1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407
//! This crate provides a hyperparameter optimization algorithm using TPE (Tree-structured Parzen Estimator).
//!
//! # Examples
//!
//! An example optimizing a simple quadratic function which has one numerical and one categorical parameters.
//! ```
//! use rand::SeedableRng as _;
//!
//! # fn main() -> anyhow::Result<()> {
//! let choices = [1, 10, 100];
//! let mut optim0 =
//! tpe::TpeOptimizer::new(tpe::parzen_estimator(), tpe::range(-5.0, 5.0)?);
//! let mut optim1 =
//! tpe::TpeOptimizer::new(tpe::histogram_estimator(), tpe::categorical_range(choices.len())?);
//!
//! fn objective(x: f64, y: i32) -> f64 {
//! x.powi(2) + y as f64
//! }
//!
//! let mut best_value = std::f64::INFINITY;
//! let mut rng = rand::rngs::StdRng::from_seed(Default::default());
//! for _ in 0..100 {
//! let x = optim0.ask(&mut rng)?;
//! let y = optim1.ask(&mut rng)?;
//!
//! let v = objective(x, choices[y as usize]);
//! optim0.tell(x, v)?;
//! optim1.tell(y, v)?;
//! best_value = best_value.min(v);
//! }
//!
//! assert_eq!(best_value, 1.000054276671888);
//! # Ok(())
//! # }
//! ```
//!
//! # References
//!
//! Please refer to the following papers about the details of TPE:
//!
//! - [Algorithms for Hyper-Parameter Optimization](https://papers.nips.cc/paper/4443-algorithms-for-hyper-parameter-optimization.pdf)
//! - [Making a Science of Model Search: Hyperparameter Optimization in Hundreds of Dimensions for Vision Architectures](http://proceedings.mlr.press/v28/bergstra13.pdf)
#![warn(missing_docs)]
use crate::density_estimation::{BuildDensityEstimator, DefaultEstimatorBuilder, DensityEstimator};
#[cfg(doc)]
use crate::density_estimation::{HistogramEstimator, ParzenEstimator};
use crate::range::{Range, RangeError};
use ordered_float::OrderedFloat;
use rand::distributions::Distribution;
use rand::Rng;
use std::num::NonZeroUsize;
pub mod density_estimation;
pub mod range;
/// Creates a [`Range`] instance.
pub fn range(start: f64, end: f64) -> Result<Range, RangeError> {
Range::new(start, end)
}
/// Creates a [`Range`] for a categorical parameter.
///
/// This is equivalent to `range(0.0, cardinality as f64)`.
pub fn categorical_range(cardinality: usize) -> Result<Range, RangeError> {
Range::new(0.0, cardinality as f64)
}
/// Creates a [`DefaultEstimatorBuilder`] to build [`ParzenEstimator`] (for categorical parameter).
pub fn parzen_estimator() -> DefaultEstimatorBuilder {
DefaultEstimatorBuilder::Parzen(Default::default())
}
/// Creates a [`DefaultEstimatorBuilder`] to build [`HistogramEstimator`] (for numerical parameter).
pub fn histogram_estimator() -> DefaultEstimatorBuilder {
DefaultEstimatorBuilder::Histogram(Default::default())
}
/// Builder of [`TpeOptimizer`].
#[derive(Debug)]
pub struct TpeOptimizerBuilder {
gamma: f64,
candidates: usize,
}
impl TpeOptimizerBuilder {
/// Makes a new [`TpeOptimizerBuilder`] instance with the default settings.
pub fn new() -> Self {
Self::default()
}
/// Sets the percentage at which the good and bad observations are split.
///
/// The default values is `0.1`.
pub fn gamma(&mut self, gamma: f64) -> &mut Self {
self.gamma = gamma;
self
}
/// Sets the number of candidates to be sampled to decide the next parameter.
///
/// The default value is `24`.
pub fn candidates(&mut self, candidates: usize) -> &mut Self {
self.candidates = candidates;
self
}
/// Builds a [`TpeOptimizer`] with the given settings.
pub fn build<T>(
&self,
estimator_builder: T,
param_range: Range,
) -> Result<TpeOptimizer<T>, BuildError>
where
T: BuildDensityEstimator,
{
if !(0.0 <= self.gamma && self.gamma <= 1.0) {
return Err(BuildError::GammaOutOfRange);
}
Ok(TpeOptimizer {
param_range,
estimator_builder,
trials: Vec::new(),
is_sorted: false,
gamma: self.gamma,
candidates: NonZeroUsize::new(self.candidates).ok_or(BuildError::ZeroCandidates)?,
})
}
}
impl Default for TpeOptimizerBuilder {
fn default() -> Self {
Self {
gamma: 0.1,
candidates: 24,
}
}
}
/// Optimizer using TPE.
