Fri, 02 Dec 2022 18:14:03 +0200
Clean up / remove various unused FB algorithm family hacks.
/*! This module provides [`RunnableExperiment`] for running chosen algorithms on a chosen experiment. */ use numeric_literals::replace_float_literals; use colored::Colorize; use serde::{Serialize, Deserialize}; use serde_json; use nalgebra::base::DVector; use std::hash::Hash; use chrono::{DateTime, Utc}; use cpu_time::ProcessTime; use clap::ValueEnum; use std::collections::HashMap; use std::time::Instant; use rand::prelude::{ StdRng, SeedableRng }; use rand_distr::Distribution; use alg_tools::bisection_tree::*; use alg_tools::iterate::{ Timed, AlgIteratorOptions, Verbose, AlgIteratorFactory, }; use alg_tools::logger::Logger; use alg_tools::error::DynError; use alg_tools::tabledump::TableDump; use alg_tools::sets::Cube; use alg_tools::mapping::RealMapping; use alg_tools::nalgebra_support::ToNalgebraRealField; use alg_tools::euclidean::Euclidean; use alg_tools::norms::{Norm, L1}; use alg_tools::lingrid::lingrid; use alg_tools::sets::SetOrd; use crate::kernels::*; use crate::types::*; use crate::measures::*; use crate::measures::merging::SpikeMerging; use crate::forward_model::*; use crate::fb::{ FBConfig, pointsource_fb, FBMetaAlgorithm, FBGenericConfig, }; use crate::pdps::{ PDPSConfig, L2Squared, pointsource_pdps, }; use crate::frank_wolfe::{ FWConfig, FWVariant, pointsource_fw, prepare_optimise_weights, optimise_weights, }; use crate::subproblem::InnerSettings; use crate::seminorms::*; use crate::plot::*; use crate::AlgorithmOverrides; /// Available algorithms and their configurations #[derive(Copy, Clone, Debug, Serialize, Deserialize)] pub enum AlgorithmConfig<F : Float> { FB(FBConfig<F>), FW(FWConfig<F>), PDPS(PDPSConfig<F>), } impl<F : ClapFloat> AlgorithmConfig<F> { /// Override supported parameters based on the command line. pub fn cli_override(self, cli : &AlgorithmOverrides<F>) -> Self { let override_fb_generic = |g : FBGenericConfig<F>| { FBGenericConfig { bootstrap_insertions : cli.bootstrap_insertions .as_ref() .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), .. g } }; use AlgorithmConfig::*; match self { FB(fb) => FB(FBConfig { τ0 : cli.tau0.unwrap_or(fb.τ0), insertion : override_fb_generic(fb.insertion), .. fb }), PDPS(pdps) => 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), .. pdps }), FW(fw) => FW(FWConfig { merging : cli.merging.clone().unwrap_or(fw.merging), .. fw }) } } } /// Helper struct for tagging and [`AlgorithmConfig`] or [`Experiment`] with a name. #[derive(Clone, Debug, Serialize, Deserialize)] pub struct Named<Data> { pub name : String, #[serde(flatten)] pub data : Data, } /// Shorthand algorithm configurations, to be used with the command line parser #[derive(ValueEnum, Debug, Copy, Clone, Eq, PartialEq, Hash)] pub enum DefaultAlgorithm { /// The μFB forward-backward method #[clap(name = "fb")] FB, /// The μFISTA inertial forward-backward method #[clap(name = "fista")] FISTA, /// The “fully corrective” conditional gradient method #[clap(name = "fw")] FW, /// The “relaxed conditional gradient method #[clap(name = "fwrelax")] FWRelax, /// The μPDPS primal-dual proximal splitting method #[clap(name = "pdps")] PDPS, } impl DefaultAlgorithm { /// Returns the algorithm configuration corresponding to the algorithm shorthand pub fn default_config<F : Float>(&self) -> AlgorithmConfig<F> { use DefaultAlgorithm::*; match *self { FB => AlgorithmConfig::FB(Default::default()), FISTA => AlgorithmConfig::FB(FBConfig{ meta : FBMetaAlgorithm::InertiaFISTA, .. Default::default() }), FW => AlgorithmConfig::FW(Default::default()), FWRelax => AlgorithmConfig::FW(FWConfig{ variant : FWVariant::Relaxed, .. Default::default() }), PDPS => AlgorithmConfig::PDPS(Default::default()), } } /// Returns the [`Named`] algorithm corresponding to the algorithm shorthand pub fn get_named<F : Float>(&self) -> Named<AlgorithmConfig<F>> { self.to_named(self.default_config()) } pub fn to_named<F : Float>(self, alg : AlgorithmConfig<F>) -> Named<AlgorithmConfig<F>> { let name = self.to_possible_value().unwrap().get_name().to_string(); Named{ name , data : alg } } } // // Floats cannot be hashed directly, so just hash the debug formatting // // for use as file identifier. // impl<F : Float> Hash for AlgorithmConfig<F> { // fn hash<H: Hasher>(&self, state: &mut H) { // format!