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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455
//! Spaces: runtime-defined types
//!
//! In addition to the spaces defined here,
//! a product space can be derived on structures containing inner spaces with
//! [`#[derive(ProductSpace)]`](ProductSpace).
#[cfg(test)]
#[macro_use]
pub mod testing;
mod array;
mod boolean;
mod index;
mod indexed_type;
mod interval;
mod ndarray;
mod nonempty_features;
mod option;
mod power;
mod singleton;
#[cfg(test)]
mod test_derive;
mod tuple;
mod wrapper;
pub use self::ndarray::{Array1Space, Array2Space, Array3Space, NdArraySpace};
pub use array::ArraySpace;
pub use boolean::BooleanSpace;
pub use index::IndexSpace;
pub use indexed_type::{Indexed, IndexedTypeSpace};
pub use interval::IntervalSpace;
pub use nonempty_features::NonEmptyFeatures;
pub use option::OptionSpace;
pub use power::PowerSpace;
pub use singleton::SingletonSpace;
pub use tuple::{TupleSpace2, TupleSpace3, TupleSpace4, TupleSpace5};
pub use wrapper::BoxSpace;
// Re-export space macros from relearn_derive
pub use relearn_derive::{
FiniteSpace, Indexed, LogElementSpace, ProductSpace, SampleSpace, Space, SubsetOrd,
};
use crate::logging::{LogError, StatsLogger};
use crate::utils::distributions::ArrayDistribution;
use crate::utils::num_array::{BuildFromArray1D, BuildFromArray2D, NumArray1D, NumArray2D};
use ::ndarray::{ArrayBase, DataMut, Ix2};
use num_traits::Float;
use rand::distributions::Distribution;
use rand::RngCore;
use std::cmp::Ordering;
use std::iter::ExactSizeIterator;
/// A space: a set of values with some added structure.
///
/// A space is effectively a runtime-defined type.
pub trait Space {
// It is awkward to constrain associated types in sub-traits so apply core constraints here.
type Element: Clone + Send;
/// Check whether a particular value is contained in the space.
fn contains(&self, value: &Self::Element) -> bool;
}
/// Implement `Space` for a deref-able wrapper type generic over `S: Space + ?Sized`.
macro_rules! impl_wrapped_space {
($wrapper:ty) => {
impl<S> Space for $wrapper
where
S: Space + ?Sized,
{
type Element = S::Element;
#[inline]
fn contains(&self, value: &Self::Element) -> bool {
S::contains(self, value)
}
}
};
}
impl_wrapped_space!(&'_ S);
impl_wrapped_space!(Box<S>);
/// Compare this space to another in terms of the subset relation.
///
/// This is a partial order and the rules for implementing this are the same as for
/// [`PartialOrd`](std::cmp::PartialOrd). In particular,
/// the comparision must return `Some(Ordering::Equal)` if and only if `self == other`.
///
/// This is distinct from [`PartialOrd`](std::cmp::PartialOrd) so that `SubsetOrd` can be defined
/// on types that already implement `PartialOrd` in a different way (e.g. lexicographically).
/// It also avoids the confusion that might arise from using comparison operators (`<`, `>`, etc.)
/// since it is not obvious that "subset" is the relationship being used.
pub trait SubsetOrd: PartialEq<Self> {
/// Compare using the subset relationship. This is a partial order.
fn subset_cmp(&self, other: &Self) -> Option<Ordering>;
/// Check if this is a strict subset of `other`.
#[inline]
fn strict_subset_of(&self, other: &Self) -> bool {
matches!(self.subset_cmp(other), Some(Ordering::Less))
}
/// Check if this is a subset (strict or equal) of `other`.
#[inline]
fn subset_of(&self, other: &Self) -> bool {
matches!(
self.subset_cmp(other),
Some(Ordering::Less | Ordering::Equal)
)
}
/// Check if this is a strict superset of `other`.