///
/// This try to search out the parameter value which could minimize the evaluation result.
///
/// Note that an instance of TpeOptimizer can handle only one hyperparameter.
/// So if you want to optimize multiple hyperparameters simultaneously,
/// please create an optimizer for each hyperparameter.
#[derive(Debug)]
pub struct TpeOptimizer<T = DefaultEstimatorBuilder> {
param_range: Range,
estimator_builder: T,
trials: Vec<Trial>,
is_sorted: bool,
gamma: f64,
candidates: NonZeroUsize,
}
impl<T: BuildDensityEstimator> TpeOptimizer<T> {
/// Makes a new [`TpeOptimizer`] with the default settings.
///
/// If you want to customize the settings, please use [`TpeOptimizerBuilder`] instead.
pub fn new(estimator_builder: T, param_range: Range) -> Self {
TpeOptimizerBuilder::new()
.build(estimator_builder, param_range)
.expect("unreachable")
}
/// Returns the next value of the optimization target parameter to be evaluated.
///
/// Note that, before the first asking, it might be worth to give some evaluation
/// results of randomly sampled observations to [`TpeOptimizer`] (via the [`tell`](TpeOptimizer::tell) method)
/// to reduce bias due to too few samples.
pub fn ask<R: Rng + ?Sized>(&mut self, rng: &mut R) -> Result<f64, T::Error> {
if !self.is_sorted {
self.trials.sort_by_key(|t| OrderedFloat(t.value));
self.is_sorted = true;
}
let split_point = self.decide_split_point();
let (superiors, inferiors) = self.trials.split_at(split_point);
let superior_estimator = self.estimator_builder.build_density_estimator(
superiors.iter().map(|t| t.param).filter(|p| p.is_finite()),
self.param_range,
)?;
let inferior_estimator = self.estimator_builder.build_density_estimator(
inferiors.iter().map(|t| t.param).filter(|p| p.is_finite()),
self.param_range,
)?;
let param = (&superior_estimator)
.sample_iter(rng)
.take(self.candidates.get())
.map(|candidate| {
let superior_log_likelihood = superior_estimator.log_pdf(candidate);
let inferior_log_likelihood = inferior_estimator.log_pdf(candidate);
let ei = superior_log_likelihood - inferior_log_likelihood;
(ei, candidate)
})
.max_by_key(|(ei, _)| OrderedFloat(*ei))
.map(|(_, param)| param)
.expect("unreachable");
Ok(param)
}
/// Tells the evaluation result of a hyperparameter value to the optimizer.
///
/// Note that the `param` should be NaN if the hyperparameter was not used in the evaluation
/// (this could be happen when the entire search space is conditional).
pub fn tell(&mut self, param: f64, value: f64) -> Result<(), TellError> {
if value.is_nan() {
return Err(TellError::NanValue);
}
if !(param.is_nan() || self.param_range.contains(param)) {
return Err(TellError::ParamOutOfRange {
param,
range: self.param_range,
});
}
self.trials.push(Trial { param, value });
self.is_sorted = false;
Ok(())
}
/// Retruns all told parameter and objective values.
///
/// Note that the order of items in the returned iterator doesn't reflect the order [`TpeOptimizer::tell`] called.
///
/// # Examples
///
/// You can this method to store and load the state of a [`TpeOptimizer`].
///
/// ```
/// # fn main() -> anyhow::Result<()> {
/// let temp_file = tempfile::NamedTempFile::new()?;
/// let mut rng = rand::thread_rng();
///
/// fn objective(x: f64) -> f64 {
/// x.powi(2)
/// }
///
/// {
/// let mut optim =
/// tpe::TpeOptimizer::new(tpe::parzen_estimator(), tpe::range(-5.0, 5.0)?);
///
/// // Runs 100 trials.
/// for _ in 0..100 {
/// let x = optim.ask(&mut rng)?;
/// let v = objective(x);
/// optim.tell(x, v)?;
/// }
///
/// // Stores the trials.
