds<F>, N>, - DiscreteMeasure<Loc<F, N>, F> : SpikeMerging<F>, - NoiseDistr : Distribution<F> + Serialize + std::fmt::Debug { +use RunError::*; + +type DoRunAllIt<'a, F, const N : usize> = LoggingIteratorFactory< + 'a, + Timed<IterInfo<F, N>>, + TimingIteratorFactory<BasicAlgIteratorFactory<IterInfo<F, N>>> +>; + +/// Helper function to run all algorithms on an experiment. +fn do_runall<F : Float + for<'b> Deserialize<'b>, Z, const N : usize>( + experiment_name : &String, + prefix : &String, + cli : &CommandLineArgs, + algorithms : Vec<Named<AlgorithmConfig<F>>>, + plotgrid : LinSpace<Loc<F, N>, [usize; N]>, + mut save_extra : impl FnMut(String, Z) -> DynError, + mut do_alg : impl FnMut( + &AlgorithmConfig<F>, + DoRunAllIt<F, N>, + SeqPlotter<F, N>, + String, + ) -> Result<(RNDM<F, N>, Z), RunError>, +) -> DynError +where + PlotLookup : Plotting<N>, +{ + let mut logs = Vec::new(); + + let iterator_options = AlgIteratorOptions{ + max_iter : cli.max_iter, + verbose_iter : cli.verbose_iter + .map_or(Verbose::LogarithmicCap{base : 10, cap : 2}, + |n| Verbose::Every(n)), + quiet : cli.quiet, + }; + + // Run the algorithm(s) + for named @ Named { name : alg_name, data : alg } in algorithms.iter() { + let this_prefix = format!("{}{}/", prefix, alg_name); + + // Create Logger and IteratorFactory + let mut logger = Logger::new(); + let iterator = iterator_options.instantiate() + .timed() + .into_log(&mut logger); + + let running = if !cli.quiet { + format!("{}\n{}\n{}\n", + format!("Running {} on experiment {}…", alg_name, experiment_name).cyan(), + format!("Iteration settings: {}", serde_json::to_string(&iterator_options)?).bright_black(), + format!("Algorithm settings: {}", serde_json::to_string(&alg)?).bright_black()) + } else { + "".to_string() + }; + // + // The following is for postprocessing, which has been disabled anyway. + // + // let reg : Box<dyn WeightOptim<_, _, _, N>> = match regularisation { + // Regularisation::Radon(α) => Box::new(RadonRegTerm(α)), + // Regularisation::NonnegRadon(α) => Box::new(NonnegRadonRegTerm(α)), + // }; + //let findim_data = reg.prepare_optimise_weights(&opA, &b); + //let inner_config : InnerSettings<F> = Default::default(); + //let inner_it = inner_config.iterator_options; + + // Create plotter and directory if needed. + let plot_count = if cli.plot >= PlotLevel::Iter { 2000 } else { 0 }; + let plotter = SeqPlotter::new(this_prefix, plot_count, plotgrid.clone()); + + let start = Instant::now(); + let start_cpu = ProcessTime::now(); - fn algorithm_defaults(&self, alg : DefaultAlgorithm, cli : &AlgorithmOverrides<F>) - -> Named<AlgorithmConfig<F>> { - alg.to_named( - self.data - .algorithm_defaults - .get(&alg) - .map_or_else(|| alg.default_config(), - |config| config.clone()) - .cli_override(cli) - ) + let (μ, z) = match do_alg(alg, iterator, plotter, running) { + Ok(μ) => μ, + Err(RunError::NotImplemented) => { + let msg = format!("Algorithm “{alg_name}” not implemented for {experiment_name}. \ + Skipping.").red(); + eprintln!("{}", msg); + continue + } + }; + + let elapsed = start.elapsed().as_secs_f64(); + let cpu_time = start_cpu.elapsed().as_secs_f64(); + + println!("{}", format!("Elapsed {elapsed}s (CPU time {cpu_time}s)… ").yellow()); + + // Save results + println!("{}", "Saving results …".green()); + + let mkname = |t| format!("{prefix}{alg_name}_{t}"); + + write_json(mkname("config.json"), &named)?; + write_json(mkname("stats.json"), &AlgorithmStats { cpu_time, elapsed })?; + μ.write_csv(mkname("reco.txt"))?; + save_extra(mkname(""), z)?; + //logger.write_csv(mkname("log.txt"))?; + logs.push((mkname("log.txt"), logger)); + } + + save_logs(logs, format!("{prefix}valuerange.json"), cli.load_valuerange) +} + +#[replace_float_literals(F::cast_from(literal))] +impl<F, NoiseDistr, S, K, P, /*PreadjointCodomain, */ const N : usize> RunnableExperiment<F> for +Named<ExperimentV2<F, NoiseDistr, S, K, P, N>> +where + F : ClapFloat + nalgebra::RealField + ToNalgebraRealField<MixedType=F> + + Default + for<'b> Deserialize<'b>, + [usize; N] : Serialize, + S : Sensor<F, N> + Copy + Serialize + std::fmt::Debug, + P : Spread<F, N> + Copy + Serialize + std::fmt::Debug, + Convolution<S, P>: Spread<F, N> + Bounded<F> + LocalAnalysis<F, Bounds<F>, N> + Copy + // TODO: shold not have differentiability as a requirement, but + // decide availability of sliding based on it. + //+ for<'b> Differentiable<&'b Loc<F, N>, Output = Loc<F, N>>, + // TODO: very weird that rust only compiles with Differentiable + // instead of the above one on references, which is required by + // poitsource_sliding_fb_reg. + + DifferentiableRealMapping<F, N> + + Lipschitz<L2, FloatType=F>, + for<'b> <Convolution<S, P> as DifferentiableMapping<Loc<F,N>>>::Differential<'b> : Lipschitz<L2, FloatType=F>, // TODO: should not be required generally, only for sliding_fb. + AutoConvolution<P> : BoundedBy<F, K>, + K : SimpleConvolutionKernel<F, N> + + LocalAnalysis<F, Bounds<F>, N> + + Copy + Serialize + std::fmt::Debug, + Cube<F, N>: P2Minimise<Loc<F, N>, F> + SetOrd, + PlotLookup : Plotting<N>, + DefaultBT<F, N> : SensorGridBT<F, S, P, N, Depth=DynamicDepth> + BTSearch<F, N>, + BTNodeLookup: BTNode<F, usize, Bounds<F>, N>, + RNDM<F, N> : SpikeMerging<F>, + NoiseDistr : Distribution<F> + Serialize + std::fmt::Debug, + // DefaultSG<F, S, P, N> : ForwardModel<RNDM<F, N>, F, PreadjointCodomain = PreadjointCodomain, Observable=DVector<F::MixedType>>, + // PreadjointCodomain : Space + Bounded<F> + DifferentiableRealMapping<F, N>, + // DefaultSeminormOp<F, K, N> : ProxPenalty<F, PreadjointCodomain, RadonRegTerm<F>, N>, + // DefaultSeminormOp<F, K, N> : ProxPenalty<F, PreadjointCodomain, NonnegRadonRegTerm<F>, N>, + // RadonSquared : ProxPenalty<F, PreadjointCodomain, RadonRegTerm<F>, N>, + // RadonSquared : ProxPenalty<F, PreadjointCodomain, NonnegRadonRegTerm<F>, N>, +{ + + fn algorithm_overrides(&self, alg : DefaultAlgorithm) -> AlgorithmOverrides<F> { + AlgorithmOverrides { + merge_radius : Some(self.data.default_merge_radius), + .. self.data.algorithm_overrides.get(&alg).cloned().unwrap_or(Default::default()) + } } fn runall(&self, cli : &CommandLineArgs, @@ -361,31 +708,15 @@ // Get experiment configuration let &Named { name : ref experiment_name, - data : $type { + data : ExperimentV2 { domain, sensor_count, ref noise_distr, sensor, spread, kernel, - ref μ_hat, /*regularisation,*/ kernel_plot_width, dataterm, noise_seed, + ref μ_hat, regularisation, kernel_plot_width, dataterm, noise_seed, .. } } = self; - #[allow(deprecated)] - let regularisation = $reg_convert(self.data.$reg_field); - - println!("{}\n{}", - format!("Performing experiment {}…", experiment_name).cyan(), - format!("{:?}", &self.data).bright_black()); - - // Set up output directory - let prefix = format!("{}/{}/", cli.outdir, self.name); // Set up algorithms - let iterator_options = AlgIteratorOptions{ - max_iter : cli.max_iter, - verbose_iter : cli.verbose_iter - .map_or(Verbose::Logarithmic(10), - |n| Verbose::Every(n)), - quiet : cli.quiet, - }; - let algorithms = match (algs, self.data.dataterm) { + let algorithms = match (algs, dataterm) { (Some(algs), _) => algs, (None, DataTerm::L2Squared) => vec![DefaultAlgorithm::FB.get_named()], (None, DataTerm::L1) => vec![DefaultAlgorithm::PDPS.get_named()], @@ -407,186 +738,492 @@ // overloading log10 and conflicting with standard NumTraits one. let stats = ExperimentStats::new(&b, &noise); - // Save experiment configuration and statistics - let mkname_e = |t| format!("{prefix}{t}.json", prefix = prefix, t = t); - std::fs::create_dir_all(&prefix)?; - write_json(mkname_e("experiment"), self)?; - write_json(mkname_e("config"), cli)?; - write_json(mkname_e("stats"), &stats)?; + let prefix = start_experiment(&self, cli, stats)?; plotall(cli, &prefix, &domain, &sensor, &kernel, &spread, &μ_hat, &op𝒟, &opA, &b_hat, &b, kernel_plot_width)?; - // Run the algorithm(s) - for named @ Named { name : alg_name, data : alg } in algorithms.iter() { - let this_prefix = format!("{}{}/", prefix, alg_name); + let plotgrid = lingrid(&domain, &[if N==1 { 1000 } else { 100 }; N]); + + let save_extra = |_, ()| Ok(()); - let running = || if !cli.quiet { - println!("{}\n{}\n{}", - format!("Running {} on experiment {}…", alg_name, experiment_name).cyan(), - format!("{:?}", iterator_options).bright_black(), - format!("{:?}", alg).bright_black()); - }; - let not_implemented = || { - let msg = format!("Algorithm “{alg_name}” not implemented for \ - dataterm {dataterm:?} and regularisation {regularisation:?}. \ - Skipping.").red(); - eprintln!("{}", msg); - }; - // Create Logger and IteratorFactory - let mut logger = Logger::new(); - let reg : Box<dyn WeightOptim<_, _, _, N>> = match regularisation { - Regularisation::Radon(α) => Box::new(RadonRegTerm(α)), - Regularisation::NonnegRadon(α) => Box::new(NonnegRadonRegTerm(α)), - }; - let findim_data = reg.prepare_optimise_weights(&opA, &b); - let inner_config : InnerSettings<F> = Default::default(); - let inner_it = inner_config.iterator_options; - let logmap = |iter, Timed { cpu_time, data }| { - let IterInfo { - value, - n_spikes, - inner_iters, - merged, - pruned, - postprocessing, - this_iters, - .. - } = data; - let post_value = match (postprocessing, dataterm) { - (Some(mut μ), DataTerm::L2Squared) => { - // Comparison postprocessing is only implemented for the case handled - // by the FW variants. - reg.optimise_weights( - &mut μ, &opA, &b, &findim_data, &inner_config, - inner_it - ); - dataterm.value_at_residual(opA.apply(&μ) - &b) - + regularisation.apply(&μ) - }, - _ => value, - }; - CSVLog { - iter, - value, - post_value, - n_spikes, - cpu_time : cpu_time.as_secs_f64(), - inner_iters, - merged, - pruned, - this_iters - } - }; - let iterator = iterator_options.instantiate() - .timed() - .mapped(logmap) - .into_log(&mut logger); - let plotgrid = lingrid(&domain, &[if N==1 { 1000 } else { 100 }; N]); - - // Create plotter and directory if needed. - let plot_count = if cli.plot >= PlotLevel::Iter { 2000 } else { 0 }; - let plotter = SeqPlotter::new(this_prefix, plot_count, plotgrid); - - // Run the algorithm - let start = Instant::now(); - let start_cpu = ProcessTime::now(); + do_runall(experiment_name, &prefix, cli, algorithms, plotgrid, save_extra, + |alg, iterator, plotter, running| + { let μ = match alg { - AlgorithmConfig::FB(ref algconfig) => { - match (regularisation, dataterm) { - (Regularisation::NonnegRadon(α), DataTerm::L2Squared) => { - running(); + AlgorithmConfig::FB(ref algconfig, prox) => { + match (regularisation, dataterm, prox) { + (Regularisation::NonnegRadon(α), DataTerm::L2Squared, ProxTerm::Wave) => Ok({ + print!("{running}"); pointsource_fb_reg( &opA, &b, NonnegRadonRegTerm(α), &op𝒟, algconfig, iterator, plotter ) - }, - (Regularisation::Radon(α), DataTerm::L2Squared) => { - running(); + }), + (Regularisation::Radon(α), DataTerm::L2Squared, ProxTerm::Wave) => Ok({ + print!("{running}"); pointsource_fb_reg( &opA, &b, RadonRegTerm(α), &op𝒟, algconfig, iterator, plotter ) - }, - _ => { - not_implemented(); - continue - } + }), + (Regularisation::NonnegRadon(α), DataTerm::L2Squared, ProxTerm::RadonSquared) => Ok({ + print!("{running}"); + pointsource_fb_reg( + &opA, &b, NonnegRadonRegTerm(α), &RadonSquared, algconfig, + iterator, plotter + ) + }), + (Regularisation::Radon(α), DataTerm::L2Squared, ProxTerm::RadonSquared) => Ok({ + print!("{running}"); + pointsource_fb_reg( + &opA, &b, RadonRegTerm(α), &RadonSquared, algconfig, + iterator, plotter + ) + }), + _ => Err(NotImplemented) } }, - AlgorithmConfig::PDPS(ref algconfig) => { - running(); - match (regularisation, dataterm) { - (Regularisation::NonnegRadon(α), DataTerm::L2Squared) => { + AlgorithmConfig::FISTA(ref algconfig, prox) => { + match (regularisation, dataterm, prox) { + (Regularisation::NonnegRadon(α), DataTerm::L2Squared, ProxTerm::Wave) => Ok({ + print!("{running}"); + pointsource_fista_reg( + &opA, &b, NonnegRadonRegTerm(α), &op𝒟, algconfig, + iterator, plotter + ) + }), + (Regularisation::Radon(α), DataTerm::L2Squared, ProxTerm::Wave) => Ok({ + print!("{running}"); + pointsource_fista_reg( + &opA, &b, RadonRegTerm(α), &op𝒟, algconfig, + iterator, plotter + ) + }), + (Regularisation::NonnegRadon(α), DataTerm::L2Squared, ProxTerm::RadonSquared) => Ok({ + print!("{running}"); + pointsource_fista_reg( + &opA, &b, NonnegRadonRegTerm(α), &RadonSquared, algconfig, + iterator, plotter + ) + }), + (Regularisation::Radon(α), DataTerm::L2Squared, ProxTerm::RadonSquared) => Ok({ + print!("{running}"); + pointsource_fista_reg( + &opA, &b, RadonRegTerm(α), &RadonSquared, algconfig, + iterator, plotter + ) + }), + _ => Err(NotImplemented), + } + }, + AlgorithmConfig::SlidingFB(ref algconfig, prox) => { + match (regularisation, dataterm, prox) { + (Regularisation::NonnegRadon(α), DataTerm::L2Squared, ProxTerm::Wave) => Ok({ + print!