#[inline]
fn strict_superset_of(&self, other: &Self) -> bool {
matches!(self.subset_cmp(other), Some(Ordering::Greater))
}
/// Check if this is a superset (strict or equal) of `other`.
#[inline]
fn superset_of(&self, other: &Self) -> bool {
matches!(
self.subset_cmp(other),
Some(Ordering::Greater | Ordering::Equal)
)
}
}
/// Implement `SubsetOrd` for a deref-able wrapper type generic over `T: SubsetOrd + ?Sized`.
macro_rules! impl_wrapped_subset_ord {
($wrapper:ty) => {
impl<S> SubsetOrd for $wrapper
where
S: SubsetOrd + ?Sized,
{
#[inline]
fn subset_cmp(&self, other: &Self) -> Option<Ordering> {
S::subset_cmp(self, other)
}
}
};
}
impl_wrapped_subset_ord!(&'_ S);
impl_wrapped_subset_ord!(Box<S>);
/// Helper function to determine the subset ordering of a product of two spaces.
///
/// Given the orderings for each of the factors, the ordering is:
/// * `Equal` if both factors are `Equal`,
/// * `Less` if both factors are `Equal` or `Less` and at least one is `Less`,
/// * `Greater` if both factors are `Equal` or `Greater` and at least one is `Greater`,
/// * `None` otherwise.
#[must_use]
#[inline]
pub const fn product_subset_ord(a: Ordering, b: Option<Ordering>) -> Option<Ordering> {
use Ordering::*;
match (a, b) {
(Equal, Some(x)) | (x, Some(Equal)) => Some(x),
(Less, Some(Less)) => Some(Less),
(Greater, Some(Greater)) => Some(Greater),
_ => None,
}
}
/// Helper function to determine the subset ordering of a product space with any number of factors.
///
/// Given the orderings for each of the factors, the ordering is:
/// * `Equal` if all factors are `Equal`,
/// * `Less` if all factors are `Equal` or `Less` and at least one is `Less`,
/// * `Greater` if all factors are `Equal` or `Greater` and at least one is `Greater`,
/// * `None` otherwise.
#[inline]
pub fn iter_product_subset_ord<I: IntoIterator<Item = Option<Ordering>>>(
ord_factors: I,
) -> Option<Ordering> {
ord_factors
.into_iter()
.try_fold(Ordering::Equal, product_subset_ord)
}
/// A space containing finitely many elements.
pub trait FiniteSpace: Space {
/// The number of elements in the space.
fn size(&self) -> usize;
/// Get the (unique) index of an element.
fn to_index(&self, element: &Self::Element) -> usize;
/// Try to convert an index to an element.
///
/// The return value is `Some(elem)` if and only if
/// `elem` is the unique element in the space with `to_index(elem) == index`.
#[allow(clippy::wrong_self_convention)] // `from_` here refers to element, not space
fn from_index(&self, index: usize) -> Option<Self::Element>;
/// Try to convert an index to an element.
///
/// If `None` is returned then the index was invalid.
/// `Some(_)` may be returned even if the index is invalid.
/// If the returned value must be validated then use [`FiniteSpace::from_index`].
#[inline]
#[allow(clippy::wrong_self_convention)] // `from_` here refers to element, not space
fn from_index_unchecked(&self, index: usize) -> Option<Self::Element> {
self.from_index(index)
}
}
/// Implement `FiniteSpace` for a deref-able wrapper type generic over `S: FiniteSpace + ?Sized`.
macro_rules! impl_wrapped_finite_space {
($wrapper:ty) => {
impl<S> FiniteSpace for $wrapper
where
S: FiniteSpace + ?Sized,
{
#[inline]
fn size(&self) -> usize {
S::size(self)
}
#[inline]
fn to_index(&self, element: &Self::Element) -> usize {
S::to_index(self, element)
}
#[inline]
fn from_index(&self, index: usize) -> Option<Self::Element> {
S::from_index(self, index)
}
#[inline]
fn from_index_unchecked(&self, index: usize) -> Option<Self::Element> {
S::from_index_unchecked(self, index)
}
}
};
}
impl_wrapped_finite_space!(&'_ S);
impl_wrapped_finite_space!(Box<S>);
/// A space containing at least one element.