/// let mut file = std::fs::OpenOptions::new()
/// .write(true)
/// .open(temp_file.path())?;
/// serde_json::to_writer(&mut file, &optim.trials().collect::<Vec<_>>())?;
/// }
///
/// {
/// let mut optim =
/// tpe::TpeOptimizer::new(tpe::parzen_estimator(), tpe::range(-5.0, 5.0)?);
///
/// // Loads the previous trials.
/// let mut file = std::fs::File::open(temp_file.path())?;
/// let trials: Vec<(f64, f64)> = serde_json::from_reader(&mut file)?;
/// for (x, v) in trials {
/// optim.tell(x, v)?;
/// }
///
/// // Runs additional 100 trials.
/// for _ in 0..100 {
/// let x = optim.ask(&mut rng)?;
/// let v = objective(x);
/// optim.tell(x, v)?;
/// }
/// assert_eq!(optim.trials().count(), 200);
/// }
/// # Ok(())
/// # }
/// ```
pub fn trials(&self) -> impl '_ + Iterator<Item = (f64, f64)> {
self.trials.iter().map(|t| (t.param, t.value))
}
fn decide_split_point(&self) -> usize {
(self.trials.len() as f64 * self.gamma).ceil() as usize
}
}
#[derive(Debug, Clone)]
struct Trial {
param: f64,
value: f64,
}
/// Possible errors during [`TpeOptimizerBuilder::build`].
#[derive(Debug, Clone, thiserror::Error)]
pub enum BuildError {
#[error("the value of `gamma` must be in the range from 0.0 to 1.0")]
/// The value of `gamma` must be in the range from `0.0` to `1.0`.
GammaOutOfRange,
#[error("the value of `candidates` must be a positive integer")]
/// The value of `candidates` must be a positive integer.
ZeroCandidates,
}
/// Possible errors during [`TpeOptimizer::tell`].
#[derive(Debug, Clone, thiserror::Error)]
pub enum TellError {
#[error("the parameter value {param} is out of the range {range}")]
/// The parameter value is out of the range.
ParamOutOfRange {
/// Actual parameter value.
param: f64,
/// Expected parameter range.
range: Range,
},
#[error("NaN value is not allowed")]
/// NaN value is not allowed.
NanValue,
}
#[cfg(test)]
mod tests {
use super::*;
use rand::SeedableRng;
#[test]
fn optimizer_works() -> anyhow::Result<()> {
let choices = [1, 10, 100];
let mut optim0 = TpeOptimizer::new(parzen_estimator(), range(-5.0, 5.0)?);
let mut optim1 =
TpeOptimizer::new(histogram_estimator(), categorical_range(choices.len())?);
fn objective(x: f64, y: i32) -> f64 {
x.powi(2) + y as f64
}
let mut best_value = std::f64::INFINITY;
let mut rng = rand::rngs::StdRng::from_seed(Default::default());
for _ in 0..100 {
let x = optim0.ask(&mut rng)?;
let y = optim1.ask(&mut rng)?;
let v = objective(x, choices[y as usize]);
optim0.tell(x, v)?;
optim1.tell(y, v)?;
best_value = best_value.min(v);
}
assert_eq!(best_value, 1.000054276671888);
Ok(())
}
#[test]
fn store_and_save_trials_worls() -> anyhow::Result<()> {
let temp_file = tempfile::NamedTempFile::new()?;
let mut rng = rand::thread_rng();
fn objective(x: f64) -> f64 {
x.powi(2)
}
{
let mut optim = TpeOptimizer::new(parzen_estimator(), range(-5.0, 5.0)?);
// Runs 100 trials.
for _ in 0..100 {
let x = optim.ask(&mut rng)?;
let v = objective(x);
optim.tell(x, v)?;
}
// Stores the trials.
let mut file = std::fs::OpenOptions::new()
.write(true)
.open(temp_file.path())?;
serde_json::to_writer(&mut file, &optim.trials().collect::<Vec<_>>())?;
}
{
let mut optim = TpeOptimizer::new(parzen_estimator(), range(-5.0, 5.0)?);
// Loads the previous trials.
let mut file = std::fs::File::open(temp_file.path())?;
let trials: Vec<(f64, f64)> = serde_json::from_reader(&mut file)?;
for (x, v) in trials {
optim.tell(x, v)?;
}
// Runs additional 100 trials.
for _ in 0..100 {
let x = optim.ask(&mut rng)?;
let v = objective(x);
optim.tell(x, v)?;
}
assert_eq!(optim.trials().count(), 200);
}
Ok(())
}
}