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
mod node;
mod var;
mod vardiff;
use ndarray::{ArrayViewMutD, Dimension, Ix, RawArrayViewMut};
use std::{
cell::{Ref, RefCell},
collections::{BTreeMap, HashSet},
hash::{Hash, Hasher},
rc::Rc,
};
pub use var::Var;
pub use vardiff::VarDiff;
pub(crate) use node::*;
pub use node::{
Backward, Constant, Convolve, ConvolveWithGroups, Data, Eval, Forward, Gradient, Input,
InputBackward, Overwrite, PaddingMode, Reflective, Replicative, Zero,
};
// ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
// ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ Global Var Identifier ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
// ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
/// Keeps track of each operations. It is also used to provide an identifier to computational nodes.
pub(crate) struct OperationsCounter {
count: usize,
}
impl OperationsCounter {
pub fn next(&mut self) -> usize {
self.count += 1;
self.count
}
}
pub(crate) static mut OPERATIONS_COUNTER: OperationsCounter = OperationsCounter { count: 0 };
// ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
// ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ Histories ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
// ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
#[derive(Clone)]
/// The computational forward-history of a variable. It keeps track of the computation up to the
/// variable to whom the struct belongs.
pub struct VarHistory {
path: BTreeMap<usize, Rc<dyn Forward>>,
buffer: RefCell<Vec<Rc<dyn Forward>>>,
changeables: HashSet<Changeable>,
}
impl VarHistory {
/// Returns a new, empty, `VarHistory`.
pub(crate) fn new() -> Self {
Self {
path: BTreeMap::new(),
buffer: RefCell::new(Vec::new()),
changeables: HashSet::new(),
}
}
/// Merges `self` and `other`. This is equivalent to a set-intersection.
///
/// # Arguments
///
/// `other` - other VarHistory.
pub(crate) fn merge(&mut self, mut other: VarHistory) {
self.path.append(&mut other.path);
}
/// Appends a new forward computational node to `self`. The new node has id `id`.
///
/// # Arguments
///
/// * `id` - id of the new node.
/// * `next` - node to append.
pub(crate) fn append_forward(&mut self, id: usize, next: Rc<dyn Forward>) {
self.path.insert(id, next);
self.buffer.borrow_mut().truncate(0);
}
/// Appends a new eval computational node to `self`. The new node has id `id`.
///
/// # Arguments
///
/// * `next` - node to append.
pub(crate) fn append_changeable(&mut self, next: Changeable) {
self.changeables.insert(next);
}
/// Returns the length of the forward path.
pub(crate) fn len(&self) -> usize {
self.path.len()
}
/// Returns `true` if the forward path contains no node.
pub(crate) fn is_empty(&self) -> bool {
self.path.is_empty()
}
/// Prepares the buffer. Clones and transfers the content of the forward path
/// into a vector. Such vector will be used to perform the actual forward pass.
pub(crate) fn prepare_buffer(&self) {
if self.buffer.borrow().is_empty() {
*self.buffer.borrow_mut() = self.path.values().cloned().collect();
}
}
/// Returns a reference to the buffer.
pub(crate) fn buffer(&self) -> Ref<[Rc<dyn Forward>]> {
Ref::map(self.buffer.borrow(), |vec| &vec[..])
}
}
#[derive(Clone)]
/// The computational backward-history of a variable. It keeps track of the computation up to the
/// variable to whom the struct belongs.
pub struct VarDiffHistory {
path: BTreeMap<usize, Rc<dyn Backward>>,
buffer: RefCell<Vec<Rc<dyn Backward>>>,
parameters: HashSet<RawParam>,
}
impl VarDiffHistory {
/// Returns a new, empty, `VarDiffHistory` with parameters `parameters`.
///
/// # Arguments
///
/// ` parameters` - parameters to store.
pub(crate) fn new(parameters: HashSet<RawParam>) -> Self {
Self {
path: BTreeMap::new(),
buffer: RefCell::new(Vec::new()),
parameters,
}
}
/// Merges `self` and `other`. This is equivalent to a set-intersection.