("{running}"); + pointsource_sliding_fb_reg( + &opA, &b, NonnegRadonRegTerm(α), &op𝒟, algconfig, + iterator, plotter + ) + }), + (Regularisation::Radon(α), DataTerm::L2Squared, ProxTerm::Wave) => Ok({ + print!("{running}"); + pointsource_sliding_fb_reg( + &opA, &b, RadonRegTerm(α), &op𝒟, algconfig, + iterator, plotter + ) + }), + (Regularisation::NonnegRadon(α), DataTerm::L2Squared, ProxTerm::RadonSquared) => Ok({ + print!("{running}"); + pointsource_sliding_fb_reg( + &opA, &b, NonnegRadonRegTerm(α), &RadonSquared, algconfig, + iterator, plotter + ) + }), + (Regularisation::Radon(α), DataTerm::L2Squared, ProxTerm::RadonSquared) => Ok({ + print!("{running}"); + pointsource_sliding_fb_reg( + &opA, &b, RadonRegTerm(α), &RadonSquared, algconfig, + iterator, plotter + ) + }), + _ => Err(NotImplemented), + } + }, + AlgorithmConfig::PDPS(ref algconfig, prox) => { + print!("{running}"); + match (regularisation, dataterm, prox) { + (Regularisation::NonnegRadon(α), DataTerm::L2Squared, ProxTerm::Wave) => Ok({ pointsource_pdps_reg( &opA, &b, NonnegRadonRegTerm(α), &op𝒟, algconfig, iterator, plotter, L2Squared ) - }, - (Regularisation::Radon(α),DataTerm::L2Squared) => { + }), + (Regularisation::Radon(α),DataTerm::L2Squared, ProxTerm::Wave) => Ok({ pointsource_pdps_reg( &opA, &b, RadonRegTerm(α), &op𝒟, algconfig, iterator, plotter, L2Squared ) - }, - (Regularisation::NonnegRadon(α), DataTerm::L1) => { + }), + (Regularisation::NonnegRadon(α), DataTerm::L1, ProxTerm::Wave) => Ok({ pointsource_pdps_reg( &opA, &b, NonnegRadonRegTerm(α), &op𝒟, algconfig, iterator, plotter, L1 ) - }, - (Regularisation::Radon(α), DataTerm::L1) => { + }), + (Regularisation::Radon(α), DataTerm::L1, ProxTerm::Wave) => Ok({ pointsource_pdps_reg( &opA, &b, RadonRegTerm(α), &op𝒟, algconfig, iterator, plotter, L1 ) - }, + }), + (Regularisation::NonnegRadon(α), DataTerm::L2Squared, ProxTerm::RadonSquared) => Ok({ + pointsource_pdps_reg( + &opA, &b, NonnegRadonRegTerm(α), &RadonSquared, algconfig, + iterator, plotter, L2Squared + ) + }), + (Regularisation::Radon(α),DataTerm::L2Squared, ProxTerm::RadonSquared) => Ok({ + pointsource_pdps_reg( + &opA, &b, RadonRegTerm(α), &RadonSquared, algconfig, + iterator, plotter, L2Squared + ) + }), + (Regularisation::NonnegRadon(α), DataTerm::L1, ProxTerm::RadonSquared) => Ok({ + pointsource_pdps_reg( + &opA, &b, NonnegRadonRegTerm(α), &RadonSquared, algconfig, + iterator, plotter, L1 + ) + }), + (Regularisation::Radon(α), DataTerm::L1, ProxTerm::RadonSquared) => Ok({ + pointsource_pdps_reg( + &opA, &b, RadonRegTerm(α), &RadonSquared, algconfig, + iterator, plotter, L1 + ) + }), + // _ => Err(NotImplemented), } }, AlgorithmConfig::FW(ref algconfig) => { match (regularisation, dataterm) { - (Regularisation::Radon(α), DataTerm::L2Squared) => { - running(); + (Regularisation::Radon(α), DataTerm::L2Squared) => Ok({ + print!("{running}"); pointsource_fw_reg(&opA, &b, RadonRegTerm(α), algconfig, iterator, plotter) - }, - (Regularisation::NonnegRadon(α), DataTerm::L2Squared) => { - running(); + }), + (Regularisation::NonnegRadon(α), DataTerm::L2Squared) => Ok({ + print!("{running}"); pointsource_fw_reg(&opA, &b, NonnegRadonRegTerm(α), algconfig, iterator, plotter) - }, - _ => { - not_implemented(); - continue - } + }), + _ => Err(NotImplemented), } - } - }; - - let elapsed = start.elapsed().as_secs_f64(); - let cpu_time = start_cpu.elapsed().as_secs_f64(); - - println!("{}", format!("Elapsed {elapsed}s (CPU time {cpu_time}s)… ").yellow()); - - // Save results - println!("{}", "Saving results…".green()); - - let mkname = |t| format!("{prefix}{alg_name}_{t}"); - - write_json(mkname("config.json"), &named)?; - write_json(mkname("stats.json"), &AlgorithmStats { cpu_time, elapsed })?; - μ.write_csv(mkname("reco.txt"))?; - logger.write_csv(mkname("log.txt"))?; - } - - Ok(()) + }, + _ => Err(NotImplemented), + }?; + Ok((μ, ())) + }) } } -// *** macro end boiler plate *** -}} -// *** actual code *** + + +#[replace_float_literals(F::cast_from(literal))] +impl<F, NoiseDistr, S, K, P, B, /*PreadjointCodomain,*/ const N : usize> RunnableExperiment<F> for +Named<ExperimentBiased<F, NoiseDistr, S, K, P, B, N>> +where + F : ClapFloat + nalgebra::RealField + ToNalgebraRealField<MixedType=F> + + Default + for<'b> Deserialize<'b>, + [usize; N] : Serialize, + S : Sensor<F, N> + Copy + Serialize + std::fmt::Debug, + P : Spread<F, N> + Copy + Serialize + std::fmt::Debug, + Convolution<S, P>: Spread<F, N> + Bounded<F> + LocalAnalysis<F, Bounds<F>, N> + Copy + // TODO: shold not have differentiability as a requirement, but + // decide availability of sliding based on it. + //+ for<'b> Differentiable<&'b Loc<F, N>, Output = Loc<F, N>>, + // TODO: very weird that rust only compiles with Differentiable + // instead of the above one on references, which is required by + // poitsource_sliding_fb_reg. + + DifferentiableRealMapping<F, N> + + Lipschitz<L2, FloatType=F>, + for<'b> <Convolution<S, P> as DifferentiableMapping<Loc<F,N>>>::Differential<'b> : Lipschitz<L2, FloatType=F>, // TODO: should not be required generally, only for sliding_fb. + AutoConvolution<P> : BoundedBy<F, K>, + K : SimpleConvolutionKernel<F, N> + + LocalAnalysis<F, Bounds<F>, N> + + Copy + Serialize + std::fmt::Debug, + Cube<F, N>: P2Minimise<Loc<F, N>, F> + SetOrd, + PlotLookup : Plotting<N>, + DefaultBT<F, N> : SensorGridBT<F, S, P, N, Depth=DynamicDepth> + BTSearch<F, N>, + BTNodeLookup: BTNode<F, usize, Bounds<F>, N>, + RNDM<F, N> : SpikeMerging<F>, + NoiseDistr : Distribution<F> + Serialize + std::fmt::Debug, + B : Mapping<Loc<F, N>, Codomain = F> + Serialize + std::fmt::Debug, + // DefaultSG<F, S, P, N> : ForwardModel<RNDM<F, N>, F, PreadjointCodomain = PreadjointCodomain, Observable=DVector<F::MixedType>>, + // PreadjointCodomain : Bounded<F> + DifferentiableRealMapping<F, N>, + // DefaultSeminormOp<F, K, N> : ProxPenalty<F, PreadjointCodomain, RadonRegTerm<F>, N>, + // DefaultSeminormOp<F, K, N> : ProxPenalty<F, PreadjointCodomain, NonnegRadonRegTerm<F>, N>, + // RadonSquared : ProxPenalty<F, PreadjointCodomain, RadonRegTerm<F>, N>, + // RadonSquared : ProxPenalty<F, PreadjointCodomain, NonnegRadonRegTerm<F>, N>, +{ + + fn algorithm_overrides(&self, alg : DefaultAlgorithm) -> AlgorithmOverrides<F> { + AlgorithmOverrides { + merge_radius : Some(self.data.base.default_merge_radius), + .. self.data.base.algorithm_overrides.get(&alg).cloned().unwrap_or(Default::default()) + } + } + + fn runall(&self, cli : &CommandLineArgs, + algs : Option<Vec<Named<AlgorithmConfig<F>>>>) -> DynError { + // Get experiment configuration + let &Named { + name : ref experiment_name, + data : ExperimentBiased { + λ, + ref bias, + base : ExperimentV2 { + domain, sensor_count, ref noise_distr, sensor, spread, kernel, + ref μ_hat, regularisation, kernel_plot_width, dataterm, noise_seed, + .. + } + } + } = self; + + // Set up algorithms + let algorithms = match (algs, dataterm) { + (Some(algs), _) => algs, + _ => vec![DefaultAlgorithm::SlidingPDPS.get_named()], + }; + + // Set up operators + let depth = DynamicDepth(8); + let opA = DefaultSG::new(domain, sensor_count, sensor, spread, depth); + let op𝒟 = DefaultSeminormOp::new(depth, domain, kernel); + let opAext = RowOp(opA.clone(), IdOp::new()); + let fnR = Zero::new(); + let h = map3(domain.span_start(), domain.span_end(), sensor_count, + |a, b, n| (b-a)/F::cast_from(n)) + .into_iter() + .reduce(NumTraitsFloat::max) + .unwrap(); + let z = DVector::zeros(sensor_count.iter().product()); + let opKz = Grad::new_for(&z, h, sensor_count, ForwardNeumann).unwrap(); + let y = opKz.apply(&z); + let fnH = Weighted{ base_fn : L1.as_mapping(), weight : λ}; // TODO: L_{2,1} + // let zero_y = y.clone(); + // let zeroBTFN = opA.preadjoint().apply(&zero_y); + // let opKμ = ZeroOp::new(&zero_y, zeroBTFN); + + // Set up random number generator. + let mut rng = StdRng::seed_from_u64(noise_seed); + + // Generate the data and calculate SSNR statistic + let bias_vec = DVector::from_vec(opA.grid() + .into_iter() + .map(|v| bias.apply(v)) + .collect::<Vec<F>>()); + let b_hat : DVector<_> = opA.apply(μ_hat) + &bias_vec; + let noise = DVector::from_distribution(b_hat.len(), &noise_distr, &mut rng); + let b = &b_hat + &noise; + // Need to wrap calc_ssnr into a function to hide ultra-lame nalgebra::RealField + // overloading log10 and conflicting with standard NumTraits one. + let stats = ExperimentStats::new(&b, &noise); + + let prefix = start_experiment(&self, cli, stats)?; + + plotall(cli, &prefix, &domain, &sensor, &kernel, &spread, + &μ_hat, &op𝒟, &opA, &b_hat, &b, kernel_plot_width)?; + + opA.write_observable(&bias_vec, format!("{prefix}bias"))?; + + let plotgrid = lingrid(&domain, &[if N==1 { 1000 } else { 100 }; N]); + + let save_extra = |prefix, z| opA.write_observable(&z, format!("{prefix}z")); -impl_experiment!(ExperimentV2, regularisation, std::convert::identity); + // Run the algorithms + do_runall(experiment_name, &prefix, cli, algorithms, plotgrid, save_extra, + |alg, iterator, plotter, running| + { + let Pair(μ, z) = match alg { + AlgorithmConfig::ForwardPDPS(ref algconfig, prox) => { + match (regularisation, dataterm, prox) { + (Regularisation::NonnegRadon(α), DataTerm::L2Squared, ProxTerm::Wave) => Ok({ + print!("{running}"); + pointsource_forward_pdps_pair( + &opAext, &b, NonnegRadonRegTerm(α), &op𝒟, algconfig, + iterator, plotter, + /* opKμ, */ &opKz, &fnR, &fnH, z.clone(), y.clone(), + ) + }), + (Regularisation::Radon(α), DataTerm::L2Squared, ProxTerm::Wave) => Ok({ + print!("{running}"); + pointsource_forward_pdps_pair( + &opAext, &b, RadonRegTerm(α), &op𝒟, algconfig, + iterator, plotter, + /* opKμ, */ &opKz, &fnR, &fnH, z.clone(), y.clone(), + ) + }), + (Regularisation::NonnegRadon(α), DataTerm::L2Squared, ProxTerm::RadonSquared) => Ok({ + print!("{running}"); + pointsource_forward_pdps_pair( + &opAext, &b, NonnegRadonRegTerm(α), &RadonSquared, algconfig, + iterator, plotter, + /* opKμ, */ &opKz, &fnR, &fnH, z.clone(), y.clone(), + ) + }), + (Regularisation::Radon(α), DataTerm::L2Squared, ProxTerm::RadonSquared) => Ok({ + print!("{running}"); + pointsource_forward_pdps_pair( + &opAext, &b, RadonRegTerm(α), &RadonSquared, algconfig, + iterator, plotter, + /* opKμ, */ &opKz, &fnR, &fnH, z.clone(), y.clone(), + ) + }), + _ => Err(NotImplemented) + } + }, + AlgorithmConfig::SlidingPDPS(ref algconfig, prox) => { + match (regularisation, dataterm, prox) { + (Regularisation::NonnegRadon(α), DataTerm::L2Squared, ProxTerm::Wave) => Ok({ + print!("{running}"); + pointsource_sliding_pdps_pair( + &opAext, &b, NonnegRadonRegTerm(α), &op𝒟, algconfig, + iterator, plotter, + /* opKμ, */ &opKz, &fnR, &fnH, z.clone(), y.clone(), + ) + }), + (Regularisation::Radon(α), DataTerm::L2Squared, ProxTerm::Wave) => Ok({ + print!("{running}"); + pointsource_sliding_pdps_pair( + &opAext, &b, RadonRegTerm(α), &op𝒟, algconfig, + iterator, plotter, + /* opKμ, */ &opKz, &fnR, &fnH, z.clone(), y.clone(), + ) + }), + (Regularisation::NonnegRadon(α), DataTerm::L2Squared, ProxTerm::RadonSquared) => Ok({ + print!("{running}"); + pointsource_sliding_pdps_pair( + &opAext, &b, NonnegRadonRegTerm(α), &RadonSquared, algconfig, + iterator, plotter, + /* opKμ, */ &opKz, &fnR, &fnH, z.clone(), y.clone(), + ) + }), + (Regularisation::Radon(α), DataTerm::L2Squared, ProxTerm::RadonSquared) => Ok({ + print!("{running}"); + pointsource_sliding_pdps_pair( + &opAext, &b, RadonRegTerm(α), &RadonSquared, algconfig, + iterator, plotter, + /* opKμ, */ &opKz, &fnR, &fnH, z.clone(), y.clone(), + ) + }), + _ => Err(NotImplemented) + } + }, + _ => Err(NotImplemented) + }?; + Ok((μ, z)) + }) + } +} + +#[derive(Copy, Clone, Debug, Serialize, Deserialize)] +struct ValueRange<F : Float> { + ini : F, + min : F, +} + +impl<F : Float> ValueRange<F> { + fn expand_with(self, other : Self) -> Self { + ValueRange { + ini : self.ini.max(other.ini), + min : self.min.min(other.min), + } + } +} + +/// Calculative minimum and maximum values of all the `logs`, and save them into +/// corresponding file names given as the first elements of the tuples in the vectors. +fn save_logs<F : Float + for<'b> Deserialize<'b>, const N : usize>( + logs : Vec<(String, Logger<Timed<IterInfo<F, N>>>)>, + valuerange_file : String, + load_valuerange : bool, +) -> DynError { + // Process logs for relative values + println!("{}", "Processing logs…"); + + // Find minimum value and initial value within a single log + let proc_single_log = |log : &Logger<Timed<IterInfo<F, N>>>| { + let d = log.data(); + let mi = d.iter() + .map(|i| i.data.value) + .reduce(NumTraitsFloat::min); + d.first() + .map(|i| i.data.value) + .zip(mi) + .map(|(ini, min)| ValueRange{ ini, min }) + }; + + // Find minimum and maximum value over all logs + let mut v = logs.iter() + .filter_map(|&(_, ref log)| proc_single_log(log)) + .reduce(|v1, v2| v1.expand_with(v2)) + .ok_or(anyhow!