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

mod cpu_kernel;

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

use super::ops::*;
use crate::{
    shapes::*,
    tensor::{HasErr, Merge, Storage, Tape, Tensor},
};

#[repr(C)]
#[derive(Debug, Default, Clone, Copy)]
pub struct BinaryAddKernelOp;

#[repr(C)]
#[derive(Debug, Clone, Copy)]
pub struct ScalarAddKernelOp<E> {
    scalar: E,
}

/// Element wise and scalar addition.
///
/// Example:
/// ```rust
/// # use dfdx::prelude::*;
/// # let dev: Cpu = Default::default();
/// let a = dev.tensor([[1.0, 2.0, 3.0], [-1.0, -2.0, -3.0]]);
/// let r = a + dev.ones();
/// assert_eq!(r.array(), [[2.0, 3.0, 4.0], [0.0, -1.0, -2.0]]);
/// ```
///
/// Adding a scalar:
/// ```rust
/// # use dfdx::prelude::*;
/// # let dev: Cpu = Default::default();
/// let a = dev.tensor([[1.0, 2.0, 3.0], [-1.0, -2.0, -3.0]]);
/// let r = a + 1.0;
/// assert_eq!(r.array(), [[2.0, 3.0, 4.0], [0.0, -1.0, -2.0]]);
/// ```
pub fn add<S: Shape, E: Dtype, D, T: Tape<E, D> + Merge<R>, R: Default>(
    lhs: Tensor<S, E, D, T>,
    rhs: Tensor<S, E, D, R>,
) -> Tensor<S, E, D, T>
where
    D: BinaryKernel<BinaryAddKernelOp, E>,
{
    lhs + rhs
}

/// Fallible version of [std::ops::Add]. See [add]
pub trait TryAdd<Rhs = Self>: HasErr {
    fn try_add(self, rhs: Rhs) -> Result<Self, Self::Err>;
}

impl<S: Shape, E: Dtype, D, LhsTape: Tape<E, D>, R> TryAdd<Tensor<S, E, D, R>>
    for Tensor<S, E, D, LhsTape>
where
    D: BinaryKernel<BinaryAddKernelOp, E>,
    LhsTape: Merge<R>,
{
    /// See [add]
    fn try_add(self, rhs: Tensor<S, E, D, R>) -> Result<Self, Self::Err> {
        try_binary_op(BinaryAddKernelOp, self, rhs)
    }
}

impl<S: Shape, E: Dtype, D: UnaryKernel<ScalarAddKernelOp<E>, E>, T: Tape<E, D>> TryAdd<E>
    for Tensor<S, E, D, T>
{
    /// See [add]
    fn try_add(self, rhs: E) -> Result<Self, Self::Err> {
        try_unary_op(ScalarAddKernelOp { scalar: rhs }, self)
    }
}

#[cfg(feature = "f16")]
impl<S: Shape, D: UnaryKernel<ScalarAddKernelOp<half::f16>, half::f16>, T: Tape<half::f16, D>>
    TryAdd<f32> for Tensor<S, half::f16, D, T>
{
    /// See [add]
    fn try_add(self, rhs: f32) -> Result<Self, Self::Err> {
        let scalar = half::f16::from_f32(rhs);
        try_unary_op(ScalarAddKernelOp { scalar }, self)
    }
}

#[cfg(feature = "f16")]
impl<
        S: Shape,
        D: UnaryKernel<
            ScalarAddKernelOp<crate::dtypes::AMP<half::f16>>,
            crate::dtypes::AMP<half::f16>,
        >,
        T: Tape<crate::dtypes::AMP<half::f16>, D>,
    > TryAdd<f32> for Tensor<S, crate::dtypes::AMP<half::f16>, D, T>
{
    /// See [add]
    fn try_add(self, rhs: f32) -> Result<Self, Self::Err> {
        let scalar = crate::dtypes::AMP(half::f16::from_f32(rhs));
        try_unary_op(ScalarAddKernelOp { scalar }, self)
    }
}

impl<S: Shape, E: Dtype, D: Storage<E>, LhsTape: Tape<E, D>, Rhs> std::ops::Add<Rhs>
    for Tensor<S, E, D, LhsTape>
where
    Self: TryAdd<Rhs>,
{
    type Output = Self;
    /// See [add]
    fn add(self, rhs: Rhs) -> Self::Output {
        self.try_add(rhs).unwrap()
    }
}