///
/// # Arguments
///
/// `other` - other VarDiffHistory.
pub(crate) fn merge(&mut self, mut other: VarDiffHistory) {
self.path.append(&mut other.path);
self.parameters.extend(other.parameters);
}
/// Appends a new backward computational node to `self`. The new node has id `id`.
///
/// # Arguments
///
/// * `id` - id of the new node.
/// * `next` - node to append.
pub(crate) fn append_backward(&mut self, id: usize, next: Rc<dyn Backward>) {
self.path.insert(id, next);
self.buffer.borrow_mut().truncate(0);
}
/// Returns the length of the backward path.
pub(crate) fn len(&self) -> usize {
self.path.len()
}
/// Returns `true` if the backward path contains no node.
pub(crate) fn is_empty(&self) -> bool {
self.path.is_empty()
}
/// Prepares the buffer. Clones and transfers the content of the backward path
/// into a vector. Such vector will be used to perform the actual backward pass.
pub(crate) fn prepare_buffer(&self) {
if self.buffer.borrow().is_empty() {
*self.buffer.borrow_mut() = self.path.values().cloned().collect();
}
}
/// Returns a reference to the buffer.
pub(crate) fn buffer(&self) -> Ref<[Rc<dyn Backward>]> {
Ref::map(self.buffer.borrow(), |vec| &vec[..])
}
}
// ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
// ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ RawParam Struct ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
// ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
/// A builder of mutable views over a differentiable variable's data and gradient.
#[derive(Clone, PartialEq, Eq, Hash)]
pub struct RawParam {
data: *mut f32,
grad: *mut f32,
shape: Vec<Ix>,
}
impl RawParam {
pub(crate) fn new(data: *mut f32, grad: *mut f32, shape: Vec<Ix>) -> Self {
Self { data, grad, shape }
}
/// Consumes the RawParam, yielding mutable views over the data and the gradient of the
/// differentiable variable that it refers to. The lifetime `'a` is for the
/// scope of the borrow.
pub(crate) fn into_param<'a>(self) -> Param<'a> {
let shape = self.shape;
unsafe {
let raw_data = RawArrayViewMut::from_shape_ptr(shape.clone(), self.data);
let raw_grad = RawArrayViewMut::from_shape_ptr(shape, self.grad);
let data = raw_data.deref_into_view_mut();
let grad = raw_grad.deref_into_view_mut();
Param { data, grad }
}
}
}
// ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
// ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ Param Struct ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
// ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
/// Mutable views over a differentiable variable's data and gradient.
///
/// See also [`.parameters()`] and [`ModelStatus`] for more details.
///
///
/// The views are [`ndarray::ArrayViewMutD`].
///
/// [`ndarray::ArrayViewMutD`]: ndarray::ArrayViewMutD
///
/// [`.parameters()`]: VarDiff::parameters()
/// [`ModelStatus`]: crate::nn::ModelStatus
#[derive(Debug)]
pub struct Param<'a> {
pub data: ArrayViewMutD<'a, f32>,
pub grad: ArrayViewMutD<'a, f32>,
}
// ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
// ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ Changeable struct ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
// ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
#[derive(Clone)]
/// Hashable and comparable wrapper for a computational node that implements the `Eval` trait.
pub(super) struct Changeable {
id: usize,
node: Rc<dyn Eval>,
}
impl PartialEq for Changeable {
fn eq(&self, other: &Self) -> bool {
self.id == other.id
}
}
impl Eq for Changeable {}
impl Hash for Changeable {
fn hash<H: Hasher>(&self, state: &mut H) {
self.id.hash(state);
}
}
// ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
// ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ Algebraic Traits ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
// ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
// ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ Matrix Multiplication ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
/// Matrix-matrix multiplication.
pub trait MatMatMul<Rhs> {
/// The type of the matrix-matrix multiplication's result. See the
/// [*differentiability arithmetic*] for more details.
///
/// [*differentiability arithmetic*]: index.html#differentiability-arithmetic
type Output;
/// Computes the matrix-matrix multiplication between `self` and `other`.
fn mm(self, other: Rhs) -> Self::Output;
}
// ~~~~~~~~~~~~~~~~~~~~~~~~~~~~ Matrix Multiplication with Transposition ~~~~~~~~~~~~~~~~~~~~~~~~~~~
/// Matrix-matrix multiplication with transposed right hand side operand.