pub trait NonEmptySpace: Space {
/// An arbitrary deterministic element from the space.
fn some_element(&self) -> Self::Element;
}
/// Implement `NonEmptySpace` for a deref-able wrapper type generic on `S: NonEmptySpace + ?Sized`.
macro_rules! impl_wrapped_non_empty_space {
($wrapper:ty) => {
impl<S> NonEmptySpace for $wrapper
where
S: NonEmptySpace + ?Sized,
{
#[inline]
fn some_element(&self) -> Self::Element {
S::some_element(self)
}
}
};
}
impl_wrapped_non_empty_space!(&'_ S);
impl_wrapped_non_empty_space!(Box<S>);
/// A space from which samples can be drawn.
///
/// No particular distribution is specified but the distribution:
/// * must have support equal to the entire space, and
/// * should be some form of reasonable "standard" distribution for the space.
///
/// # Note
/// This re-implements sample method of [`Distribution`] rather than set
/// `Distribution<Self::Element>` as a super-trait so that `SampleSpace` is object-safe since
/// * `Distribution<T>` is not object-safe, and even if it was,
/// * generic super traits using `<Self::AssocType>` are not object safe due to a bug / issue:
/// <https://github.com/rust-lang/rust/issues/40533>.
pub trait SampleSpace: NonEmptySpace {
/// Sample a random element.
fn sample(&self, rng: &mut dyn RngCore) -> Self::Element;
}
impl<S> SampleSpace for S
where
S: NonEmptySpace + Distribution<<Self as Space>::Element>,
{
#[inline]
fn sample(&self, rng: &mut dyn RngCore) -> Self::Element {
Distribution::sample(&self, rng)
}
}
/// A space whose elements can be represented as value of type `T`
///
/// This representation is generally minimal, in contrast to [`FeatureSpace`],
/// which produces a representation suited for use as input to a machine learning model.
pub trait ReprSpace<T, T0 = T>: Space {
/// Representation of a single element.
fn repr(&self, element: &Self::Element) -> T0;
/// Represent a batch of elements as an array.
fn batch_repr<'a, I>(&self, elements: I) -> T
where
I: IntoIterator<Item = &'a Self::Element>,
I::IntoIter: ExactSizeIterator + Clone,
Self::Element: 'a;
}
/// Implement `ReprSpace<T, T0>` for a deref-able wrapper type generic over `S`.
macro_rules! impl_wrapped_repr_space {
($wrapper:ty) => {
impl<S, T, T0> ReprSpace<T, T0> for $wrapper
where
S: ReprSpace<T, T0> + ?Sized,
{
#[inline]
fn repr(&self, element: &Self::Element) -> T0 {
S::repr(self, element)
}
#[inline]
fn batch_repr<'a, I>(&self, elements: I) -> T
where
I: IntoIterator<Item = &'a Self::Element>,
I::IntoIter: ExactSizeIterator + Clone,
Self::Element: 'a,
{
S::batch_repr(self, elements)
}
}
};
}
impl_wrapped_repr_space!(&'_ S);
impl_wrapped_repr_space!(Box<S>);
/// A space whose elements can be encoded as floating-point feature vectors.
pub trait FeatureSpace: Space {
/// Length of the encoded feature vectors.
fn num_features(&self) -> usize;
/// Encode the feature vector of an element into a mutable slice.
///
/// # Args
/// * `element` - The element to encode.
/// * `out` - A slice of length at least `num_features()` in which the features are written.
/// Only the first `num_features()` values are written to.
/// * `zeroed` - Whether `out` is zero-initialized.