("No algorithms found"))?; + + // Load existing range + if load_valuerange && std::fs::metadata(&valuerange_file).is_ok() { + let data = std::fs::read_to_string(&valuerange_file)?; + v = v.expand_with(serde_json::from_str(&data)?); + } + + let logmap = |Timed { cpu_time, iter, data }| { + let IterInfo { + value, + n_spikes, + inner_iters, + merged, + pruned, + //postprocessing, + this_iters, + .. + } = data; + // let post_value = match (postprocessing, dataterm) { + // (Some(mut μ), DataTerm::L2Squared) => { + // // Comparison postprocessing is only implemented for the case handled + // // by the FW variants. + // reg.optimise_weights( + // &mut μ, &opA, &b, &findim_data, &inner_config, + // inner_it + // ); + // dataterm.value_at_residual(opA.apply(&μ) - &b) + // + regularisation.apply(&μ) + // }, + // _ => value, + // }; + let relative_value = (value - v.min)/(v.ini - v.min); + CSVLog { + iter, + value, + relative_value, + //post_value, + n_spikes, + cpu_time : cpu_time.as_secs_f64(), + inner_iters, + merged, + pruned, + this_iters + } + }; + + println!("{}", "Saving logs …".green()); + + serde_json::to_writer_pretty(std::fs::File::create(&valuerange_file)?, &v)?; + + for (name, logger) in logs { + logger.map(logmap).write_csv(name)?; + } + + Ok(()) +} + /// Plot experiment setup #[replace_float_literals(F::cast_from(literal))] @@ -597,7 +1234,7 @@ sensor : &Sensor, kernel : &Kernel, spread : &Spread, - μ_hat : &DiscreteMeasure<Loc<F, N>, F>, + μ_hat : &RNDM<F, N>, op𝒟 : &𝒟, opA : &A, b_hat : &A::Observable, @@ -608,11 +1245,12 @@ Sensor : RealMapping<F, N> + Support<F, N> + Clone, Spread : RealMapping<F, N> + Support<F, N> + Clone, Kernel : RealMapping<F, N> + Support<F, N>, - Convolution<Sensor, Spread> : RealMapping<F, N> + Support<F, N>, + Convolution<Sensor, Spread> : DifferentiableRealMapping<F, N> + Support<F, N>, 𝒟 : DiscreteMeasureOp<Loc<F, N>, F>, 𝒟::Codomain : RealMapping<F, N>, - A : ForwardModel<Loc<F, N>, F>, - A::PreadjointCodomain : RealMapping<F, N> + Bounded<F>, + A : ForwardModel<RNDM<F, N>, F>, + for<'a> &'a A::Observable : Instance<A::Observable>, + A::PreadjointCodomain : DifferentiableRealMapping<F, N> + Bounded<F>, PlotLookup : Plotting<N>, Cube<F, N> : SetOrd { @@ -623,79 +1261,36 @@ let base = Convolution(sensor.clone(), spread.clone()); let resolution = if N==1 { 100 } else { 40 }; - let pfx = |n| format!("{}{}", prefix, n); + let pfx = |n| format!("{prefix}{n}"); let plotgrid = lingrid(&[[-kernel_plot_width, kernel_plot_width]; N].into(), &[resolution; N]); - PlotLookup::plot_into_file(sensor, plotgrid, pfx("sensor"), "sensor".to_string()); - PlotLookup::plot_into_file(kernel, plotgrid, pfx("kernel"), "kernel".to_string()); - PlotLookup::plot_into_file(spread, plotgrid, pfx("spread"), "spread".to_string()); - PlotLookup::plot_into_file(&base, plotgrid, pfx("base_sensor"), "base_sensor".to_string()); + PlotLookup::plot_into_file(sensor, plotgrid, pfx("sensor")); + PlotLookup::plot_into_file(kernel, plotgrid, pfx("kernel")); + PlotLookup::plot_into_file(spread, plotgrid, pfx("spread")); + PlotLookup::plot_into_file(&base, plotgrid, pfx("base_sensor")); let plotgrid2 = lingrid(&domain, &[resolution; N]); let ω_hat = op𝒟.apply(μ_hat); let noise = opA.preadjoint().apply(opA.apply(μ_hat) - b); - PlotLookup::plot_into_file(&ω_hat, plotgrid2, pfx("omega_hat"), "ω̂".to_string()); - PlotLookup::plot_into_file(&noise, plotgrid2, pfx("omega_noise"), - "noise Aᵀ(Aμ̂ - b)".to_string()); + PlotLookup::plot_into_file(&ω_hat, plotgrid2, pfx("omega_hat")); + PlotLookup::plot_into_file(&noise, plotgrid2, pfx("omega_noise")); let preadj_b = opA.preadjoint().apply(b); let preadj_b_hat = opA.preadjoint().apply(b_hat); //let bounds = preadj_b.bounds().common(&preadj_b_hat.bounds()); PlotLookup::plot_into_file_spikes( - "Aᵀb".to_string(), &preadj_b, - "Aᵀb̂".to_string(), Some(&preadj_b_hat), - plotgrid2, None, &μ_hat, + Some(&preadj_b), + Some(&preadj_b_hat), + plotgrid2, + &μ_hat, pfx("omega_b") ); + PlotLookup::plot_into_file(&preadj_b, plotgrid2, pfx("preadj_b")); + PlotLookup::plot_into_file(&preadj_b_hat, plotgrid2, pfx("preadj_b_hat")); // Save true solution and observables - let pfx = |n| format!("{}{}", prefix, n); μ_hat.write_csv(pfx("orig.txt"))?; opA.write_observable(&b_hat, pfx("b_hat"))?; opA.write_observable(&b, pfx("b_noisy")) } - -// -// Deprecated interface -// - -/// Struct for experiment configurations -#[derive(Debug, Clone, Serialize)] -pub struct Experiment<F, NoiseDistr, S, K, P, const N : usize> -where F : Float, - [usize; N] : Serialize, - NoiseDistr : Distribution<F>, - S : Sensor<F, N>, - P : Spread<F, N>, - K : SimpleConvolutionKernel<F, N>, -{ - /// Domain $Ω$. - pub domain : Cube<F, N>, - /// Number of sensors along each dimension - pub sensor_count : [usize; N], - /// Noise distribution - pub noise_distr : NoiseDistr, - /// Seed for random noise generation (for repeatable experiments) - pub noise_seed : u64, - /// Sensor $θ$; $θ * ψ$ forms the forward operator $𝒜$. - pub sensor : S, - /// Spread $ψ$; $θ * ψ$ forms the forward operator $𝒜$. - pub spread : P, - /// Kernel $ρ$ of $𝒟$. - pub kernel : K, - /// True point sources - pub μ_hat : DiscreteMeasure<Loc<F, N>, F>, - /// Regularisation parameter - #[deprecated(note = "Use [`ExperimentV2`], which replaces `α` by more generic `regularisation`")] - pub α : F, - /// For plotting : how wide should the kernels be plotted - pub kernel_plot_width : F, - /// Data term - pub dataterm : DataTerm, - /// A map of default configurations for algorithms - #[serde(skip)] - pub algorithm_defaults : HashMap<DefaultAlgorithm, AlgorithmConfig<F>>, -} - -impl_experiment!(Experiment, α, Regularisation::NonnegRadon);
--- a/src/seminorms.rs Tue Aug 01 10:25:09 2023 +0300 +++ b/src/seminorms.rs Mon Feb 17 13:54:53 2025 -0500 @@ -12,9 +12,11 @@ use alg_tools::bisection_tree::*; use alg_tools::mapping::RealMapping; use alg_tools::iter::{Mappable, FilterMapX}; -use alg_tools::linops::{Apply, Linear, BoundedLinear}; +use alg_tools::linops::{Mapping, Linear, BoundedLinear}; +use alg_tools::instance::Instance; use alg_tools::nalgebra_support::ToNalgebraRealField; -use crate::measures::{DiscreteMeasure, DeltaMeasure, SpikeIter}; +use alg_tools::norms::Linfinity; +use crate::measures::{DiscreteMeasure, DeltaMeasure, SpikeIter, Radon, RNDM}; use nalgebra::DMatrix; use std::marker::PhantomData; use itertools::Itertools; @@ -22,9 +24,12 @@ /// Abstraction for operators $𝒟 ∈ 𝕃(𝒵(Ω); C_c(Ω))$. /// /// Here $𝒵(Ω) ⊂ ℳ(Ω)$ is the space of sums of delta measures, presented by [`DiscreteMeasure`]. -pub trait DiscreteMeasureOp<Domain, F> : BoundedLinear<DiscreteMeasure<Domain, F>, FloatType=F> -where F : Float + ToNalgebraRealField, - Domain : 'static { +pub trait DiscreteMeasureOp<Domain, F> + : BoundedLinear<DiscreteMeasure<Domain, F>, Radon, Linfinity, F> +where + F : Float + ToNalgebraRealField, + Domain : 'static + Clone + PartialEq, +{ /// The output type of [`Self::preapply`]. type PreCodomain; @@ -38,7 +43,7 @@ fn findim_matrix<'a, I>(&self, points : I) -> DMatrix<F::MixedType> where I : ExactSizeIterator<Item=&'a Domain> + Clone; - /// [`Apply::apply`] that typically returns an uninitialised [`PreBTFN`] + /// [`Mapping`] that typically returns an uninitialised [`PreBTFN`] /// instead of a full [`BTFN`]. fn preapply(&self, μ : DiscreteMeasure<Domain, F>) -> Self::PreCodomain; } @@ -73,7 +78,7 @@ pub struct ConvolutionSupportGenerator<F : Float, K, const N : usize> where K : SimpleConvolutionKernel<F, N> { kernel : K, - centres : DiscreteMeasure<Loc<F, N>, F>, + centres : RNDM<F, N>, } impl<F : Float, K, const N : usize> ConvolutionSupportGenerator<F, K, N> @@ -130,9 +135,9 @@ where F : Float + ToNalgebraRealField, BT : BTImpl<F, N, Data=usize>, K : SimpleConvolutionKernel<F, N> { - /// Depth of the [`BT`] bisection tree for the outputs [`Apply::apply`]. + /// Depth of the [`BT`] bisection tree for the outputs [`Mapping::apply`]. depth : BT::Depth, - /// Domain of the [`BT`] bisection tree for the outputs [`Apply::apply`]. + /// Domain of the [`BT`] bisection tree for the outputs [`Mapping::apply`]. domain : Cube<F, N>, /// The convolution kernel kernel : K, @@ -146,7 +151,7 @@ /// Creates a new convolution operator $𝒟$ with `kernel` on `domain`. /// - /// The output of [`Apply::apply`] is a [`BT`] of given `depth`. + /// The output of [`Mapping::apply`] is a [`BT`] of given `depth`. pub fn new(depth : BT::Depth, domain : Cube<F, N>, kernel : K) -> Self { ConvolutionOp { depth : depth, @@ -157,7 +162,7 @@ } /// Returns the support generator for this convolution operator. - fn support_generator(&self, μ : DiscreteMeasure<Loc<F, N>, F>) + fn support_generator(&self, μ : RNDM<F, N>) -> ConvolutionSupportGenerator<F, K, N> { // TODO: can we avoid cloning μ? @@ -173,94 +178,43 @@ } } -impl<F, K, BT, const N : usize> Apply<DiscreteMeasure<Loc<F, N>, F>> +impl<F, K, BT, const N : usize> Mapping<RNDM<F, N>> for ConvolutionOp<F, K, BT, N> -where F : Float + ToNalgebraRealField, - BT : BTImpl<F, N, Data=usize>, - K : SimpleConvolutionKernel<F, N>, - Weighted<Shift<K, F, N>, F> : LocalAnalysis<F, BT::Agg, N> { +where + F : Float + ToNalgebraRealField, + BT : BTImpl<F, N, Data=usize>, + K : SimpleConvolutionKernel<F, N>, + Weighted<Shift<K, F, N>, F> : LocalAnalysis<F, BT::Agg, N> +{ - type Output = BTFN<F, ConvolutionSupportGenerator<F, K, N>, BT, N>; + type Codomain = BTFN<F, ConvolutionSupportGenerator<F, K, N>, BT, N>; - fn apply(&self, μ : DiscreteMeasure<Loc<F, N>, F>) -> Self::Output { - let g = self.support_generator(μ); + fn apply<I>(&self, μ : I) -> Self::Codomain + where I : Instance<RNDM<F, N>> { + let g = self.support_generator(μ.own()); BTFN::construct(self.domain.clone(), self.depth, g) } } -impl<'a, F, K, BT, const N : usize> Apply<&'a DiscreteMeasure<Loc<F, N>, F>> +/// [`ConvolutionOp`]s as linear operators over [`DiscreteMeasure`]s. +impl<F, K, BT, const N : usize> Linear<RNDM<F, N>> +for ConvolutionOp<F, K, BT, N> +where + F : Float + ToNalgebraRealField, + BT : BTImpl<F, N, Data=usize>, + K : SimpleConvolutionKernel<F, N>, + Weighted<Shift<K, F, N>, F> : LocalAnalysis<F, BT::Agg, N> +{ } + +impl<F, K, BT, const N : usize> +BoundedLinear<RNDM<F, N>, Radon, Linfinity, F> for ConvolutionOp<F, K, BT, N> where F : Float + ToNalgebraRealField, BT : BTImpl<F, N, Data=usize>, K : SimpleConvolutionKernel<F, N>, Weighted<Shift<K, F, N>, F> : LocalAnalysis<F, BT::Agg, N> { - type Output = BTFN<F, ConvolutionSupportGenerator<F, K, N>, BT, N>; - - fn apply(&self, μ : &'a DiscreteMeasure<Loc<F, N>, F>) -> Self::Output { - self.apply(μ.clone()) - } -} - -/// [`ConvolutionOp`]s as linear operators over [`DiscreteMeasure`]s. -impl<F, K, BT, const N : usize> Linear<DiscreteMeasure<Loc<F, N>, F>> -for ConvolutionOp<F, K, BT, N> -where F : Float + ToNalgebraRealField, - BT : BTImpl<F, N, Data=usize>, - K : SimpleConvolutionKernel<F, N>, - Weighted<Shift<K, F, N>, F> : LocalAnalysis<F, BT::Agg, N> { - type Codomain = BTFN<F, ConvolutionSupportGenerator<F, K, N>, BT, N>; -} - -impl<F, K, BT, const N : usize> Apply<DeltaMeasure<Loc<F, N>, F>> -for ConvolutionOp<F, K, BT, N> -where F : Float + ToNalgebraRealField, - BT : BTImpl<F, N, Data=usize>, - K : SimpleConvolutionKernel<F, N> { - - type Output = Weighted<Shift<K, F, N>, F>; - - #[inline] - fn apply(&self, δ : DeltaMeasure<Loc<F, N>, F>) -> Self::Output { - self.kernel.clone().shift(δ.x).weigh(δ.α) - } -} - -impl<'a, F, K, BT, const N : usize> Apply<&'a DeltaMeasure<Loc<F, N>, F>> -for ConvolutionOp<F, K, BT, N> -where F : Float + ToNalgebraRealField, - BT : BTImpl<F, N, Data=usize>, - K : SimpleConvolutionKernel<F, N> { - - type Output = Weighted<Shift<K, F, N>, F>; - - #[inline] - fn apply(&self, δ : &'a DeltaMeasure<Loc<F, N>, F>) -> Self::Output { - self.kernel.clone().shift(δ.x).weigh(δ.α) - } -} - -/// [`ConvolutionOp`]s as linear operators over [`DeltaMeasure`]s. -/// -/// The codomain is different from the implementation for [`DiscreteMeasure`]. -impl<F, K, BT, const N : usize> Linear<DeltaMeasure<Loc<F, N>, F>> -for ConvolutionOp<F, K, BT, N> -where F : Float + ToNalgebraRealField, - BT : BTImpl<F, N, Data=usize>, - K : SimpleConvolutionKernel<F, N> { - type Codomain = Weighted<Shift<K, F, N>, F>; -} - -impl<F, K, BT, const N : usize> BoundedLinear<DiscreteMeasure<Loc<F, N>, F>> -for ConvolutionOp<F, K, BT, N> -where F : Float + ToNalgebraRealField, - BT : BTImpl<F, N, Data=usize>, - K : SimpleConvolutionKernel<F, N>, - Weighted<Shift<K, F, N>, F> : LocalAnalysis<F, BT::Agg, N> { - - type FloatType = F; - - fn opnorm_bound(&self) -> F { + fn opnorm_bound(&self, _ : Radon, _ : Linfinity) -> F { // With μ = ∑_i α_i δ_{x_i}, we have // |𝒟μ|_∞ // = sup_z |∑_i α_i φ(z - x_i)| @@ -292,10 +246,10 @@ DMatrix::from_iterator(n, n, values) } - /// A version of [`Apply::apply`] that does not instantiate the [`BTFN`] codomain with + /// A version of [`Mapping::apply`] that does not instantiate the [`BTFN`] codomain with /// a bisection tree, instead returning a [`PreBTFN`]. This can improve performance when /// the output is to be added as the right-hand-side operand to a proper BTFN. - fn preapply(&self, μ : DiscreteMeasure<Loc<F, N>, F>) -> Self::PreCodomain { + fn preapply(&self, μ : RNDM<F, N>) -> Self::PreCodomain { BTFN::new_pre(self.support_generator(μ)) } } @@ -368,11 +322,3 @@ make_convolutionsupportgenerator_unaryop!(Neg, neg); -/// Trait for indicating that `Self` is Lipschitz with respect to the seminorm `D`. -pub trait Lipschitz<D> { - /// The type of floats - type FloatType : Float; - - /// Returns the Lipschitz factor of `self` with respect to the seminorm `D`. - fn lipschitz_factor(&self, seminorm : &D) -> Option<Self::FloatType>; -}