("{:?}", self).hash(state); // } // } /// Plotting level configuration #[derive(Copy, Clone, Eq, PartialEq, Ord, PartialOrd, Serialize, ValueEnum, Debug)] pub enum PlotLevel { /// Plot nothing #[clap(name = "none")] None, /// Plot problem data #[clap(name = "data")] Data, /// Plot iterationwise state #[clap(name = "iter")] Iter, } /// Algorithm and iterator config for the experiments #[derive(Clone, Debug, Serialize)] #[serde(default)] pub struct Configuration<F : Float> { /// Algorithms to run pub algorithms : Vec<Named<AlgorithmConfig<F>>>, /// Options for algorithm step iteration (verbosity, etc.) pub iterator_options : AlgIteratorOptions, /// Plotting level pub plot : PlotLevel, /// Directory where to save results pub outdir : String, /// Bisection tree depth pub bt_depth : DynamicDepth, } type DefaultBT<F, const N : usize> = BT< DynamicDepth, F, usize, Bounds<F>, N >; type DefaultSeminormOp<F, K, const N : usize> = ConvolutionOp<F, K, DefaultBT<F, N>, N>; type DefaultSG<F, Sensor, Spread, const N : usize> = SensorGrid::< F, Sensor, Spread, DefaultBT<F, N>, N >; /// This is a dirty workaround to rust-csv not supporting struct flattening etc. #[derive(Serialize)] struct CSVLog<F> { iter : usize, cpu_time : f64, value : F, post_value : F, n_spikes : usize, inner_iters : usize, merged : usize, pruned : usize, this_iters : usize, } /// Collected experiment statistics #[derive(Clone, Debug, Serialize)] struct ExperimentStats<F : Float> { /// Signal-to-noise ratio in decibels ssnr : F, /// Proportion of noise in the signal as a number in $[0, 1]$. noise_ratio : F, /// When the experiment was run (UTC) when : DateTime<Utc>, } #[replace_float_literals(F::cast_from(literal))] impl<F : Float> ExperimentStats<F> { /// Calculate [`ExperimentStats`] based on a noisy `signal` and the separated `noise` signal. fn new<E : Euclidean<F>>(signal : &E, noise : &E) -> Self { let s = signal.norm2_squared(); let n = noise.norm2_squared(); let noise_ratio = (n / s).sqrt(); let ssnr = 10.0 * (s / n).log10(); ExperimentStats { ssnr, noise_ratio, when : Utc::now(), } } } /// Collected algorithm statistics #[derive(Clone, Debug, Serialize)] struct AlgorithmStats<F : Float> { /// Overall CPU time spent cpu_time : F, /// Real time spent elapsed : F } /// A wrapper for [`serde_json::to_writer_pretty`] that takes a filename as input /// and outputs a [`DynError`]. fn write_json<T : Serialize>(filename : String, data : &T) -> DynError { serde_json::to_writer_pretty(std::fs::File::create(filename)?, data)?; Ok(()) } /// 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 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>>, } /// Trait for runnable experiments pub trait RunnableExperiment<F : ClapFloat> { /// Run all algorithms of the [`Configuration`] `config` on the experiment. fn runall(&self, config : Configuration<F>) -> DynError; /// Returns the default configuration fn default_config(&self) -> Configuration<F>; /// Return algorithm default config fn algorithm_defaults(&self, alg : DefaultAlgorithm, cli : &AlgorithmOverrides<F>) -> Named<AlgorithmConfig<F>>; } impl<F, NoiseDistr, S, K, P, const N : usize> RunnableExperiment<F> for Named<Experiment<F, NoiseDistr, S, K, P, N>> where F : ClapFloat + nalgebra::RealField + ToNalgebraRealField<MixedType=F>, [usize; N] : Serialize, S : Sensor<F, N> + Copy + Serialize, P : Spread<F, N> + Copy + Serialize, 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, 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 { 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) ) } fn default_config(&self) -> Configuration<F> { let default_alg = match self.data.dataterm { DataTerm::L2Squared => DefaultAlgorithm::FB.get_named(), DataTerm::L1 => DefaultAlgorithm::PDPS.get_named(), }; Configuration{ algorithms : vec![default_alg], iterator_options : AlgIteratorOptions{ max_iter : 2000, verbose_iter : Verbose::Logarithmic(10), quiet : false, }, plot : PlotLevel::Data, outdir : "out".to_string(), bt_depth : DynamicDepth(8), } } fn runall(&self, config : Configuration<F>) -> DynError { let &Named { name : ref experiment_name, data : Experiment { domain, sensor_count, ref noise_distr, sensor, spread, kernel, ref μ_hat, α, kernel_plot_width, dataterm, noise_seed, .. } } = self; // Set path let prefix = format!