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
use rand::distributions::Distribution;
use super::*;
use crate::shapes::*;
use std::sync::Arc;
/// The single tensor struct that stores nd arrays and tapes.
///
/// See module level documentation on how to create and use tensors.
///
/// Generics:
/// 1. [Shape] - the shape of the underlying nd array
/// 2. [Dtype] - the type of the datas stored in the array
/// 3. [Storage] - the device the array is stored on
/// 4. [Tape] - the tape the tensor has
///
/// Examples:
/// ```rust
/// # use dfdx::prelude::*;
/// # let dev: Cpu = Default::default();
/// // A 1d tensor with 1000 f32 elements, stored on the Cpu
/// type A = Tensor<Rank1<1000>, f32, Cpu>;
///
/// // A 2d tensor with bool elements, stored on the Cpu
/// type B = Tensor<Rank2<2, 3>, bool, Cpu>;
///
/// // A 3d tensor with usize elements, stored on the Cpu, without any tape
/// type C = Tensor<Rank3<4, 2, 3>, usize, Cpu, NoneTape>;
/// ```
#[derive(Debug, Clone)]
pub struct Tensor<S: Shape, E, D: Storage<E>, T = NoneTape> {
pub(crate) id: UniqueId,
pub(crate) data: Arc<D::Vec>,
pub(crate) shape: S,
pub(crate) strides: S::Concrete,
pub(crate) device: D,
pub(crate) tape: T,
}
impl<S: Shape, E, D: Storage<E>, T> HasShape for Tensor<S, E, D, T> {
type WithShape<New: Shape> = Tensor<New, E, D, T>;
type Shape = S;
fn shape(&self) -> &Self::Shape {
&self.shape
}
}
impl<S: Shape, E: Unit, D: Storage<E>, T> HasUnitType for Tensor<S, E, D, T> {
type Unit = E;
}
impl<S: Shape, E: Dtype, D: Storage<E>, T> HasDtype for Tensor<S, E, D, T> {
type Dtype = E;
}
impl<S: Shape, E, D: Storage<E>, T> HasErr for Tensor<S, E, D, T> {
type Err = D::Err;
}
/// Something that can trace gradients
pub trait Trace<E, D: Storage<E>>: Clone {
type Traced;
/// Start tracking gradients, clones self. The gradients will never free
/// temporary gradients - See [Gradients::leaky()] for more info.
///
/// Prefer to use [Tensor::trace()] with gradients allocated
/// with [crate::nn::ZeroGrads::alloc_grads()].
fn leaky_trace(&self) -> Self::Traced {
self.clone().leaky_traced()
}
/// Start tracking gradients. The gradients will never free
/// temporary gradients - See [Gradients::leaky()] for more info.
///
/// Prefer to use [Tensor::traced()] with gradients allocated
/// with [crate::nn::ZeroGrads::alloc_grads()].
fn leaky_traced(self) -> Self::Traced;
/// Accumulates gradients into `gradients`, clones self. Use [crate::nn::ZeroGrads::alloc_grads()]
/// to create gradients.
fn trace(&self, gradients: Gradients<E, D>) -> Self::Traced {
self.clone().traced(gradients)
}
/// Accumulates gradients into `gradients`. Use [crate::nn::ZeroGrads::alloc_grads()]
/// to create gradients.
fn traced(self, gradients: Gradients<E, D>) -> Self::Traced;
}
impl<S: Shape, E: Unit, F: Unit, D: Storage<F> + Storage<E>> Trace<E, D>
for Tensor<S, F, D, NoneTape>
{
type Traced = Tensor<S, F, D, OwnedTape<E, D>>;
fn leaky_traced(self) -> Self::Traced {
self.put_tape(Default::default())
}
fn traced(self, gradients: Gradients<E, D>) -> Self::Traced {
self.put_tape(OwnedTape {
gradients,
operations: std::vec::Vec::new(),
})
}
}
impl<S: Shape, E, D: Storage<E>, T> Tensor<S, E, D, T> {
/// Clone and insert a new tape of type `New` into the tensor
pub fn retaped<New: Tape<E, D>>(&self) -> Tensor<S, E, D, New> {
Tensor {
id: self.id,
data: self.data.clone(),
shape: self.shape,
strides: self.strides,
device: self.device.clone(),
tape: Default::default(),
}
}
/// Get a reference to the tensor's `Storage`
pub fn device(&self) -> &D {
&self.device
}
}
/// Put a tape of type `T` into the tensor
pub trait PutTape<T> {
type Output;
/// ```rust
/// # use dfdx::prelude::*;
/// # let dev: Cpu = Default::default();
/// let a: Tensor<Rank2<2, 3>, f32, _, NoneTape> = dev.zeros();
/// let a: Tensor<Rank2<2, 3>, f32, _, OwnedTape<f32, Cpu>> = a.put_tape(Default::default());
/// ```
fn put_tape(self, tape: T) -> Self::Output;
}
impl<S: Shape, E, D: Storage<E>, T> PutTape<T> for Tensor<S, E, D> {
type Output = Tensor<S, E, D, T>;
fn put_tape(self, tape: T) -> Self::Output {
Tensor {
id: self.id,
data: self.data,
shape: self.shape,
strides: self.strides,
device: self.device,
tape,
}
}
}
/// Remove the tape from a tensor
pub trait SplitTape {
/// The type of tape the tensor has now
type Tape;
// The type of Self without the tape.
