dfdx 0.13.0

Ergonomic auto differentiation in Rust, with pytorch like apis.
Documentation 
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

use crate::{shapes::*, tensor::*};

use std::vec::Vec;

mod cpu_kernel;
#[cfg(feature = "cuda")]
mod cuda_kernel;

/// Stack an array or vec of tensors together along a new dimension.
///
/// An array of tensors will be turned into a [Const] dim, and
/// a `Vec` of tensors will be turned into a [usize] dim.
///
/// **Pytorch equivalent** `torch.stack`.
///
/// Stacking with an array:
/// ```rust
/// # use dfdx::prelude::*;
/// # let dev: Cpu = Default::default();
/// let a: Tensor<Rank2<3, 4>, f32, _> = dev.zeros();
/// let b: Tensor<Rank2<3, 4>, f32, _> = dev.zeros();
/// let _: Tensor<Rank3<2, 3, 4>, f32, _> = [a, b].stack();
/// ```
///
/// Stacking with a vec:
/// ```rust
/// # use dfdx::prelude::*;
/// # let dev: Cpu = Default::default();
/// let a: Tensor<Rank2<3, 4>, f32, _> = dev.zeros();
/// let b: Tensor<Rank2<3, 4>, f32, _> = dev.zeros();
/// let _: Tensor<(usize, Const<3>, Const<4>), f32, _> = vec![a, b].stack();
/// ```
pub trait TryStack: Sized {
    type Stacked;
    type Err: std::fmt::Debug;

    /// Stack an array or vec of tensors together along a new dimension.
    fn stack(self) -> Self::Stacked {
        self.try_stack().unwrap()
    }
    /// Fallible version of [TryStack::stack]
    fn try_stack(self) -> Result<Self::Stacked, Self::Err>;
}

impl<S: Shape, E: Dtype, D: StackKernel<E>, T, const N: usize> TryStack for [Tensor<S, E, D, T>; N]
where
    S: AddDim<Const<N>>,
    T: Tape<E, D>,
{
    type Stacked = Tensor<S::Larger, E, D, T>;
    type Err = D::Err;
    fn try_stack(self) -> Result<Self::Stacked, Self::Err> {
        try_stack(self)
    }
}

impl<S: Shape, E: Dtype, D: StackKernel<E>, T> TryStack for std::vec::Vec<Tensor<S, E, D, T>>
where
    S: AddDim<usize>,
    T: Tape<E, D>,
{
    type Stacked = Tensor<S::Larger, E, D, T>;
    type Err = D::Err;
    fn try_stack(self) -> Result<Self::Stacked, Self::Err> {
        try_stack(self)
    }
}

impl<A: TryStack, B: TryStack<Err = A::Err>> TryStack for (A, B) {
    type Stacked = (A::Stacked, B::Stacked);
    type Err = A::Err;
    fn try_stack(self) -> Result<Self::Stacked, Self::Err> {
        Ok((self.0.try_stack()?, self.1.try_stack()?))
    }
}

pub trait AddDim<D: Dim>: Shape {
    type Larger: Shape;
    fn add_dim(&self, dim: D) -> Self::Larger;
}

impl<New: Dim> AddDim<New> for () {
    type Larger = (New,);
    fn add_dim(&self, dim: New) -> Self::Larger {
        (dim,)
    }
}
impl<D1: Dim, New: Dim> AddDim<New> for (D1,) {
    type Larger = (New, D1);
    fn add_dim(&self, dim: New) -> Self::Larger {
        (dim, self.0)
    }
}
impl<D1: Dim, D2: Dim, New: Dim> AddDim<New> for (D1, D2) {
    type Larger = (New, D1, D2);
    fn add_dim(&self, dim: New) -> Self::Larger {
        (dim, self.0, self.1)
    }
}
impl<D1: Dim, D2: Dim, D3: Dim, New: Dim> AddDim<New> for (D1, D2, D3) {
    type Larger = (New, D1, D2, D3);
    fn add_dim(&self, dim: New) -> Self::Larger {
        (dim, self.0, self.1, self.2)
    }
}
impl<D1: Dim, D2: Dim, D3: Dim, D4: Dim, New: Dim> AddDim<New> for (D1, D2, D3, D4) {
    type Larger = (New, D1, D2, D3, D4);
    fn add_dim(&self, dim: New) -> Self::Larger {
        (dim, self.0, self.1, self.2, self.3)
    }
}

pub trait StackKernel<E: Dtype>: Storage<E> {
    fn forward<S: Shape, Num: Dim>(
        &self,
        num: Num,
        inp: &[Tensor<S, E, Self>],
    ) -> Result<Tensor<S::Larger, E, Self>, Self::Err>
    where
        S: AddDim<Num>;
    fn backward(
        &self,
        grad_inp: Vec<&mut Self::Vec>,
        grad_out: &Self::Vec,
    ) -> Result<(), Self::Err>;
}

fn try_stack<S: Shape, E: Dtype, D: StackKernel<E>, T, Items>(
    items: Items,
) -> Result<Tensor<S::Larger, E, D, T>, D::Err>
where
    Items: Array<Tensor<S, E, D, T>>,
    S: AddDim<Items::Dim>,
    T: Tape<E, D> + Merge<T>,
{
    let new_dim = items.dim();
    assert!(new_dim.size() > 0);