("{}/{}/", config.outdir, experiment_name); // Set up operators let depth = config.bt_depth; let opA = DefaultSG::new(domain, sensor_count, sensor, spread, depth); let op𝒟 = DefaultSeminormOp::new(depth, domain, kernel); // Set up random number generator. let mut rng = StdRng::seed_from_u64(noise_seed); // Generate the data and calculate SSNR statistic let b_hat = opA.apply(μ_hat); 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); // 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"), &config)?; write_json(mkname_e("stats"), &stats)?; plotall(&config, &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 config.algorithms.iter() { let this_prefix = format!("{}{}/", prefix, alg_name); let running = || { println!("{}\n{}\n{}", format!("Running {} on experiment {}…", alg_name, experiment_name).cyan(), format!("{:?}", config.iterator_options).bright_black(), format!("{:?}", alg).bright_black()); }; // Create Logger and IteratorFactory let mut logger = Logger::new(); let findim_data = prepare_optimise_weights(&opA); 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 { None => value, Some(mut μ) => { match dataterm { DataTerm::L2Squared => { optimise_weights( &mut μ, &opA, &b, α, &findim_data, &inner_config, inner_it ); dataterm.value_at_residual(opA.apply(&μ) - &b) + α * μ.norm(Radon) }, _ => value, } } }; CSVLog { iter, value, post_value, n_spikes, cpu_time : cpu_time.as_secs_f64(), inner_iters, merged, pruned, this_iters } }; let iterator = config.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 config.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(); let μ : DiscreteMeasure<Loc<F, N>, F> = match (alg, dataterm) { (AlgorithmConfig::FB(ref algconfig), DataTerm::L2Squared) => { running(); pointsource_fb(&opA, &b, α, &op𝒟, &algconfig, iterator, plotter) }, (AlgorithmConfig::FW(ref algconfig), DataTerm::L2Squared) => { running(); pointsource_fw(&opA, &b, α, &algconfig, iterator, plotter) }, (AlgorithmConfig::PDPS(ref algconfig), DataTerm::L2Squared) => { running(); pointsource_pdps(&opA, &b, α, &op𝒟, &algconfig, iterator, plotter, L2Squared) }, (AlgorithmConfig::PDPS(ref algconfig), DataTerm::L1) => { running(); pointsource_pdps(&opA, &b, α, &op𝒟, &algconfig, iterator, plotter, L1) }, _ => { let msg = format!("Algorithm “{}” not implemented for dataterm {:?}. Skipping.", alg_name, dataterm).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!("{p}{n}_{t}", p = prefix, n = alg_name, t = 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(()) } } /// Plot experiment setup #[replace_float_literals(F::cast_from(literal))] fn plotall<F, Sensor, Kernel, Spread, 𝒟, A, const N : usize>( config : &Configuration<F>, prefix : &String, domain : &Cube<F, N>, sensor : &Sensor, kernel : &Kernel, spread : &Spread, μ_hat : &DiscreteMeasure<Loc<F, N>, F>, op𝒟 : &𝒟, opA : &A, b_hat : &A::Observable, b : &A::Observable, kernel_plot_width : F, ) -> DynError where F : Float + ToNalgebraRealField, 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>, 𝒟 : DiscreteMeasureOp<Loc<F, N>, F>, 𝒟::Codomain : RealMapping<F, N>, A : ForwardModel<Loc<F, N>, F>, A::PreadjointCodomain : RealMapping<F, N> + Bounded<F>, PlotLookup : Plotting<N>, Cube<F, N> : SetOrd { if config.plot < PlotLevel::Data { return Ok(()) } let base = Convolution(sensor.clone(), spread.clone()); let resolution = if N==1 { 100 } else { 40 }; 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()); 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()); 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, pfx("omega_b") ); // 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")) }