--- /dev/null Thu Jan 01 00:00:00 1970 +0000 +++ b/src/sliding_fb.rs Mon Feb 17 13:54:53 2025 -0500 @@ -0,0 +1,444 @@ +/*! +Solver for the point source localisation problem using a sliding +forward-backward splitting method. +*/ + +use numeric_literals::replace_float_literals; +use serde::{Deserialize, Serialize}; +//use colored::Colorize; +//use nalgebra::{DVector, DMatrix}; +use itertools::izip; +use std::iter::Iterator; + +use alg_tools::euclidean::Euclidean; +use alg_tools::iterate::AlgIteratorFactory; +use alg_tools::mapping::{DifferentiableRealMapping, Instance, Mapping}; +use alg_tools::nalgebra_support::ToNalgebraRealField; +use alg_tools::norms::Norm; + +use crate::forward_model::{AdjointProductBoundedBy, BoundedCurvature, ForwardModel}; +use crate::measures::merging::SpikeMerging; +use crate::measures::{DiscreteMeasure, Radon, RNDM}; +use crate::types::*; +//use crate::tolerance::Tolerance; +use crate::dataterm::{calculate_residual, calculate_residual2, DataTerm, L2Squared}; +use crate::fb::*; +use crate::plot::{PlotLookup, Plotting, SeqPlotter}; +use crate::regularisation::SlidingRegTerm; +//use crate::transport::TransportLipschitz; + +/// Transport settings for [`pointsource_sliding_fb_reg`]. +#[derive(Clone, Copy, Eq, PartialEq, Serialize, Deserialize, Debug)] +#[serde(default)] +pub struct TransportConfig<F: Float> { + /// Transport step length $θ$ normalised to $(0, 1)$. + pub θ0: F, + /// Factor in $(0, 1)$ for decreasing transport to adapt to tolerance. + pub adaptation: F, + /// A posteriori transport tolerance multiplier (C_pos) + pub tolerance_mult_con: F, +} + +#[replace_float_literals(F::cast_from(literal))] +impl<F: Float> TransportConfig<F> { + /// Check that the parameters are ok. Panics if not. + pub fn check(&self) { + assert!(self.θ0 > 0.0); + assert!(0.0 < self.adaptation && self.adaptation < 1.0); + assert!(self.tolerance_mult_con > 0.0); + } +} + +#[replace_float_literals(F::cast_from(literal))] +impl<F: Float> Default for TransportConfig<F> { + fn default() -> Self { + TransportConfig { + θ0: 0.9, + adaptation: 0.9, + tolerance_mult_con: 100.0, + } + } +} + +/// Settings for [`pointsource_sliding_fb_reg`]. +#[derive(Clone, Copy, Eq, PartialEq, Serialize, Deserialize, Debug)] +#[serde(default)] +pub struct SlidingFBConfig<F: Float> { + /// Step length scaling + pub τ0: F, + /// Transport parameters + pub transport: TransportConfig<F>, + /// Generic parameters + pub insertion: FBGenericConfig<F>, +} + +#[replace_float_literals(F::cast_from(literal))] +impl<F: Float> Default for SlidingFBConfig<F> { + fn default() -> Self { + SlidingFBConfig { + τ0: 0.99, + transport: Default::default(), + insertion: Default::default(), + } + } +} + +/// Internal type of adaptive transport step length calculation +pub(crate) enum TransportStepLength<F: Float, G: Fn(F, F) -> F> { + /// Fixed, known step length + #[allow(dead_code)] + Fixed(F), + /// Adaptive step length, only wrt. maximum transport. + /// Content of `l` depends on use case, while `g` calculates the step length from `l`. + AdaptiveMax { l: F, max_transport: F, g: G }, + /// Adaptive step length. + /// Content of `l` depends on use case, while `g` calculates the step length from `l`. + FullyAdaptive { l: F, max_transport: F, g: G }, +} + +/// Constrution of initial transport `γ1` from initial measure `μ` and `v=F'(μ)` +/// with step lengh τ and transport step length `θ_or_adaptive`. +#[replace_float_literals(F::cast_from(literal))] +pub(crate) fn initial_transport<F, G, D, const N: usize>( + γ1: &mut RNDM<F, N>, + μ: &mut RNDM<F, N>, + τ: F, + θ_or_adaptive: &mut TransportStepLength<F, G>, + v: D, +) -> (Vec<F>, RNDM<F, N>) +where + F: Float + ToNalgebraRealField, + G: Fn(F, F) -> F, + D: DifferentiableRealMapping<F, N>, +{ + use TransportStepLength::*; + + // Save current base point and shift μ to new positions. Idea is that + // μ_base(_masses) = μ^k (vector of masses) + // μ_base_minus_γ0 = μ^k - π_♯^0γ^{k+1} + // γ1 = π_♯^1γ^{k+1} + // μ = μ^{k+1} + let μ_base_masses: Vec<F> = μ.iter_masses().collect(); + let mut μ_base_minus_γ0 = μ.clone(); // Weights will be set in the loop below. + // Construct μ^{k+1} and π_♯^1γ^{k+1} initial candidates + //let mut sum_norm_dv = 0.0; + let γ_prev_len = γ1.len(); + assert!(μ.len() >= γ_prev_len); + γ1.extend(μ[γ_prev_len..].iter().cloned()); + + // Calculate initial transport and step length. + // First calculate initial transported weights + for (δ, ρ) in izip!(μ.iter_spikes(), γ1.iter_spikes_mut()) { + // If old transport has opposing sign, the new transport will be none. + ρ.α = if (ρ.α > 0.0 && δ.α < 0.0) || (ρ.α < 0.0 && δ.α > 0.0) { + 0.0 + } else { + δ.α + }; + } + + // Calculate transport rays. + match *θ_or_adaptive { + Fixed(θ) => { + let θτ = τ * θ; + for (δ, ρ) in izip!(μ.iter_spikes(), γ1.iter_spikes_mut()) { + ρ.x = δ.x - v.differential(&δ.x) * (ρ.α.signum() * θτ); + } + } + AdaptiveMax { + l: ℓ_F, + ref mut max_transport, + g: ref calculate_θ, + } => { + *max_transport = max_transport.max(γ1.norm(Radon)); + let θτ = τ * calculate_θ(ℓ_F, *max_transport); + for (δ, ρ) in izip!(μ.iter_spikes(), γ1.iter_spikes_mut()) { + ρ.x = δ.x - v.differential(&δ.x) * (ρ.α.signum() * θτ); + } + } + FullyAdaptive { + l: ref mut adaptive_ℓ_F, + ref mut max_transport, + g: ref calculate_θ, + } => { + *max_transport = max_transport.max(γ1.norm(Radon)); + let mut θ = calculate_θ(*adaptive_ℓ_F, *max_transport); + // Do two runs through the spikes to update θ, breaking if first run did not cause + // a change. + for _i in 0..=1 { + let mut changes = false; + for (δ, ρ) in izip!(μ.iter_spikes(), γ1.iter_spikes_mut()) { + let dv_x = v.differential(&δ.x); + let g = &dv_x * (ρ.α.signum() * θ * τ); + ρ.x = δ.x - g; + let n = g.norm2(); + if n >= F::EPSILON { + // Estimate Lipschitz factor of ∇v + let this_ℓ_F = (dv_x - v.differential(&ρ.x)).norm2() / n; + *adaptive_ℓ_F = adaptive_ℓ_F.max(this_ℓ_F); + θ = calculate_θ(*adaptive_ℓ_F, *max_transport); + changes = true + } + } + if !changes { + break; + } + } + } + } + + // Set initial guess for μ=μ^{k+1}. + for (δ, ρ, &β) in izip!(μ.iter_spikes_mut(), γ1.iter_spikes(), μ_base_masses.iter()) { + if ρ.α.abs() > F::EPSILON { + δ.x = ρ.x; + //δ.α = ρ.α; // already set above + } else { + δ.α = β; + } + } + // Calculate μ^k-π_♯^0γ^{k+1} and v̆ = A_*(A[μ_transported + μ_transported_base]-b) + μ_base_minus_γ0.set_masses( + μ_base_masses + .iter() + .zip(γ1.iter_masses()) + .map(|(&a, b)| a - b), + ); + (μ_base_masses, μ_base_minus_γ0) +} + +/// A posteriori transport adaptation. +#[replace_float_literals(F::cast_from(literal))] +pub(crate) fn aposteriori_transport<F, const N: usize>( + γ1: &mut RNDM<F, N>, + μ: &mut RNDM<F, N>, + μ_base_minus_γ0: &mut RNDM<F, N>, + μ_base_masses: &Vec<F>, + extra: Option<F>, + ε: F, + tconfig: &TransportConfig<F>, +) -> bool +where + F: Float + ToNalgebraRealField, +{ + // 1. If π_♯^1γ^{k+1} = γ1 has non-zero mass at some point y, but μ = μ^{k+1} does not, + // then the ansatz ∇w̃_x(y) = w^{k+1}(y) may not be satisfied. So set the mass of γ1 + // at that point to zero, and retry. + let mut all_ok = true; + for (α_μ, α_γ1) in izip!(μ.iter_masses(), γ1.iter_masses_mut()) { + if α_μ == 0.0 && *α_γ1 != 0.0 { + all_ok = false; + *α_γ1 = 0.0; + } + } + + // 2. Through bounding ∫ B_ω(y, z) dλ(x, y, z). + // through the estimate ≤ C ‖Δ‖‖γ^{k+1}‖ for Δ := μ^{k+1}-μ^k-(π_♯^1-π_♯^0)γ^{k+1}, + // which holds for some some C if the convolution kernel in 𝒟 has Lipschitz gradient. + let nγ = γ1.norm(Radon); + let nΔ = μ_base_minus_γ0.norm(Radon) + μ.dist_matching(&γ1) + extra.unwrap_or(0.0); + let t = ε * tconfig.tolerance_mult_con; + if nγ * nΔ > t { + // Since t/(nγ*nΔ)<1, and the constant tconfig.adaptation < 1, + // this will guarantee that eventually ‖γ‖ decreases sufficiently that we + // will not enter here. + *γ1 *= tconfig.adaptation * t / (nγ * nΔ); + all_ok = false + } + + if !all_ok { + // Update weights for μ_base_minus_γ0 = μ^k - π_♯^0γ^{k+1} + μ_base_minus_γ0.set_masses( + μ_base_masses + .iter() + .zip(γ1.iter_masses()) + .map(|(&a, b)| a - b), + ); + } + + all_ok +} + +/// Iteratively solve the pointsource localisation problem using sliding forward-backward +/// splitting +/// +/// The parametrisation is as for [`pointsource_fb_reg`]. +/// Inertia is currently not supported. +#[replace_float_literals(F::cast_from(literal))] +pub fn pointsource_sliding_fb_reg<F, I, A, Reg, P, const N: usize>( + opA: &A, + b: &A::Observable, + reg: Reg, + prox_penalty: &P, + config: &SlidingFBConfig<F>, + iterator: I, + mut plotter: SeqPlotter<F, N>, +) -> RNDM<F, N> +where + F: Float + ToNalgebraRealField, + I: AlgIteratorFactory<IterInfo<F, N>>, + A: ForwardModel<RNDM<F, N>, F> + + AdjointProductBoundedBy<RNDM<F, N>, P, FloatType = F> + + BoundedCurvature<FloatType = F>, + for<'b> &'b A::Observable: std::ops::Neg<Output = A::Observable> + Instance<A::Observable>, + A::PreadjointCodomain: DifferentiableRealMapping<F, N>, + RNDM<F, N>: SpikeMerging<F>, + Reg: SlidingRegTerm<F, N>, + P: ProxPenalty<F, A::PreadjointCodomain, Reg, N>, + PlotLookup: Plotting<N>, +{ + // Check parameters + assert!(config.τ0 > 0.0, "Invalid step length parameter"); + config.transport.check(); + + // Initialise iterates + let mut μ = DiscreteMeasure::new(); + let mut γ1 = DiscreteMeasure::new(); + let mut residual = -b; // Has to equal $Aμ-b$. + + // Set up parameters + // let opAnorm = opA.opnorm_bound(Radon, L2); + //let max_transport = config.max_transport.scale + // * reg.radon_norm_bound(b.norm2_squared() / 2.0); + //let ℓ = opA.transport.lipschitz_factor(L2Squared) * max_transport; + let ℓ = 0.0; + let τ = config.τ0 / opA.adjoint_product_bound(prox_penalty).unwrap(); + let (maybe_ℓ_F0, maybe_transport_lip) = opA.curvature_bound_components(); + let transport_lip = maybe_transport_lip.unwrap(); + let calculate_θ = |ℓ_F, max_transport| { + let ℓ_r = transport_lip * max_transport; + config.transport.θ0 / (τ * (ℓ + ℓ_F + ℓ_r)) + }; + let mut θ_or_adaptive = match maybe_ℓ_F0 { + //Some(ℓ_F0) => TransportStepLength::Fixed(calculate_θ(ℓ_F0 * b.norm2(), 0.0)), + Some(ℓ_F0) => TransportStepLength::AdaptiveMax { + l: ℓ_F0 * b.norm2(), // TODO: could estimate computing the real reesidual + max_transport: 0.0, + g: calculate_θ, + }, + None => TransportStepLength::FullyAdaptive { + l: 10.0 * F::EPSILON, // Start with something very small to estimate differentials + max_transport: 0.0, + g: calculate_θ, + }, + }; + // We multiply tolerance by τ for FB since our subproblems depending on tolerances are scaled + // by τ compared to the conditional gradient approach. + let tolerance = config.insertion.tolerance * τ * reg.tolerance_scaling(); + let mut ε = tolerance.initial(); + + // Statistics + let full_stats = |residual: &A::Observable, μ: &RNDM<F, N>, ε, stats| IterInfo { + value: residual.norm2_squared_div2() + reg.apply(μ), + n_spikes: μ.len(), + ε, + // postprocessing: config.insertion.postprocessing.then(|| μ.clone()), + ..stats + }; + let mut stats = IterInfo::new(); + + // Run the algorithm + for state in iterator.iter_init(|| full_stats(&residual, &μ, ε, stats.clone())) { + // Calculate initial transport + let v = opA.preadjoint().apply(residual); + let (μ_base_masses, mut μ_base_minus_γ0) = + initial_transport(&mut γ1, &mut μ, τ, &mut θ_or_adaptive, v); + + // Solve finite-dimensional subproblem several times until the dual variable for the + // regularisation term conforms to the assumptions made for the transport above. + let (maybe_d, _within_tolerances, mut τv̆) = 'adapt_transport: loop { + // Calculate τv̆ = τA_*(A[μ_transported + μ_transported_base]-b) + let residual_μ̆ = calculate_residual2(&γ1, &μ_base_minus_γ0, opA, b); + let mut τv̆ = opA.preadjoint().apply(residual_μ̆ * τ); + + // Construct μ^{k+1} by solving finite-dimensional subproblems and insert new spikes. + let (maybe_d, within_tolerances) = prox_penalty.insert_and_reweigh( + &mut μ, + &mut τv̆, + &γ1, + Some(&μ_base_minus_γ0), + τ, + ε, + &config.insertion, + ®, + &state, + &mut stats, + ); + + // A posteriori transport adaptation. + if aposteriori_transport( + &mut γ1, + &mut μ, + &mut μ_base_minus_γ0, + &μ_base_masses, + None, + ε, + &config.transport, + ) { + break 'adapt_transport (maybe_d, within_tolerances, τv̆); + } + }; + + stats.untransported_fraction = Some({ + assert_eq!(μ_base_masses.len(), γ1.len()); + let (a, b) = stats.untransported_fraction.unwrap_or((0.0, 0.0)); + let source = μ_base_masses.iter().map(|v| v.abs()).sum(); + (a + μ_base_minus_γ0.norm(Radon), b + source) + }); + stats.transport_error = Some({ + assert_eq!(μ_base_masses.len(), γ1.len()); + let (a, b) = stats.transport_error.unwrap_or((0.0, 0.0)); + (a + μ.dist_matching(&γ1), b + γ1.norm(Radon)) + }); + + // Merge spikes. + // This crucially expects the merge routine to be stable with respect to spike locations, + // and not to performing any pruning. That is be to done below simultaneously for γ. + let ins = &config.insertion; + if ins.merge_now(&state) { + stats.merged += prox_penalty.merge_spikes( + &mut μ, + &mut τv̆, + &γ1, + Some(&μ_base_minus_γ0), + τ, + ε, + ins, + ®, + Some(|μ̃: &RNDM<F, N>| L2Squared.calculate_fit_op(μ̃, opA, b)), + ); + } + + // Prune spikes with zero weight. To maintain correct ordering between μ and γ1, also the + // latter needs to be pruned when μ is. + // TODO: This could do with a two-vector Vec::retain to avoid copies. + let μ_new = DiscreteMeasure::from_iter(μ.iter_spikes().filter(|δ| δ.α != F::ZERO).cloned()); + if μ_new.len() != μ.len() { + let mut μ_iter = μ.iter_spikes(); + γ1.prune_by(|_| μ_iter.next().unwrap().α != F::ZERO); + stats.pruned += μ.len() - μ_new.len(); + μ = μ_new; + } + + // Update residual + residual = calculate_residual(&μ, opA, b); + + let iter = state.iteration(); + stats.this_iters += 1; + + // Give statistics if requested + state.if_verbose(|| { + plotter.plot_spikes(iter, maybe_d.as_ref(), Some(&τv̆), &μ); + full_stats( + &residual, + &μ, + ε, + std::mem::replace(&mut stats, IterInfo::new()), + ) + }); + + // Update main tolerance for next iteration + ε = tolerance.update(ε, iter); + } + + postprocess(μ, &config.insertion, L2Squared, opA, b) +}