/// Helps avoid redundant writes for sparse feature vectors.
///
/// # Returns
/// A reference to the remainder of out: `&mut out[num_features()..]`.
///
/// # Panics
/// If the slice is not large enough to fit the feature vector.
fn features_out<'a, F: Float>(
&self,
element: &Self::Element,
out: &'a mut [F],
zeroed: bool,
) -> &'a mut [F];
/// Encode the feature vector of an element into an array.
#[inline]
fn features<T>(&self, element: &Self::Element) -> T
where
T: BuildFromArray1D,
<T::Array as NumArray1D>::Elem: Float,
{
let mut array = T::Array::zeros(self.num_features());
self.features_out(element, array.as_slice_mut(), true);
array.into()
}
/// Encode the feature vectors of multiple elements into rows of a two-dimensional array.
///
/// # Args
/// * `elements` - Elements to encode.
/// * `out` - A two-dimensional array of shape at least `[elements.len(), num_features()]`.
/// Only the left `[.., 0..num_features()]` subarray may be written to.
/// * `zeroed` - Whether `out` is zero-initialized.
/// Helps avoid redundant writes for sparse feature vectors.
///
/// # Panics
/// If the array is not large enough to fit the feature vectors.
#[inline]
fn batch_features_out<'a, I, A>(&self, elements: I, out: &mut ArrayBase<A, Ix2>, zeroed: bool)
where
I: IntoIterator<Item = &'a Self::Element>,
Self::Element: 'a,
A: DataMut,
A::Elem: Float,
{
// Don't zip rows so that we can check whether there are too few rows.
let mut rows = out.rows_mut().into_iter();
for element in elements {
let mut row = rows.next().expect("fewer rows than elements");
self.features_out(
element,
row.as_slice_mut().expect("could not view row as slice"),
zeroed,
);
}
}
/// Encode the feature vectors of multiple elements as rows of a two-dimensional array.
#[inline]
fn batch_features<'a, I, T>(&self, elements: I) -> T
where
I: IntoIterator<Item = &'a Self::Element>,
I::IntoIter: ExactSizeIterator,
Self::Element: 'a,
T: BuildFromArray2D,
<T::Array as NumArray2D>::Elem: Float,
{
let elements = elements.into_iter();
let mut array = T::Array::zeros((elements.len(), self.num_features()));
self.batch_features_out(elements, &mut array.view_mut(), true);
array.into()
}
}
/// A space whose elements parameterize a distribution
pub trait ParameterizedDistributionSpace<T, T2 = T>: ReprSpace<T, T2> {
/// Batched distribution type.
///
/// The element representation must match the format of [`ReprSpace`].
/// That is, `batch_repr(&[...])` must be a valid input for [`ArrayDistribution::log_probs`].
type Distribution: ArrayDistribution<T, T>;
/// Size of the parameter vector for which elements are sampled.
fn num_distribution_params(&self) -> usize;
// TODO Take Prng?
/// Sample a single element given a parameter vector.
///
/// # Args
/// * `params` - A one-dimensional parameter vector of length `self.num_distribution_params()`.
///
/// # Panics
/// Panics if `params` does not have the correct shape.
fn sample_element(&self, params: &T) -> Self::Element;
/// The distribution parameterized by the given parameter vector.
///
/// # Args
/// * `params` - Batched parameter vectors.
/// An array with shape `[BATCH_SIZE.., self.num_distribution_params()]`.
///
/// # Returns
/// The distribution(s) parameterized by `params`.
fn distribution(&self, params: &T2) -> Self::Distribution;
}
/// A space whose elements can be logged to a [`StatsLogger`]
pub trait LogElementSpace: Space {
/// Log an element of the space
fn log_element<L: StatsLogger + ?Sized>(
&self,
name: &'static str,
element: &Self::Element,
logger: &mut L,
) -> Result<(), LogError>;
}