type NoTape: Clone + PutTape<Self::Tape, Output = Self>;
/// Splits tape off of self
/// ```rust
/// # use dfdx::prelude::*;
/// # let dev: Cpu = Default::default();
/// # let grads = Gradients::leaky();
/// let a: Tensor<Rank1<5>, f32, _, OwnedTape<f32, _>> = dev.zeros().traced(grads);
/// let (a, tape): (Tensor<_, _, _, NoneTape>, OwnedTape<f32, _>) = a.split_tape();
/// ```
fn split_tape(self) -> (Self::NoTape, Self::Tape);
}
impl<S: Shape, E: Clone, D: Storage<E>, T> SplitTape for Tensor<S, E, D, T> {
type Tape = T;
type NoTape = Tensor<S, E, D>;
fn split_tape(self) -> (Self::NoTape, Self::Tape) {
(
Tensor {
id: self.id,
data: self.data,
shape: self.shape,
strides: self.strides,
device: self.device,
tape: NoneTape,
},
self.tape,
)
}
}
/// Clones self and inserts a new empty tape into the clone
pub trait WithEmptyTape {
/// Clones self and inserts a new empty tape into the clone
fn with_empty_tape(&self) -> Self;
}
impl<S: Shape, E, D: Storage<E>, T: Default> WithEmptyTape for Tensor<S, E, D, T> {
fn with_empty_tape(&self) -> Self {
Tensor {
id: self.id,
data: self.data.clone(),
shape: self.shape,
strides: self.strides,
device: self.device.clone(),
tape: Default::default(),
}
}
}
impl<S: Shape, E: Dtype, D: ZeroFillStorage<E>, T> Tensor<S, E, D, T> {
/// Fills the tensor with zeros
pub fn fill_with_zeros(&mut self) {
self.try_fill_with_zeros().unwrap()
}
/// Fallible version of [Tensor::fill_with_zeros]
pub fn try_fill_with_zeros(&mut self) -> Result<(), D::Err> {
self.device
.try_fill_with_zeros(Arc::make_mut(&mut self.data))
}
}
impl<S: Shape, E: Dtype, D: OneFillStorage<E>, T> Tensor<S, E, D, T> {
/// Fills the tensor with ones
pub fn fill_with_ones(&mut self) {
self.try_fill_with_ones().unwrap()
}
/// Fallible version of [Tensor::fill_with_ones]
pub fn try_fill_with_ones(&mut self) -> Result<(), D::Err> {
self.device
.try_fill_with_ones(Arc::make_mut(&mut self.data))
}
}
impl<S: Shape, E: Unit, D: SampleTensor<E>, T> Tensor<S, E, D, T> {
/// Fills the tensor with random data from the distribution
pub fn fill_with_distr<Distr: Distribution<E>>(&mut self, distr: Distr) {
self.try_fill_with_distr(distr).unwrap()
}
/// Fallible version of [Tensor::fill_with_distr]
pub fn try_fill_with_distr<Distr: Distribution<E>>(
&mut self,
distr: Distr,
) -> Result<(), D::Err> {
self.device
.try_fill_with_distr(Arc::make_mut(&mut self.data), distr)
}
}
pub type Tensor0D<Tape = NoneTape> = Tensor<Rank0, f32, Cpu, Tape>;
pub type Tensor1D<const M: usize, Tape = NoneTape> = Tensor<Rank1<M>, f32, Cpu, Tape>;
pub type Tensor2D<const M: usize, const N: usize, Tape = NoneTape> =
Tensor<Rank2<M, N>, f32, Cpu, Tape>;
pub type Tensor3D<const M: usize, const N: usize, const O: usize, Tape = NoneTape> =
Tensor<Rank3<M, N, O>, f32, Cpu, Tape>;
pub type Tensor4D<const M: usize, const N: usize, const O: usize, const P: usize, Tape = NoneTape> =
Tensor<Rank4<M, N, O, P>, f32, Cpu, Tape>;
pub type Tensor5D<
const M: usize,
const N: usize,
const O: usize,
const P: usize,
const Q: usize,
Tape = NoneTape,
> = Tensor<Rank5<M, N, O, P, Q>, f32, Cpu, Tape>;
pub type Tensor6D<
const M: usize,
const N: usize,
const O: usize,
const P: usize,
const Q: usize,
const R: usize,
Tape = NoneTape,
> = Tensor<Rank6<M, N, O, P, Q, R>, f32, Cpu, Tape>;