--- /dev/null Thu Jan 01 00:00:00 1970 +0000 +++ b/src/sliding_pdps.rs Mon Feb 17 13:54:53 2025 -0500 @@ -0,0 +1,373 @@ +/*! +Solver for the point source localisation problem using a sliding +primal-dual proximal splitting method. +*/ + +use numeric_literals::replace_float_literals; +use serde::{Deserialize, Serialize}; +//use colored::Colorize; +//use nalgebra::{DVector, DMatrix}; +use std::iter::Iterator; + +use alg_tools::convex::{Conjugable, Prox}; +use alg_tools::direct_product::Pair; +use alg_tools::euclidean::Euclidean; +use alg_tools::iterate::AlgIteratorFactory; +use alg_tools::linops::{Adjointable, BoundedLinear, IdOp, AXPY, GEMV}; +use alg_tools::mapping::{DifferentiableRealMapping, Instance, Mapping}; +use alg_tools::nalgebra_support::ToNalgebraRealField; +use alg_tools::norms::{Dist, Norm}; +use alg_tools::norms::{PairNorm, L2}; + +use crate::forward_model::{AdjointProductPairBoundedBy, BoundedCurvature, ForwardModel}; +use crate::measures::merging::SpikeMerging; +use crate::measures::{DiscreteMeasure, Radon, RNDM}; +use crate::types::*; +// use crate::transport::TransportLipschitz; +//use crate::tolerance::Tolerance; +use crate::fb::*; +use crate::plot::{PlotLookup, Plotting, SeqPlotter}; +use crate::regularisation::SlidingRegTerm; +// use crate::dataterm::L2Squared; +use crate::dataterm::{calculate_residual, calculate_residual2}; +use crate::sliding_fb::{ + aposteriori_transport, initial_transport, TransportConfig, TransportStepLength, +}; + +/// Settings for [`pointsource_sliding_pdps_pair`]. +#[derive(Clone, Copy, Eq, PartialEq, Serialize, Deserialize, Debug)] +#[serde(default)] +pub struct SlidingPDPSConfig<F: Float> { + /// Primal step length scaling. + pub τ0: F, + /// Primal step length scaling. + pub σp0: F, + /// Dual step length scaling. + pub σd0: F, + /// Transport parameters + pub transport: TransportConfig<F>, + /// Generic parameters + pub insertion: FBGenericConfig<F>, +} + +#[replace_float_literals(F::cast_from(literal))] +impl<F: Float> Default for SlidingPDPSConfig<F> { + fn default() -> Self { + SlidingPDPSConfig { + τ0: 0.99, + σd0: 0.05, + σp0: 0.99, + transport: TransportConfig { + θ0: 0.9, + ..Default::default() + }, + insertion: Default::default(), + } + } +} + +type MeasureZ<F, Z, const N: usize> = Pair<RNDM<F, N>, Z>; + +/// Iteratively solve the pointsource localisation with an additional variable +/// using sliding primal-dual proximal splitting +/// +/// The parametrisation is as for [`crate::forward_pdps::pointsource_forward_pdps_pair`]. +#[replace_float_literals(F::cast_from(literal))] +pub fn pointsource_sliding_pdps_pair< + F, + I, + A, + S, + Reg, + P, + Z, + R, + Y, + /*KOpM, */ KOpZ, + H, + const N: usize, +>( + opA: &A, + b: &A::Observable, + reg: Reg, + prox_penalty: &P, + config: &SlidingPDPSConfig<F>, + iterator: I, + mut plotter: SeqPlotter<F, N>, + //opKμ : KOpM, + opKz: &KOpZ, + fnR: &R, + fnH: &H, + mut z: Z, + mut y: Y, +) -> MeasureZ<F, Z, N> +where + F: Float + ToNalgebraRealField, + I: AlgIteratorFactory<IterInfo<F, N>>, + A: ForwardModel<MeasureZ<F, Z, N>, F, PairNorm<Radon, L2, L2>, PreadjointCodomain = Pair<S, Z>> + + AdjointProductPairBoundedBy<MeasureZ<F, Z, N>, P, IdOp<Z>, FloatType = F> + + BoundedCurvature<FloatType = F>, + S: DifferentiableRealMapping<F, N>, + for<'b> &'b A::Observable: std::ops::Neg<Output = A::Observable> + Instance<A::Observable>, + PlotLookup: Plotting<N>, + RNDM<F, N>: SpikeMerging<F>, + Reg: SlidingRegTerm<F, N>, + P: ProxPenalty<F, S, Reg, N>, + // KOpM : Linear<RNDM<F, N>, Codomain=Y> + // + GEMV<F, RNDM<F, N>> + // + Preadjointable< + // RNDM<F, N>, Y, + // PreadjointCodomain = S, + // > + // + TransportLipschitz<L2Squared, FloatType=F> + // + AdjointProductBoundedBy<RNDM<F, N>, 𝒟, FloatType=F>, + // for<'b> KOpM::Preadjoint<'b> : GEMV<F, Y>, + // Since Z is Hilbert, we may just as well use adjoints for K_z. + KOpZ: BoundedLinear<Z, L2, L2, F, Codomain = Y> + + GEMV<F, Z> + + Adjointable<Z, Y, AdjointCodomain = Z>, + for<'b> KOpZ::Adjoint<'b>: GEMV<F, Y>, + Y: AXPY<F> + Euclidean<F, Output = Y> + Clone + ClosedAdd, + for<'b> &'b Y: Instance<Y>, + Z: AXPY<F, Owned = Z> + Euclidean<F, Output = Z> + Clone + Norm<F, L2> + Dist<F, L2>, + for<'b> &'b Z: Instance<Z>, + R: Prox<Z, Codomain = F>, + H: Conjugable<Y, F, Codomain = F>, + for<'b> H::Conjugate<'b>: Prox<Y>, +{ + // Check parameters + assert!( + config.τ0 > 0.0 + && config.τ0 < 1.0 + && config.σp0 > 0.0 + && config.σp0 < 1.0 + && config.σd0 > 0.0 + && config.σp0 * config.σd0 <= 1.0, + "Invalid step length parameters" + ); + config.transport.check(); + + // Initialise iterates + let mut μ = DiscreteMeasure::new(); + let mut γ1 = DiscreteMeasure::new(); + let mut residual = calculate_residual(Pair(&μ, &z), opA, b); + let zero_z = z.similar_origin(); + + // Set up parameters + // TODO: maybe this PairNorm doesn't make sense here? + // let opAnorm = opA.opnorm_bound(PairNorm(Radon, L2, L2), L2); + let bigθ = 0.0; //opKμ.transport_lipschitz_factor(L2Squared); + let bigM = 0.0; //opKμ.adjoint_product_bound(&op𝒟).unwrap().sqrt(); + let nKz = opKz.opnorm_bound(L2, L2); + let ℓ = 0.0; + let opIdZ = IdOp::new(); + let (l, l_z) = opA + .adjoint_product_pair_bound(prox_penalty, &opIdZ) + .unwrap(); + // We need to satisfy + // + // τσ_dM(1-σ_p L_z)/(1 - τ L) + [σ_p L_z + σ_pσ_d‖K_z‖^2] < 1 + // ^^^^^^^^^^^^^^^^^^^^^^^^^ + // with 1 > σ_p L_z and 1 > τ L. + // + // To do so, we first solve σ_p and σ_d from standard PDPS step length condition + // ^^^^^ < 1. then we solve τ from the rest. + let σ_d = config.σd0 / nKz; + let σ_p = config.σp0 / (l_z + config.σd0 * nKz); + // Observe that = 1 - ^^^^^^^^^^^^^^^^^^^^^ = 1 - σ_{p,0} + // We get the condition τσ_d M (1-σ_p L_z) < (1-σ_{p,0})*(1-τ L) + // ⟺ τ [ σ_d M (1-σ_p L_z) + (1-σ_{p,0}) L ] < (1-σ_{p,0}) + let φ = 1.0 - config.σp0; + let a = 1.0 - σ_p * l_z; + let τ = config.τ0 * φ / (σ_d * bigM * a + φ * l); + let ψ = 1.0 - τ * l; + let β = σ_p * config.σd0 * nKz / a; // σ_p * σ_d * (nKz * nK_z) / a; + assert!(β < 1.0); + // Now we need κ‖K_μ(π_♯^1 - π_♯^0)γ‖^2 ≤ (1/θ - τ[ℓ_F + ℓ]) ∫ c_2 dγ for κ defined as: + let κ = τ * σ_d * ψ / ((1.0 - β) * ψ - τ * σ_d * bigM); + // The factor two in the manuscript disappears due to the definition of 𝚹 being + // for ‖x-y‖₂² instead of c_2(x, y)=‖x-y‖₂²/2. + let (maybe_ℓ_F0, maybe_transport_lip) = opA.curvature_bound_components(); + let transport_lip = maybe_transport_lip.unwrap(); + let calculate_θ = |ℓ_F, max_transport| { + let ℓ_r = transport_lip * max_transport; + config.transport.θ0 / (τ * (ℓ + ℓ_F + ℓ_r) + κ * bigθ * max_transport) + }; + let mut θ_or_adaptive = match maybe_ℓ_F0 { + // We assume that the residual is decreasing. + Some(ℓ_F0) => TransportStepLength::AdaptiveMax { + l: ℓ_F0 * b.norm2(), // TODO: could estimate computing the real reesidual + max_transport: 0.0, + g: calculate_θ, + }, + None => TransportStepLength::FullyAdaptive { + l: F::EPSILON, + max_transport: 0.0, + g: calculate_θ, + }, + }; + // Acceleration is not currently supported + // let γ = dataterm.factor_of_strong_convexity(); + let ω = 1.0; + + // We multiply tolerance by τ for FB since our subproblems depending on tolerances are scaled + // by τ compared to the conditional gradient approach. + let tolerance = config.insertion.tolerance * τ * reg.tolerance_scaling(); + let mut ε = tolerance.initial(); + + let starH = fnH.conjugate(); + + // Statistics + let full_stats = |residual: &A::Observable, μ: &RNDM<F, N>, z: &Z, ε, stats| IterInfo { + value: residual.norm2_squared_div2() + + fnR.apply(z) + + reg.apply(μ) + + fnH.apply(/* opKμ.apply(μ) + */ opKz.apply(z)), + n_spikes: μ.len(), + ε, + // postprocessing: config.insertion.postprocessing.then(|| μ.clone()), + ..stats + }; + let mut stats = IterInfo::new(); + + // Run the algorithm + for state in iterator.iter_init(|| full_stats(&residual, &μ, &z, ε, stats.clone())) { + // Calculate initial transport + let Pair(v, _) = opA.preadjoint().apply(&residual); + //opKμ.preadjoint().apply_add(&mut v, y); + // We want to proceed as in Example 4.12 but with v and v̆ as in §5. + // With A(ν, z) = A_μ ν + A_z z, following Example 5.1, we have + // P_ℳ[F'(ν, z) + Ξ(ν, z, y)]= A_ν^*[A_ν ν + A_z z] + K_μ ν = A_ν^*A(ν, z) + K_μ ν, + // where A_ν^* becomes a multiplier. + // This is much easier with K_μ = 0, which is the only reason why are enforcing it. + // TODO: Write a version of initial_transport that can deal with K_μ ≠ 0. + + let (μ_base_masses, mut μ_base_minus_γ0) = + initial_transport(&mut γ1, &mut μ, τ, &mut θ_or_adaptive, v); + + // Solve finite-dimensional subproblem several times until the dual variable for the + // regularisation term conforms to the assumptions made for the transport above. + let (maybe_d, _within_tolerances, mut τv̆, z_new) = 'adapt_transport: loop { + // Calculate τv̆ = τA_*(A[μ_transported + μ_transported_base]-b) + let residual_μ̆ = + calculate_residual2(Pair(&γ1, &z), Pair(&μ_base_minus_γ0, &zero_z), opA, b); + let Pair(mut τv̆, τz̆) = opA.preadjoint().apply(residual_μ̆ * τ); + // opKμ.preadjoint().gemv(&mut τv̆, τ, y, 1.0); + + // Construct μ^{k+1} by solving finite-dimensional subproblems and insert new spikes. + let (maybe_d, within_tolerances) = prox_penalty.insert_and_reweigh( + &mut μ, + &mut τv̆, + &γ1, + Some(&μ_base_minus_γ0), + τ, + ε, + &config.insertion, + ®, + &state, + &mut stats, + ); + + // Do z variable primal update here to able to estimate B_{v̆^k-v^{k+1}} + let mut z_new = τz̆; + opKz.adjoint().gemv(&mut z_new, -σ_p, &y, -σ_p / τ); + z_new = fnR.prox(σ_p, z_new + &z); + + // A posteriori transport adaptation. + if aposteriori_transport( + &mut γ1, + &mut μ, + &mut μ_base_minus_γ0, + &μ_base_masses, + Some(z_new.dist(&z, L2)), + ε, + &config.transport, + ) { + break 'adapt_transport (maybe_d, within_tolerances, τv̆, z_new); + } + }; + + stats.untransported_fraction = Some({ + assert_eq!(μ_base_masses.len(), γ1.len()); + let (a, b) = stats.untransported_fraction.unwrap_or((0.0, 0.0)); + let source = μ_base_masses.iter().map(|v| v.abs()).sum(); + (a + μ_base_minus_γ0.norm(Radon), b + source) + }); + stats.transport_error = Some({ + assert_eq!(μ_base_masses.len(), γ1.len()); + let (a, b) = stats.transport_error.unwrap_or((0.0, 0.0)); + (a + μ.dist_matching(&γ1), b + γ1.norm(Radon)) + }); + + // Merge spikes. + // This crucially expects the merge routine to be stable with respect to spike locations, + // and not to performing any pruning. That is be to done below simultaneously for γ. + let ins = &config.insertion; + if ins.merge_now(&state) { + stats.merged += prox_penalty.merge_spikes_no_fitness( + &mut μ, + &mut τv̆, + &γ1, + Some(&μ_base_minus_γ0), + τ, + ε, + ins, + ®, + //Some(|μ̃ : &RNDM<F, N>| calculate_residual(Pair(μ̃, &z), opA, b).norm2_squared_div2()), + ); + } + + // Prune spikes with zero weight. To maintain correct ordering between μ and γ1, also the + // latter needs to be pruned when μ is. + // TODO: This could do with a two-vector Vec::retain to avoid copies. + let μ_new = DiscreteMeasure::from_iter(μ.iter_spikes().filter(|δ| δ.α != F::ZERO).cloned()); + if μ_new.len() != μ.len() { + let mut μ_iter = μ.iter_spikes(); + γ1.prune_by(|_| μ_iter.next().unwrap().α != F::ZERO); + stats.pruned += μ.len() - μ_new.len(); + μ = μ_new; + } + + // Do dual update + // opKμ.gemv(&mut y, σ_d*(1.0 + ω), &μ, 1.0); // y = y + σ_d K[(1+ω)(μ,z)^{k+1}] + opKz.gemv(&mut y, σ_d * (1.0 + ω), &z_new, 1.0); + // opKμ.gemv(&mut y, -σ_d*ω, μ_base, 1.0);// y = y + σ_d K[(1+ω)(μ,z)^{k+1} - ω (μ,z)^k]-b + opKz.gemv(&mut y, -σ_d * ω, z, 1.0); // y = y + σ_d K[(1+ω)(μ,z)^{k+1} - ω (μ,z)^k]-b + y = starH.prox(σ_d, y); + z = z_new; + + // Update residual + residual = calculate_residual(Pair(&μ, &z), opA, b); + + // Update step length parameters + // let ω = pdpsconfig.acceleration.accelerate(&mut τ, &mut σ, γ); + + // Give statistics if requested + let iter = state.iteration(); + stats.this_iters += 1; + + state.if_verbose(|| { + plotter.plot_spikes(iter, maybe_d.as_ref(), Some(&τv̆), &μ); + full_stats( + &residual, + &μ, + &z, + ε, + std::mem::replace(&mut stats, IterInfo::new()), + ) + }); + + // Update main tolerance for next iteration + ε = tolerance.update(ε, iter); + } + + let fit = |μ̃: &RNDM<F, N>| { + (opA.apply(Pair(μ̃, &z))-b).norm2_squared_div2() + //+ fnR.apply(z) + reg.apply(μ) + + fnH.apply(/* opKμ.apply(&μ̃) + */ opKz.apply(&z)) + }; + + μ.merge_spikes_fitness(config.insertion.final_merging_method(), fit, |&v| v); + μ.prune(); + Pair(μ, z) +}