    // need to split tape and transform into Vec for ease of implementation
    let mut tensors = Vec::with_capacity(new_dim.size());
    let mut tape: T = Default::default();
    for item in items.into_iter() {
        let (item, rhs): (Tensor<S, E, D>, T) = item.split_tape();
        tape = tape.merge(rhs);
        tensors.push(item);
    }

    // check that all the shapes are equal
    let device = tensors[0].device.clone();
    let shape = *tensors[0].shape();
    for t in tensors.iter() {
        assert_eq!(t.shape(), &shape);
    }

    // we map to Storage<E> refs so kernels don't have to know about tensors
    let out = device.forward(new_dim, &tensors)?;

    let inp_ghosts: Vec<_> = tensors.iter().map(|t| t.ghost()).collect();
    let out_ghost = out.ghost();
    tape.add_backward_op(move |grads| {
        for t in inp_ghosts.iter() {
            grads.try_alloc_for(t)?;
        }
        grads.try_alloc_for(&out_ghost)?;
        let (grad_inp, grad_out) = grads.many_and_ref(&inp_ghosts, &out_ghost);
        device.backward(grad_inp, grad_out)
    });
    Ok(out.put_tape(tape))
}

#[cfg(test)]
mod tests {
    use super::*;
    use crate::{tensor_ops::*, tests::*};

    #[test]
    fn test_valid_stacks() {
        let dev: TestDevice = Default::default();

        {
            let x: Tensor<(), TestDtype, _> = dev.sample_normal();
            let y: Tensor<(), TestDtype, _> = dev.sample_normal();
            let _: Tensor<Rank1<2>, TestDtype, _> = [x, y].stack();
        }

        {
            let x: Tensor<Rank1<3>, TestDtype, _> = dev.sample_normal();
            let y: Tensor<Rank1<3>, TestDtype, _> = dev.sample_normal();
            let z: Tensor<Rank1<3>, TestDtype, _> = dev.sample_normal();
            let _: Tensor<Rank2<3, 3>, TestDtype, _> = [x, y, z].stack();
        }

        {
            let x: Tensor<Rank2<2, 3>, TestDtype, _> = dev.sample_normal();
            let y: Tensor<Rank2<2, 3>, TestDtype, _> = dev.sample_normal();
            let z: Tensor<Rank2<2, 3>, TestDtype, _> = dev.sample_normal();
            let r: Tensor<(usize, Const<2>, Const<3>), TestDtype, _> = std::vec![x, y, z].stack();
            assert_eq!(r.shape().0, 3);
        }
    }

    #[test]
    #[should_panic]
    fn test_stack_with_diff_sizes() {
        let dev: TestDevice = Default::default();
        let x: Tensor<_, TestDtype, _> = dev.sample_like(&(2, 3), rand_distr::StandardNormal);
        let y: Tensor<_, TestDtype, _> = dev.sample_like(&(3, 4), rand_distr::StandardNormal);
        let _ = [x, y].stack();
    }

    #[test]
    #[should_panic]
    fn test_stack_with_diff_strides() {
        let dev: TestDevice = Default::default();
        let x: Tensor<Rank2<2, 3>, TestDtype, _> = dev.sample_normal();
        let y: Tensor<Rank1<3>, TestDtype, _> = dev.sample_normal();
        let _ = [x, y.broadcast()].stack();
    }

    #[test]
    fn test_stack_with_all_broadcasted() {
        let dev: TestDevice = Default::default();
        let x: Tensor<Rank1<3>, TestDtype, _> = dev.sample_normal();
        let y: Tensor<Rank1<3>, TestDtype, _> = dev.sample_normal();
        let r = [
            x.leaky_trace().broadcast::<Rank2<4, 3>, _>(),
            y.leaky_trace().broadcast(),
        ]
        .stack();
        assert_eq!(r.array(), [[x.array(); 4], [y.array(); 4]]);
        let g = r.exp().mean().backward();
        let g1 = [x.leaky_trace(), y.leaky_trace()]
            .stack()
            .exp()
            .mean()
            .backward();
        assert_eq!(g.get(&x).array(), g1.get(&x).array());
        assert_eq!(g.get(&y).array(), g1.get(&y).array());
    }

    #[test]
    fn test_stack_backwards() {
        let dev: TestDevice = Default::default();

        let x: Tensor<Rank2<2, 3>, TestDtype, _> = dev.sample_normal();
        let y: Tensor<Rank2<2, 3>, TestDtype, _> = dev.sample_normal();
        let z: Tensor<Rank2<2, 3>, TestDtype, _> = dev.sample_normal();
        let r = [x.leaky_trace(), y.leaky_trace(), z.leaky_trace()].stack();
        assert_eq!(r.array(), [x.array(), y.array(), z.array()]);
        let r1 = r.retaped::<NoneTape>();
        let g1 = r1.leaky_trace().exp().mean().backward();
        let g = r.exp().mean().backward();
        let r_grad = g1.get(&r1).array();
        assert_eq!(r_grad[0], g.get(&x).array());
        assert_eq!(r_grad[1], g.get(&y).array());
        assert_eq!(r_grad[2], g.get(&z).array());
    }
}