///
/// This fused operation is marginally faster than performing the matrix-matrix multiplication
/// and transposition separately.
pub trait MatMatMulT<Rhs> {
/// The type of the matrix-matrix multiplication with transposed right hand side operand's
/// result. See the [*differentiability arithmetic*] for more details.
///
/// [*differentiability arithmetic*]: index.html#differentiability-arithmetic
type Output;
/// Computes the matrix-matrix multiplication between `self` and transposed `other`.
fn mm_t(self, other: Rhs) -> Self::Output;
}
// ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ Matrix Vector Multiplication ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
/// Matrix-vector multiplication.
pub trait MatVecMul<Rhs> {
/// The type of the matrix-vector multiplication's result. See the
/// [*differentiability arithmetic*] for more details.
///
/// [*differentiability arithmetic*]: index.html#differentiability-arithmetic
type Output;
/// Computes the matrix-vector multiplication between `self` and `other`.
fn mv(self, other: Rhs) -> Self::Output;
}
// ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ Vector Matrix Multiplication ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
/// Vector-matrix multiplication.
pub trait VecMatMul<Rhs> {
/// The type of the vector-matrix multiplication's result. See the
/// [*differentiability arithmetic*] for more details.
///
/// [*differentiability arithmetic*]: index.html#differentiability-arithmetic
type Output;
/// Computes the vector-matrix multiplication between `self` and `other`.
fn vm(self, other: Rhs) -> Self::Output;
}
// ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ Vector Vector Multiplication ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
/// Vector-vector multiplication, *a.k.a. dot product or inner product*.
pub trait VecVecMul<Rhs> {
/// The type of the dot product's result. See the [*differentiability arithmetic*] for
/// more details.
///
/// [*differentiability arithmetic*]: index.html#differentiability-arithmetic
type Output;
/// Computes the dot product between `self` and `other`.
fn vv(self, other: Rhs) -> Self::Output;
}
// ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
// ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ Cat and Stack traits ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
// ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
/// Concatenation.
pub trait Cat<Rhs> {
/// The type of the concatenation's result. See the [*differentiability arithmetic*] for
/// more details.
///
/// [*differentiability arithmetic*]: index.html#differentiability-arithmetic
type Output;
/// Concatenates variables along the given axis.
fn cat(self, other: Rhs, axis: usize) -> Self::Output;
}
/// Stacking.
pub trait Stack<Rhs> {
/// The type of the stacking's result. See the [*differentiability arithmetic*] for
/// more details.
///
/// [*differentiability arithmetic*]: index.html#differentiability-arithmetic
type Output;
/// Stacks variables along the given axis.
fn stack(self, other: Rhs, axis: usize) -> Self::Output;
}
// ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
// ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ Higher abstraction traits ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
// ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
/// Defines the interface required to build computational graph's leaves using a dynamically typed
/// non-differentiable variables.
pub trait Variable<D: Dimension> {
fn get_node(&self) -> Rc<dyn Data<Dim = D>>;
fn get_past(&self) -> VarHistory;
}
impl<T: Data<Dim = D>, D: Dimension> Variable<D> for Var<T> {
fn get_node(&self) -> Rc<dyn Data<Dim = D>> {
self.node.clone()
}
fn get_past(&self) -> VarHistory {
self.past.clone()
}
}
/// Defines the interface required to build computational graph's leaves using a dynamically typed
/// differentiable variables.
pub trait DifferentiableVariable<D: Dimension> {
fn get_var(&self) -> Box<dyn Variable<D>>;
fn get_node(&self) -> Rc<dyn GradientOverwrite<D>>;
fn get_past(&self) -> VarDiffHistory;
}
impl<T: Data<Dim = D>, U: GradientOverwrite<D>, D: Dimension> DifferentiableVariable<D>
for VarDiff<T, U>
{
fn get_var(&self) -> Box<dyn Variable<D>> {
Box::new(self.var.clone())
}
fn get_node(&self) -> Rc<dyn GradientOverwrite<D>> {
self.node.clone()
}
fn get_past(&self) -> VarDiffHistory {
self.past.clone()
}
}
// ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
// ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ Tests ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
// ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
#[cfg(test)]
mod test;