--- a/src/subproblem.rs Tue Aug 01 10:25:09 2023 +0300 +++ b/src/subproblem.rs Mon Feb 17 13:54:53 2025 -0500 @@ -14,15 +14,12 @@ pub mod nonneg; pub mod unconstrained; +pub mod l1squared_unconstrained; +pub mod l1squared_nonneg; -#[deprecated(since = "1.0.1", note = "Moved to submodule nonneg")] -pub use nonneg::{ - quadratic_nonneg, - quadratic_nonneg_ssn, - quadratic_nonneg_fb -}; -/// Method for solving finite-dimensional subproblems +/// Method for solving finite-dimensional subproblems. +/// Not all outer methods necessarily support all options. #[derive(Clone, Copy, Eq, PartialEq, Serialize, Deserialize, Debug)] #[allow(dead_code)] pub enum InnerMethod { @@ -30,6 +27,8 @@ FB, /// Semismooth Newton SSN, + /// PDPS + PDPS, } /// Settings for the solution of finite-dimensional subproblems @@ -37,8 +36,10 @@ pub struct InnerSettings<F : Float> { /// Method pub method : InnerMethod, - /// Proportional step length (∈ [0, 1) for `InnerMethod::FB`). - pub τ0 : F, + /// Proportional step length ∈ [0, 1) for `InnerMethod::FB`. + pub fb_τ0 : F, + /// Proportional primal and dual step lengths for `InnerMethod::PDPS`. + pub pdps_τσ0 : (F, F), /// Fraction of `tolerance` given to inner algorithm pub tolerance_mult : F, /// Iterator options @@ -50,7 +51,8 @@ impl<F : Float> Default for InnerSettings<F> { fn default() -> Self { InnerSettings { - τ0 : 0.99, + fb_τ0 : 0.99, + pdps_τσ0 : (1.98, 0.5), iterator_options : AlgIteratorOptions { // max_iter cannot be very small, as initially FB needs many iterations, although // on later invocations even one or two tends to be enough @@ -62,7 +64,7 @@ quiet : true, .. Default::default() }, - method : InnerMethod::FB, + method : InnerMethod::SSN, tolerance_mult : 0.01, } }
--- /dev/null Thu Jan 01 00:00:00 1970 +0000 +++ b/src/subproblem/l1squared_nonneg.rs Mon Feb 17 13:54:53 2025 -0500 @@ -0,0 +1,437 @@ +/*! +Iterative algorithms for solving the finite-dimensional subproblem with constraint. +*/ + +use nalgebra::DVector; +use numeric_literals::replace_float_literals; +use itertools::izip; +//use std::iter::zip; +use std::cmp::Ordering::*; + +use alg_tools::iterate::{ + AlgIteratorFactory, + AlgIteratorState, +}; +use alg_tools::nalgebra_support::ToNalgebraRealField; +use alg_tools::norms::{Dist, L1}; +use alg_tools::nanleast::NaNLeast; + +use crate::types::*; +use super::{ + InnerMethod, + InnerSettings +}; +use super::nonneg::nonneg_soft_thresholding; +use super::l1squared_unconstrained::l1squared_prox; + +/// Return maximum of `dist` and distnce of inteval `[lb, ub]` to zero. +#[replace_float_literals(F::cast_from(literal))] +pub(super) fn max_interval_dist_to_zero<F : Float>(dist : F, lb : F, ub : F) -> F { + if lb < 0.0 { + if ub > 0.0 { + dist + } else { + dist.max(-ub) + } + } else /* ub ≥ 0.0*/ { + dist.max(lb) + } +} + +/// Returns the ∞-norm minimal subdifferential of $x ↦ (β/2)|x-y|_1^2 - g^⊤ x + λ\|x\|₁ +δ_{≥}(x)$ at $x$. +/// +/// `v` will be modified and cannot be trusted to contain useful values afterwards. +#[replace_float_literals(F::cast_from(literal))] +fn min_subdifferential<F : Float + nalgebra::RealField>( + y : &DVector<F>, + x : &DVector<F>, + g : &DVector<F>, + λ : F, + β : F +) -> F { + let mut val = 0.0; + let tmp = β*y.dist(x, L1); + for (&g_i, &x_i, y_i) in izip!(g.iter(), x.iter(), y.iter()) { + let (mut lb, mut ub) = (-g_i, -g_i); + match x_i.partial_cmp(y_i) { + Some(Greater) => { lb += tmp; ub += tmp }, + Some(Less) => { lb -= tmp; ub -= tmp }, + Some(Equal) => { lb -= tmp; ub += tmp }, + None => {}, + } + match x_i.partial_cmp(&0.0) { + Some(Greater) => { lb += λ; ub += λ }, + // Less should not happen + Some(Less|Equal) => { lb = F::NEG_INFINITY; ub += λ }, + None => {}, + }; + val = max_interval_dist_to_zero(val, lb, ub); + } + val +} + +#[replace_float_literals(F::cast_from(literal))] +fn lbd_soft_thresholding<F : Float>(v : F, λ : F, b : F) -> F +{ + match (b >= 0.0, v >= b) { + (true, false) => b, + (true, true) => b.max(v - λ), // soft-to-b from above + (false, true) => super::unconstrained::soft_thresholding(v, λ), + (false, false) => 0.0.min(b.max(v - λ)), // soft-to-0 with lower bound + } +} + +/// Calculate $prox_f(x)$ for $f(x)=\frac{β}{2}\norm{x-y}_1^2 + δ_{≥0}(x)$. +/// +/// To derive an algorithm for this, we can use +/// $prox_f(x) = prox_{f_0}(x - y) - y$ for +/// $f_0(z)=\frac{β}{2}\norm{z}_1^2 + δ_{≥-y}(z)$. +/// Now, the optimality conditions for $w = prox_{f_0}(x)$ are +/// $$\tag{*} +/// x ∈ w + β\norm{w}_1\sign w + N_{≥ -y}(w). +/// $$ +/// If we know $\norm{w}_1$, then this is easily solved by lower-bounded soft-thresholding. +/// We find this by sorting the elements by the distance to the 'locked' lower-bounded +/// soft-thresholding target ($0$ or $-y_i$). +/// Then we loop over this sorted vector, increasing our estimate of $\norm{w}_1$ as we decide +/// that the soft-thresholding parameter `β\norm{w}_1` has to be such that the passed elements +/// will reach their locked value (after which they cannot change anymore, for a larger +/// soft-thresholding parameter. This has to be slightly more fine-grained for account +/// for the case that $-y_i<0$ and $x_i < -y_i$. +/// +/// Indeed, we denote by x' and w' the subset of elements such that w_i ≠ 0 and w_i > -y_i, +/// we can calculate by applying $⟨\cdot, \sign w'⟩$ to the corresponding lines of (*) that +/// $$ +/// \norm{x'} = \norm{w'} + β \norm{w}_1 m. +/// $$ +/// Having a value for $t = \norm{w}-\norm{w'}$, we can then calculate +/// $$ +/// \norm{x'} - t = (1+β m)\norm{w}_1, +/// $$ +/// from where we can calculate the soft-thresholding parameter $λ=β\norm{w}_1$. +/// Since we do not actually know the unlocked elements, but just loop over all the possibilities +/// for them, we have to check that $λ$ is above the current lower bound for this parameter +/// (`shift` in the code), and below a value that would cause changes in the locked set +/// (`max_shift` in the code). +#[replace_float_literals(F::cast_from(literal))] +pub(super) fn l1squared_nonneg_prox<F :Float + nalgebra::RealField>( + sorted : &mut Vec<(F, F, F, Option<(F, F)>)>, + x : &mut DVector<F>, + y : &DVector<F>, + β : F +) { + // nalgebra double-definition bullshit workaround + //let max = alg_tools::NumTraitsFloat::max; + let abs = alg_tools::NumTraitsFloat::abs; + + *x -= y; + + for (az_x_i, &x_i, &y_i) in izip!(sorted.iter_mut(), x.iter(), y.iter()) { + // The first component of each az_x_i contains the distance of x_i to the + // soft-thresholding limit. If it is negative, it is always reached. + // The second component contains the absolute value of the result for that component + // w_i of the solution, if the soft-thresholding limit is reached. + // This is stored here due to the sorting, although could otherwise be computed directly. + // Likewise the third component contains the absolute value of x_i. + // The fourth component contains an initial lower bound. + let a_i = abs(x_i); + let b = -y_i; + *az_x_i = match (b >= 0.0, x_i >= b) { + (true, false) => (x_i-b, b, a_i, None), // w_i=b, so sorting element negative! + (true, true) => (x_i-b, b, a_i, None), // soft-to-b from above + (false, true) => (a_i, 0.0, a_i, None), // soft-to-0 + (false, false) => (a_i, 0.0, a_i, Some((b, b-x_i))), // soft-to-0 with initial limit + }; + } + sorted.as_mut_slice() + .sort_unstable_by(|(a, _, _, _), (b, _, _, _)| NaNLeast(*a).cmp(&NaNLeast(*b))); + + let mut nwlow = 0.0; + let mut shift = 0.0; + // This main loop is over different combinations of elements of the solution locked + // to the soft-thresholding lower bound (`0` or `-y_i`), in the sorted order of locking. + for (i, az_x_i) in izip!(0.., sorted.iter()) { + // This 'attempt is over different combinations of elements locked to the + // lower bound (`-y_i ≤ 0`). It calculates `max_shift` as the maximum shift that + // can be done until the locking would change (or become non-strictly-complementary). + // If the main rule (*) gives an estimate of `λ` that stays below `max_shift`, it is + // accepted. Otherwise `shift` is updated to `max_shift`, and we attempt again, + // with the non-locking set that participates in the calculation of `λ` then including + // the elements that are no longer locked to the lower bound. + 'attempt: loop { + let mut nwthis = 0.0; // contribution to ‖w‖ from elements with locking + //soft-thresholding parameter = `shift` + let mut nxmore = 0.0; // ‖x'‖ for those elements through to not be locked to + // neither the soft-thresholding limit, nor the lower bound + let mut nwlbd = 0.0; // contribution to ‖w‖ from those elements locked to their + // lower bound + let mut m = 0; + let mut max_shift = F::INFINITY; // maximal shift for which our estimate of the set of + // unlocked elements is valid. + let mut max_shift_from_lbd = false; // Whether max_shift comes from the next element + // or from a lower bound being reached. + for az_x_j in sorted[i as usize..].iter() { + if az_x_j.0 <= shift { + nwthis += az_x_j.1; + } else { + match az_x_j.3 { + Some((l, s)) if shift < s => { + if max_shift > s { + max_shift_from_lbd = true; + max_shift = s; + } + nwlbd += -l; + }, + _ => { + nxmore += az_x_j.2; + if m == 0 && max_shift > az_x_j.0 { + max_shift = az_x_j.0; + max_shift_from_lbd = false; + } + m += 1; + } + } + } + } + + // We need ‖x'‖ = ‖w'‖ + β m ‖w‖, i.e. ‖x'‖ - (‖w‖-‖w'‖)= (1 + β m)‖w‖. + let tmp = β*(nxmore - (nwlow + nwthis + nwlbd))/(1.0 + β*F::cast_from(m)); + if tmp > max_shift { + if max_shift_from_lbd { + shift = max_shift; + continue 'attempt; + } else { + break 'attempt + } + } else if tmp < shift { + // TODO: this should probably be an assert!(false) + break 'attempt; + } else { + // success + x.zip_apply(y, |x_i, y_i| *x_i = y_i + lbd_soft_thresholding(*x_i, tmp, -y_i)); + return + } + } + shift = az_x_i.0; + nwlow += az_x_i.1; + } + // TODO: this fallback has not been verified to be correct + x.zip_apply(y, |x_i, y_i| *x_i = y_i + lbd_soft_thresholding(*x_i, shift, -y_i)); +} + +/// Proximal point method implementation of [`l1squared_nonneg`]. +/// For detailed documentation of the inputs and outputs, refer to there. +/// +/// The `λ` component of the model is handled in the proximal step instead of the gradient step +/// for potential performance improvements. +#[replace_float_literals(F::cast_from(literal).to_nalgebra_mixed())] +pub fn l1squared_nonneg_pp<F, I>( + y : &DVector<F::MixedType>, + g : &DVector<F::MixedType>, + λ_ : F, + β_ : F, + x : &mut DVector<F::MixedType>, + τ_ : F, + θ_ : F, + iterator : I +) -> usize +where F : Float + ToNalgebraRealField, + I : AlgIteratorFactory<F> +{ + let λ = λ_.to_nalgebra_mixed(); + let β = β_.to_nalgebra_mixed(); + let mut τ = τ_.to_nalgebra_mixed(); + let θ = θ_.to_nalgebra_mixed(); + let mut tmp = std::iter::repeat((0.0, 0.0, 0.0, None)).take(x.len()).collect(); + let mut iters = 0; + + iterator.iterate(|state| { + // Primal step: x^{k+1} = prox_{(τβ/2)|.-y|_1^2+δ_{≥0}+}(x^k - τ(λ𝟙^⊤-g)) + x.apply(|x_i| *x_i -= τ*λ); + x.axpy(τ, g, 1.0); + l1squared_nonneg_prox(&mut tmp, x, y, τ*β); + + iters += 1; + // This gives O(1/N^2) rates due to monotonicity of function values. + // Higher acceleration does not seem to be numerically stable. + τ += θ; + + // This gives O(1/N^3) rates due to monotonicity of function values. + // Higher acceleration does not seem to be numerically stable. + //τ + = F::cast_from(iters).to_nalgebra_mixed()*θ; + + state.if_verbose(|| { + F::from_nalgebra_mixed(min_subdifferential(y, x, g, λ, β)) + }) + }); + + iters +} + +/// PDPS implementation of [`l1squared_nonneg`]. +/// For detailed documentation of the inputs and outputs, refer to there. +/// +/// The `λ` component of the model is handled in the proximal step instead of the gradient step +/// for potential performance improvements. +/// The parameter `θ` is used to multiply the rescale the operator (identity) of the PDPS model. +#[replace_float_literals(F::cast_from(literal).to_nalgebra_mixed())] +pub fn l1squared_nonneg_pdps<F, I>( + y : &DVector<F::MixedType>, + g : &DVector<F::MixedType>, + λ_ : F, + β_ : F, + x : &mut DVector<F::MixedType>, + τ_ : F, + σ_ : F, + θ_ : F, + iterator : I +) -> usize +where F : Float + ToNalgebraRealField, + I : AlgIteratorFactory<F> +{ + let λ = λ_.to_nalgebra_mixed(); + let β = β_.to_nalgebra_mixed(); + let τ = τ_.to_nalgebra_mixed(); + let σ = σ_.to_nalgebra_mixed(); + let θ = θ_.to_nalgebra_mixed(); + let mut w = DVector::zeros(x.len()); + let mut tmp = DVector::zeros(x.len()); + let mut xprev = x.clone(); + let mut iters = 0; + + iterator.iterate(|state| { + // Primal step: x^{k+1} = prox_{(τβ/2)|.-y|_1^2}(x^k - τ (w^k - g)) + x.axpy(-τ*θ, &w, 1.0); + x.axpy(τ, g, 1.0); + l1squared_prox(&mut tmp, x, y, τ*β); + + // Dual step: w^{k+1} = proj_{[-∞,λ]}(w^k + σ(2x^{k+1}-x^k)) + w.axpy(2.0*σ*θ, x, 1.0); + w.axpy(-σ*θ, &xprev, 1.0); + w.apply(|w_i| *w_i = w_i.min(λ)); + xprev.copy_from(x); + + iters +=1; + + state.if_verbose(|| { + F::from_nalgebra_mixed(min_subdifferential(y, x, g, λ, β)) + }) + }); + + iters +} + +/// Alternative PDPS implementation of [`l1squared_nonneg`]. +/// For detailed documentation of the inputs and outputs, refer to there. +/// +/// By not dualising the 1-norm, this should produce more sparse solutions than +/// [`l1squared_nonneg_pdps`]. +/// +/// The `λ` component of the model is handled in the proximal step instead of the gradient step +/// for potential performance improvements. +/// The parameter `θ` is used to multiply the rescale the operator (identity) of the PDPS model. +/// We rewrite +/// <div>$$ +/// \begin{split} +/// & \min_{x ∈ ℝ^n} \frac{β}{2} |x-y|_1^2 - g^⊤ x + λ\|x\|₁ + δ_{≥ 0}(x) \\ +/// & = \min_{x ∈ ℝ^n} \max_{w} ⟨θ w, x⟩ - g^⊤ x + λ\|x\|₁ + δ_{≥ 0}(x) +/// - \left(x ↦ \frac{β}{2θ} |x-y|_1^2 \right)^*(w). +/// \end{split} +/// $$</div> +#[replace_float_literals(F::cast_from(literal).to_nalgebra_mixed())] +pub fn l1squared_nonneg_pdps_alt<F, I>( + y : &DVector<F::MixedType>, + g : &DVector<F::MixedType>, + λ_ : F, + β_ : F, + x : &mut DVector<F::MixedType>, + τ_ : F, + σ_ : F, + θ_ : F, + iterator : I +) -> usize +where F : Float + ToNalgebraRealField, + I : AlgIteratorFactory<F> +{ + let λ = λ_.to_nalgebra_mixed(); + let τ = τ_.to_nalgebra_mixed(); + let σ = σ_.to_nalgebra_mixed(); + let θ = θ_.to_nalgebra_mixed(); + let β = β_.to_nalgebra_mixed(); + let σθ = σ*θ; + let τθ = τ*θ; + let mut w = DVector::zeros(x.len()); + let mut tmp = DVector::zeros(x.len()); + let mut xprev = x.clone(); + let mut iters = 0; + + iterator.iterate(|state| { + // Primal step: x^{k+1} = nonnegsoft_τλ(x^k - τ(θ w^k -g)) + x.axpy(-τθ, &w, 1.0); + x.axpy(τ, g, 1.0); + x.apply(|x_i| *x_i = nonneg_soft_thresholding(*x_i, τ*λ)); + + // Dual step: with g(x) = (β/(2θ))‖x-y‖₁² and q = w^k + σ(2x^{k+1}-x^k), + // we compute w^{k+1} = prox_{σg^*}(q) for + // = q - σ prox_{g/σ}(q/σ) + // = q - σ prox_{(β/(2θσ))‖.-y‖₁²}(q/σ) + // = σ(q/σ - prox_{(β/(2θσ))‖.-y‖₁²}(q/σ)) + // where q/σ = w^k/σ + (2x^{k+1}-x^k), + w /= σ; + w.axpy(2.0, x, 1.0); + w.axpy(-1.0, &xprev, 1.0); + xprev.copy_from(&w); // use xprev as temporary variable + l1squared_prox(&mut tmp, &mut xprev, y, β/σθ); + w -= &xprev; + w *= σ; + xprev.copy_from(x); + + iters += 1; + + state.if_verbose(|| { + F::from_nalgebra_mixed(min_subdifferential(y, x, g, λ, β)) + }) + }); + + iters +} + + +/// This function applies an iterative method for the solution of the problem +/// <div>$$ +/// \min_{x ∈ ℝ^n} \frac{β}{2} |x-y|_1^2 - g^⊤ x + λ\|x\|₁ + δ_{≥ 0}(x). +/// $$</div> +/// +/// This function returns the number of iterations taken. +#[replace_float_literals(F::cast_from(literal))] +pub fn l1squared_nonneg<F, I>( + y : &DVector<F::MixedType>, + g : &DVector<F::MixedType>, + λ : F, + β : F, + x : &mut DVector<F::MixedType>, + inner : &InnerSettings<F>, + iterator : I +) -> usize +where F : Float + ToNalgebraRealField, + I : AlgIteratorFactory<F> +{ + match inner.method { + InnerMethod::PDPS => { + let inner_θ = 1.0; + // Estimate of ‖K‖ for K=θ\Id. + let normest = inner_θ; + let (inner_τ, inner_σ) = (inner.pdps_τσ0.0 / normest, inner.pdps_τσ0.1 / normest); + l1squared_nonneg_pdps_alt(y, g, λ, β, x, inner_τ, inner_σ, inner_θ, iterator) + }, + InnerMethod::FB => { + // The Lipschitz factor of ∇[x ↦ g^⊤ x + λ∑x]=g - λ𝟙 is FB is just a proximal point + // method with on constraints on τ. We “accelerate” it by adding to τ the constant θ + // on each iteration. Exponential growth does not seem stable. + let inner_τ = inner.fb_τ0; + let inner_θ = inner_τ; + l1squared_nonneg_pp(y, g, λ, β, x, inner_τ, inner_θ, iterator) + }, + other => unimplemented!("${other:?} is unimplemented"), + } +}
--- /dev/null Thu Jan 01 00:00:00 1970 +0000 +++ b/src/subproblem/l1squared_unconstrained.rs Mon Feb 17 13:54:53 2025 -0500 @@ -0,0 +1,271 @@ +/*! +Iterative algorithms for solving the finite-dimensional subproblem without constraints. +*/ + +use nalgebra::DVector; +use numeric_literals::replace_float_literals; +use itertools::izip; +use std::cmp::Ordering::*; + +use std::iter::zip; +use alg_tools::iterate::{ + AlgIteratorFactory, + AlgIteratorState, +}; +use alg_tools::nalgebra_support::ToNalgebraRealField; +use alg_tools::nanleast::NaNLeast; +use alg_tools::norms::{Dist, L1}; + +use crate::types::*; +use super::{ + InnerMethod, + InnerSettings +}; +use super::unconstrained::soft_thresholding; +use super::l1squared_nonneg::max_interval_dist_to_zero; + +/// Calculate $prox_f(x)$ for $f(x)=\frac{β}{2}\norm{x-y}_1^2$. +/// +/// To derive an algorithm for this, we can assume that $y=0$, as +/// $prox_f(x) = prox_{f_0}(x - y) - y$ for $f_0=\frac{β}{2}\norm{x}_1^2$. +/// Now, the optimality conditions for $w = prox_f(x)$ are +/// $$\tag{*} +/// 0 ∈ w-x + β\norm{w}_1\sign w. +/// $$ +/// Clearly then $w = \soft_{β\norm{w}_1}(x)$. +/// Thus the components of $x$ with smallest absolute value will be zeroed out. +/// Denoting by $w'$ the non-zero components, and by $x'$ the corresponding components +/// of $x$, and by $m$ their count, multipying the corresponding lines of (*) by $\sign x'$, +/// we obtain +/// $$ +/// \norm{x'}_1 = (1+βm)\norm{w'}_1. +/// $$ +/// That is, $\norm{w}_1=\norm{w'}_1=\norm{x'}_1/(1+βm)$. +/// Thus, sorting $x$ by absolute value, and sequentially in order eliminating the smallest +/// elements, we can easily calculate what $\norm{w}_1$ should be for that choice, and +/// then easily calculate $w = \soft_{β\norm{w}_1}(x)$. We just have to verify that +/// the resulting $w$ has the same norm. There's a shortcut to this, as we work +/// sequentially: just check that the smallest assumed-nonzero component $i$ satisfies the +/// condition of soft-thresholding to remain non-zero: $|x_i|>τ\norm{x'}/(1+τm)$. +/// Clearly, if this condition fails for x_i, it will fail for all the components +/// already exluced. While, if it holds, it will hold for all components not excluded. +#[replace_float_literals(F::cast_from(literal))] +pub(super) fn l1squared_prox<F :Float + nalgebra::RealField>( + sorted_abs : &mut DVector<F>, + x : &mut DVector<F>, + y : &DVector<F>, + β : F +) { + sorted_abs.copy_from(x); + sorted_abs.axpy(-1.0, y, 1.0); + sorted_abs.apply(|z_i| *z_i = num_traits::abs(*z_i)); + sorted_abs.as_mut_slice().sort_unstable_by(|a, b| NaNLeast(*a).cmp(&NaNLeast(*b))); + + let mut n = sorted_abs.sum(); + for (m, az_i) in zip((1..=x.len() as u32).rev(), sorted_abs) { + // test first + let tmp = β*n/(1.0 + β*F::cast_from(m)); + if *az_i <= tmp { + // Fail + n -= *az_i; + } else { + // Success + x.zip_apply(y, |x_i, y_i| *x_i = y_i + soft_thresholding(*x_i-y_i, tmp)); + return + } + } + // m = 0 should always work, but x is zero. + x.fill(0.0); +} + +/// Returns the ∞-norm minimal subdifferential of $x ↦ (β/2)|x-y|_1^2 - g^⊤ x + λ\|x\|₁$ at $x$. +/// +/// `v` will be modified and cannot be trusted to contain useful values afterwards. +#[replace_float_literals(F::cast_from(literal))] +fn min_subdifferential<F : Float + nalgebra::RealField>( + y : &DVector<F>, + x : &DVector<F>, + g : &DVector<F>, + λ : F, + β : F +) -> F { + let mut val = 0.0; + let tmp = β*y.dist(x, L1); + for (&g_i, &x_i, y_i) in izip!(g.iter(), x.iter(), y.iter()) { + let (mut lb, mut ub) = (-g_i, -g_i); + match x_i.partial_cmp(y_i) { + Some(Greater) => { lb += tmp; ub += tmp }, + Some(Less) => { lb -= tmp; ub -= tmp }, + Some(Equal) => { lb -= tmp; ub += tmp }, + None => {}, + } + match x_i.partial_cmp(&0.0) { + Some(Greater) => { lb += λ; ub += λ }, + Some(Less) => { lb -= λ; ub -= λ }, + Some(Equal) => { lb -= λ; ub += λ }, + None => {}, + }; + val = max_interval_dist_to_zero(val, lb, ub); + } + val +} + + +/// PDPS implementation of [`l1squared_unconstrained`]. +/// For detailed documentation of the inputs and outputs, refer to there. +/// +/// The `λ` component of the model is handled in the proximal step instead of the gradient step +/// for potential performance improvements. +#[replace_float_literals(F::cast_from(literal).to_nalgebra_mixed())] +pub fn l1squared_unconstrained_pdps<F, I>( + y : &DVector<F::MixedType>, + g : &DVector<F::MixedType>, + λ_ : F, + β_ : F, + x : &mut DVector<F::MixedType>, + τ_ : F, + σ_ : F, + iterator : I +) -> usize +where F : Float + ToNalgebraRealField, + I : AlgIteratorFactory<F> +{ + let λ = λ_.to_nalgebra_mixed(); + let β = β_.to_nalgebra_mixed(); + let τ = τ_.to_nalgebra_mixed(); + let σ = σ_.to_nalgebra_mixed(); + let mut w = DVector::zeros(x.len()); + let mut tmp = DVector::zeros(x.len()); + let mut xprev = x.clone(); + let mut iters = 0; + + iterator.iterate(|state| { + // Primal step: x^{k+1} = prox_{τ|.-y|_1^2}(x^k - τ (w^k - g)) + x.axpy(-τ, &w, 1.0); + x.axpy(τ, g, 1.0); + l1squared_prox(&mut tmp, x, y, τ*β); + + // Dual step: w^{k+1} = proj_{[-λ,λ]}(w^k + σ(2x^{k+1}-x^k)) + w.axpy(2.0*σ, x, 1.0); + w.axpy(-σ, &xprev, 1.0); + w.apply(|w_i| *w_i = num_traits::clamp(*w_i, -λ, λ)); + xprev.copy_from(x); + + iters +=1; + + state.if_verbose(|| { + F::from_nalgebra_mixed(min_subdifferential(y, x, g, λ, β)) + }) + }); + + iters +} + +/// Alternative PDPS implementation of [`l1squared_unconstrained`]. +/// For detailed documentation of the inputs and outputs, refer to there. +/// +/// By not dualising the 1-norm, this should produce more sparse solutions than +/// [`l1squared_unconstrained_pdps`]. +/// +/// The `λ` component of the model is handled in the proximal step instead of the gradient step +/// for potential performance improvements. +/// The parameter `θ` is used to multiply the rescale the operator (identity) of the PDPS model. +/// We rewrite +/// <div>$$ +/// \begin{split} +/// & \min_{x ∈ ℝ^n} \frac{β}{2} |x-y|_1^2 - g^⊤ x + λ\|x\|₁ \\ +/// & = \min_{x ∈ ℝ^n} \max_{w} ⟨θ w, x⟩ - g^⊤ x + λ\|x\|₁ +/// - \left(x ↦ \frac{β}{2θ} |x-y|_1^2 \right)^*(w). +/// \end{split} +/// $$</div> +#[replace_float_literals(F::cast_from(literal).to_nalgebra_mixed())] +pub fn l1squared_unconstrained_pdps_alt<F, I>( + y : &DVector<F::MixedType>, + g : &DVector<F::MixedType>, + λ_ : F, + β_ : F, + x : &mut DVector<F::MixedType>, + τ_ : F, + σ_ : F, + θ_ : F, + iterator : I +) -> usize +where F : Float + ToNalgebraRealField, + I : AlgIteratorFactory<F> +{ + let λ = λ_.to_nalgebra_mixed(); + let τ = τ_.to_nalgebra_mixed(); + let σ = σ_.to_nalgebra_mixed(); + let θ = θ_.to_nalgebra_mixed(); + let β = β_.to_nalgebra_mixed(); + let σθ = σ*θ; + let τθ = τ*θ; + let mut w = DVector::zeros(x.len()); + let mut tmp = DVector::zeros(x.len()); + let mut xprev = x.clone(); + let mut iters = 0; + + iterator.iterate(|state| { + // Primal step: x^{k+1} = soft_τλ(x^k - τ(θ w^k -g)) + x.axpy(-τθ, &w, 1.0); + x.axpy(τ, g, 1.0); + x.apply(|x_i| *x_i = soft_thresholding(*x_i, τ*λ)); + + // Dual step: with g(x) = (β/(2θ))‖x-y‖₁² and q = w^k + σ(2x^{k+1}-x^k), + // we compute w^{k+1} = prox_{σg^*}(q) for + // = q - σ prox_{g/σ}(q/σ) + // = q - σ prox_{(β/(2θσ))‖.-y‖₁²}(q/σ) + // = σ(q/σ - prox_{(β/(2θσ))‖.-y‖₁²}(q/σ)) + // where q/σ = w^k/σ + (2x^{k+1}-x^k), + w /= σ; + w.axpy(2.0, x, 1.0); + w.axpy(-1.0, &xprev, 1.0); + xprev.copy_from(&w); // use xprev as temporary variable + l1squared_prox(&mut tmp, &mut xprev, y, β/σθ); + w -= &xprev; + w *= σ; + xprev.copy_from(x); + + iters += 1; + + state.if_verbose(|| { + F::from_nalgebra_mixed(min_subdifferential(y, x, g, λ, β)) + }) + }); + + iters +} + + +/// This function applies an iterative method for the solution of the problem +/// <div>$$ +/// \min_{x ∈ ℝ^n} \frac{β}{2} |x-y|_1^2 - g^⊤ x + λ\|x\|₁. +/// $$</div> +/// Only PDPS is supported. +/// +/// This function returns the number of iterations taken. +#[replace_float_literals(F::cast_from(literal))] +pub fn l1squared_unconstrained<F, I>( + y : &DVector<F::MixedType>, + g : &DVector<F::MixedType>, + λ : F, + β : F, + x : &mut DVector<F::MixedType>, + inner : &InnerSettings<F>, + iterator : I +) -> usize +where F : Float + ToNalgebraRealField, + I : AlgIteratorFactory<F> +{ + // Estimate of ‖K‖ for K=θ Id. + let inner_θ = 1.0; + let normest = inner_θ; + + let (inner_τ, inner_σ) = (inner.pdps_τσ0.0 / normest, inner.pdps_τσ0.1 / normest); + + match inner.method { + InnerMethod::PDPS => + l1squared_unconstrained_pdps_alt(y, g, λ, β, x, inner_τ, inner_σ, inner_θ, iterator), + other => unimplemented!("${other:?} is unimplemented"), + } +}
--- a/src/subproblem/nonneg.rs Tue Aug 01 10:25:09 2023 +0300 +++ b/src/subproblem/nonneg.rs Mon Feb 17 13:54:53 2025 -0500 @@ -6,6 +6,7 @@ use numeric_literals::replace_float_literals; use itertools::{izip, Itertools}; use colored::Colorize; +use std::cmp::Ordering::*; use alg_tools::iter::Mappable; use alg_tools::error::NumericalError; @@ -22,38 +23,42 @@ InnerMethod, InnerSettings }; +use super::l1squared_nonneg::max_interval_dist_to_zero; /// Compute the proximal operator of $x \mapsto x + \delta\_{[0, \infty)}$, i.e., /// the non-negativity contrained soft-thresholding operator. #[inline] #[replace_float_literals(F::cast_from(literal))] -fn nonneg_soft_thresholding<F : Float>(v : F, λ : F) -> F { +pub(super) fn nonneg_soft_thresholding<F : Float>(v : F, λ : F) -> F { (v - λ).max(0.0) } -/// Returns the ∞-norm minimal subdifferential of $x ↦ x^⊤Ax - g^⊤ x + λ\vec 1^⊤ x δ_{≥ 0}(x)$ +/// Returns the ∞-norm minimal subdifferential of $x ↦ x^⊤Ax/2 - g^⊤ x + λ\vec 1^⊤ x + δ_{≥ 0}(x)$ /// at $x$. /// /// `v` will be modified and cannot be trusted to contain useful values afterwards. -#[replace_float_literals(F::cast_from(literal).to_nalgebra_mixed())] -fn min_subdifferential<F : Float + ToNalgebraRealField>( - v : &mut DVector<F::MixedType>, - mA : &DMatrix<F::MixedType>, - x : &DVector<F::MixedType>, - g : &DVector<F::MixedType>, - λ : F::MixedType +#[replace_float_literals(F::cast_from(literal))] +fn min_subdifferential<F : Float + nalgebra::RealField>( + v : &mut DVector<F>, + mA : &DMatrix<F>, + x : &DVector<F>, + g : &DVector<F>, + λ : F ) -> F { v.copy_from(g); mA.gemv(v, 1.0, x, -1.0); // v = Ax - g let mut val = 0.0; for (&v_i, &x_i) in izip!(v.iter(), x.iter()) { - // The subdifferential of the objective is $Ax - g + λ + ∂ δ_{≥ 0}(x)$. - let d = v_i + λ; - if x_i > 0.0 || d < 0.0 { - val = val.max(d.abs()); + let (mut lb, mut ub) = (v_i, v_i); + match x_i.partial_cmp(&0.0) { + Some(Greater) => { lb += λ; ub += λ }, + // Less should not happen + Some(Less|Equal) => { lb = F::NEG_INFINITY; ub += λ }, + None => {}, } + val = max_interval_dist_to_zero(val, lb, ub); } - F::from_nalgebra_mixed(val) + val } /// Forward-backward splitting implementation of [`quadratic_nonneg`]. @@ -98,7 +103,7 @@ iters +=1; backup.map(|_| { - min_subdifferential(&mut v, mA, x, g, λ) + F::from_nalgebra_mixed(min_subdifferential(&mut v, mA, x, g, λ)) }) }); @@ -281,7 +286,7 @@ // 4. Report solution quality state.if_verbose(|| { // Calculate subdifferential at the FB step `x` that hasn't yet had `s` yet added. - min_subdifferential(&mut v, mA, x, g, λ) + F::from_nalgebra_mixed(min_subdifferential(&mut v, mA, x, g, λ)) }) }); @@ -291,7 +296,7 @@ /// This function applies an iterative method for the solution of the quadratic non-negativity /// constrained problem /// <div>$$ -/// \min_{x ∈ ℝ^n} \frac{1}{2} x^⊤Ax - g^⊤ x + λ{\vec 1}^⊤ x + c + δ_{≥ 0}(x). +/// \min_{x ∈ ℝ^n} \frac{1}{2} x^⊤Ax - g^⊤ x + λ{\vec 1}^⊤ x + δ_{≥ 0}(x). /// $$</div> /// Semismooth Newton or forward-backward are supported based on the setting in `method`. /// The parameter `mA` is matrix $A$, and `g` and `λ` are as in the mathematical formulation. @@ -307,27 +312,28 @@ /// /// This function returns the number of iterations taken. pub fn quadratic_nonneg<F, I>( - method : InnerMethod, mA : &DMatrix<F::MixedType>, g : &DVector<F::MixedType>, - //c_ : F, λ : F, x : &mut DVector<F::MixedType>, - τ : F, + mA_normest : F, + inner : &InnerSettings<F>, iterator : I ) -> usize where F : Float + ToNalgebraRealField, I : AlgIteratorFactory<F> { - - match method { + let inner_τ = inner.fb_τ0 / mA_normest; + + match inner.method { InnerMethod::FB => - quadratic_nonneg_fb(mA, g, λ, x, τ, iterator), + quadratic_nonneg_fb(mA, g, λ, x, inner_τ, iterator), InnerMethod::SSN => - quadratic_nonneg_ssn(mA, g, λ, x, τ, iterator).unwrap_or_else(|e| { + quadratic_nonneg_ssn(mA, g, λ, x, inner_τ, iterator).unwrap_or_else(|e| { println!("{}", format!("{e}. Using FB fallback.").red()); let ins = InnerSettings::<F>::default(); - quadratic_nonneg_fb(mA, g, λ, x, τ, ins.iterator_options) - }) + quadratic_nonneg_fb(mA, g, λ, x, inner_τ, ins.iterator_options) + }), + other => unimplemented!("${other:?} is unimplemented"), } }
--- a/src/subproblem/unconstrained.rs Tue Aug 01 10:25:09 2023 +0300 +++ b/src/subproblem/unconstrained.rs Mon Feb 17 13:54:53 2025 -0500 @@ -23,11 +23,12 @@ InnerMethod, InnerSettings }; +use super::l1squared_nonneg::max_interval_dist_to_zero; /// Compute the proximal operator of $x \mapsto |x|$, i.e., the soft-thresholding operator. #[inline] #[replace_float_literals(F::cast_from(literal))] -fn soft_thresholding<F : Float>(v : F, λ : F) -> F { +pub(crate) fn soft_thresholding<F : Float>(v : F, λ : F) -> F { if v > λ { v - λ } else if v < -λ { @@ -40,27 +41,28 @@ /// Returns the ∞-norm minimal subdifferential of $x ↦ x^⊤Ax - g^⊤ x + λ\|x\|₁$ at $x$. /// /// `v` will be modified and cannot be trusted to contain useful values afterwards. -#[replace_float_literals(F::cast_from(literal).to_nalgebra_mixed())] -fn min_subdifferential<F : Float + ToNalgebraRealField>( - v : &mut DVector<F::MixedType>, - mA : &DMatrix<F::MixedType>, - x : &DVector<F::MixedType>, - g : &DVector<F::MixedType>, - λ : F::MixedType +#[replace_float_literals(F::cast_from(literal))] +fn min_subdifferential<F : Float + nalgebra::RealField>( + v : &mut DVector<F>, + mA : &DMatrix<F>, + x : &DVector<F>, + g : &DVector<F>, + λ : F ) -> F { v.copy_from(g); mA.gemv(v, 1.0, x, -1.0); // v = Ax - g let mut val = 0.0; for (&v_i, &x_i) in izip!(v.iter(), x.iter()) { - // The subdifferential at x is $Ax - g + λ ∂‖·‖₁(x)$. - val = val.max(match x_i.partial_cmp(&0.0) { - Some(Greater) => v_i + λ, - Some(Less) => v_i - λ, - Some(Equal) => soft_thresholding(v_i, λ), - None => F::MixedType::nan(), - }) + let (mut lb, mut ub) = (v_i, v_i); + match x_i.partial_cmp(&0.0) { + Some(Greater) => { lb += λ; ub += λ }, + Some(Less) => { lb -= λ; ub -= λ }, + Some(Equal) => { lb -= λ; ub += λ }, + None => {}, + } + val = max_interval_dist_to_zero(val, lb, ub); } - F::from_nalgebra_mixed(val) + val } @@ -106,7 +108,7 @@ iters +=1; backup.map(|_| { - min_subdifferential(&mut v, mA, x, g, λ) + F::from_nalgebra_mixed(min_subdifferential(&mut v, mA, x, g, λ)) }) }); @@ -242,7 +244,7 @@ // 4. Report solution quality state.if_verbose(|| { // Calculate subdifferential at the FB step `x` that hasn't yet had `s` yet added. - min_subdifferential(&mut v, mA, x, g, λ) + F::from_nalgebra_mixed(min_subdifferential(&mut v, mA, x, g, λ)) }) }); @@ -251,7 +253,7 @@ /// This function applies an iterative method for the solution of the problem /// <div>$$ -/// \min_{x ∈ ℝ^n} \frac{1}{2} x^⊤Ax - g^⊤ x + λ\|x\|₁ + c. +/// \min_{x ∈ ℝ^n} \frac{1}{2} x^⊤Ax - g^⊤ x + λ\|x\|₁. /// $$</div> /// Semismooth Newton or forward-backward are supported based on the setting in `method`. /// The parameter `mA` is matrix $A$, and `g` and `λ` are as in the mathematical formulation. @@ -262,27 +264,28 @@ /// /// This function returns the number of iterations taken. pub fn quadratic_unconstrained<F, I>( - method : InnerMethod, mA : &DMatrix<F::MixedType>, g : &DVector<F::MixedType>, - //c_ : F, λ : F, x : &mut DVector<F::MixedType>, - τ : F, + mA_normest : F, + inner : &InnerSettings<F>, iterator : I ) -> usize where F : Float + ToNalgebraRealField, I : AlgIteratorFactory<F> { + let inner_τ = inner.fb_τ0 / mA_normest; - match method { + match inner.method { InnerMethod::FB => - quadratic_unconstrained_fb(mA, g, λ, x, τ, iterator), + quadratic_unconstrained_fb(mA, g, λ, x, inner_τ, iterator), InnerMethod::SSN => - quadratic_unconstrained_ssn(mA, g, λ, x, τ, iterator).unwrap_or_else(|e| { + quadratic_unconstrained_ssn(mA, g, λ, x, inner_τ, iterator).unwrap_or_else(|e| { println!("{}", format!("{e}. Using FB fallback.").red()); let ins = InnerSettings::<F>::default(); - quadratic_unconstrained_fb(mA, g, λ, x, τ, ins.iterator_options) - }) + quadratic_unconstrained_fb(mA, g, λ, x, inner_τ, ins.iterator_options) + }), + other => unimplemented!("${other:?} is unimplemented"), } }
--- a/src/types.rs Tue Aug 01 10:25:09 2023 +0300 +++ b/src/types.rs Mon Feb 17 13:54:53 2025 -0500 @@ -4,7 +4,6 @@ use colored::ColoredString; use serde::{Serialize, Deserialize}; -use clap::ValueEnum; use alg_tools::iterate::LogRepr; use alg_tools::euclidean::Euclidean; use alg_tools::norms::{Norm, L1}; @@ -13,7 +12,7 @@ pub use alg_tools::loc::Loc; pub use alg_tools::sets::Cube; -use crate::measures::DiscreteMeasure; +// use crate::measures::DiscreteMeasure; /// [`Float`] with extra display and string conversion traits such that [`clap`] doesn't choke up. pub trait ClapFloat : Float @@ -27,31 +26,74 @@ pub struct IterInfo<F : Float, const N : usize> { /// Function value pub value : F, - /// Number of speaks + /// Number of spikes pub n_spikes : usize, /// Number of iterations this statistic covers pub this_iters : usize, + /// Number of spikes inserted since last IterInfo statistic + pub inserted : usize, /// Number of spikes removed by merging since last IterInfo statistic pub merged : usize, /// Number of spikes removed by pruning since last IterInfo statistic pub pruned : usize, /// Number of inner iterations since last IterInfo statistic pub inner_iters : usize, + /// Tuple of (transported mass, source mass) + pub untransported_fraction : Option<(F, F)>, + /// Tuple of (|destination mass - untransported_mass|, transported mass) + pub transport_error : Option<(F, F)>, /// Current tolerance pub ε : F, - /// Solve fin.dim problem for this measure to get the optimal `value`. - pub postprocessing : Option<DiscreteMeasure<Loc<F, N>, F>>, + // /// Solve fin.dim problem for this measure to get the optimal `value`. + // pub postprocessing : Option<RNDM<F, N>>, } +impl<F : Float, const N : usize> IterInfo<F, N> { + /// Initialise statistics with zeros. `ε` and `value` are unspecified. + pub fn new() -> Self { + IterInfo { + value : F::NAN, + n_spikes : 0, + this_iters : 0, + merged : 0, + inserted : 0, + pruned : 0, + inner_iters : 0, + ε : F::NAN, + // postprocessing : None, + untransported_fraction : None, + transport_error : None, + } + } +} + +#[replace_float_literals(F::cast_from(literal))] impl<F, const N : usize> LogRepr for IterInfo<F, N> where F : LogRepr + Float { fn logrepr(&self) -> ColoredString { - format!("{}\t| N = {}, ε = {:.8}, inner_iters_mean = {}, merged+pruned_mean = {}+{}", + format!("{}\t| N = {}, ε = {:.8}, 𝔼inner_it = {}, 𝔼ins/mer/pru = {}/{}/{}{}{}", self.value.logrepr(), self.n_spikes, self.ε, - self.inner_iters as float / self.this_iters as float, - self.merged as float / self.this_iters as float, - self.pruned as float / self.this_iters as float, + self.inner_iters as float / self.this_iters.max(1) as float, + self.inserted as float / self.this_iters.max(1) as float, + self.merged as float / self.this_iters.max(1) as float, + self.pruned as float / self.this_iters.max(1) as float, + match self.untransported_fraction { + None => format!(""), + Some((a, b)) => if b > 0.0 { + format!(", untransported {:.2}%", 100.0*a/b) + } else { + format!("") + } + }, + match self.transport_error { + None => format!(""), + Some((a, b)) => if b > 0.0 { + format!(", transport error {:.2}%", 100.0*a/b) + } else { + format!("") + } + } ).as_str().into() } } @@ -78,7 +120,7 @@ } /// Data term type -#[derive(Clone, Copy, Eq, PartialEq, Serialize, Deserialize, Debug, ValueEnum)] +#[derive(Clone, Copy, PartialEq, Serialize, Deserialize, Debug)] pub enum DataTerm { /// $\\|z\\|\_2^2/2$ L2Squared, @@ -95,3 +137,24 @@ } } } + +/// Type for indicating norm-2-squared data fidelity or transport cost. +#[derive(Clone, Copy, Serialize, Deserialize)] +pub struct L2Squared; + +/// Trait for indicating that `Self` is Lipschitz with respect to the (semi)norm `D`. +pub trait Lipschitz<M> { + /// The type of floats + type FloatType : Float; + + /// Returns the Lipschitz factor of `self` with respect to the (semi)norm `D`. + fn lipschitz_factor(&self, seminorm : M) -> Option<Self::FloatType>; +} + +/// Trait for norm-bounded functions. +pub trait NormBounded<M> { + type FloatType : Float; + + /// Returns a bound on the values of this function object in the `M`-norm. + fn norm_bound(&self, m : M) -> Self::FloatType; +}