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

    #[test]
    fn test_add_0d() {
        let dev: TestDevice = Default::default();
        let a = dev.tensor(1.0f64).to_dtype::<TestDtype>();
        let b = dev.tensor(1.0f64).to_dtype::<TestDtype>();

        let r = a.leaky_trace() + b.clone();
        assert_close_to_literal!(r, 2.0);
        let g = r.backward();
        assert_close_to_literal!(g.get(&a), 1.0);
        assert_close_to_literal!(g.get(&b), 1.0);
    }

    #[test]
    fn test_add_1d() {
        let dev: TestDevice = Default::default();
        let a = dev.tensor([1.0f64, 2.0, 3.0]).to_dtype::<TestDtype>();
        let b = dev.tensor([1.0f64, -1.0, 0.0]).to_dtype::<TestDtype>();

        let r = a.leaky_trace() + b.clone();
        assert_close_to_literal!(r, [2.0, 1.0, 3.0]);
        let g = r.mean().backward();
        assert_close_to_literal!(g.get(&a), [1.0 / 3.0; 3]);
        assert_close_to_literal!(g.get(&b), [1.0 / 3.0; 3]);
    }

    #[test]
    fn test_add_2d() {
        let dev: TestDevice = Default::default();
        let a = dev
            .tensor([[0.6570f64, 0.1708, 0.1500], [0.5658, 0.7010, 0.8342]])
            .to_dtype::<TestDtype>();
        let b = dev
            .tensor([[0.5199f64, 0.3844, 0.3759], [0.8259, 0.3682, 0.0388]])
            .to_dtype::<TestDtype>();

        let r = a.leaky_trace() + b.clone();
        assert_close_to_literal!(r, [[1.1769, 0.5552, 0.5259], [1.3917, 1.0692, 0.873]]);
        let g = r.mean().backward();
        assert_close_to_literal!(g.get(&a), [[1.0 / 6.0; 3]; 2]);
        assert_close_to_literal!(g.get(&b), [[1.0 / 6.0; 3]; 2]);
    }

    #[test]
    fn test_add_broadcast_bottom() {
        let dev: TestDevice = Default::default();
        let a = dev
            .tensor([[0.6570f64, 0.1708, 0.1500], [0.5658, 0.7010, 0.8342]])
            .to_dtype::<TestDtype>();
        let b = dev
            .tensor([[0.5199f64, 0.3844, 0.3759], [0.8259, 0.3682, 0.0388]])
            .to_dtype::<TestDtype>();

        let a2 = a.broadcast::<Rank3<2, 3, 4>, _>();
        let b2 = b.broadcast::<Rank3<2, 3, 4>, _>();

        let r = a2.leaky_trace() + b2.clone();
        assert_close_to_literal!(
            r,
            [
                [[1.1769; 4], [0.5552; 4], [0.5259; 4]],
                [[1.3917; 4], [1.0692; 4], [0.873; 4]]
            ]
        );
        let g = r.mean().backward();
        assert_close_to_literal!(g.get(&a2), [[[1.0 / 6.0; 4]; 3]; 2]);
        assert_close_to_literal!(g.get(&b2), [[[1.0 / 6.0; 4]; 3]; 2]);
    }

    #[test]
    fn test_add_broadcast_top() {
        let dev: TestDevice = Default::default();
        let a = dev
            .tensor([[0.6570f64, 0.1708, 0.1500], [0.5658, 0.7010, 0.8342]])
            .to_dtype::<TestDtype>();
        let b = dev
            .tensor([[0.5199f64, 0.3844, 0.3759], [0.8259, 0.3682, 0.0388]])
            .to_dtype::<TestDtype>();

        let a2 = a.broadcast::<Rank3<4, 2, 3>, _>();
        let b2 = b.broadcast::<Rank3<4, 2, 3>, _>();

        let r = a2.leaky_trace() + b2.clone();
        assert_close_to_literal!(r, [[[1.1769, 0.5552, 0.5259], [1.3917, 1.0692, 0.873]]; 4]);
        let g = r.mean().backward();
        assert_close_to_literal!(g.get(&a2), [[[1.0 / 6.0; 3]; 2]; 4]);
        assert_close_to_literal!(g.get(&b2), [[[1.0 / 6.0; 3]; 2]; 4]);
    }

    #[test]
    fn test_scalar_add_0d() {
        let dev: TestDevice = Default::default();
        let x: Tensor<(), TestDtype, _> = dev.zeros();
        let r = x.leaky_trace() + 1.0;
        assert_close_to_literal!(r, 1.0);
        let g = r.exp().backward();
        assert_close_to_literal!(g.get(&x), f64::exp(1.0));
    }

    #[test]
    fn test_scalar_add_1d() {
        let dev: TestDevice = Default::default();
        let x = dev.tensor([0.0, 1.0, 2.0]).to_dtype::<TestDtype>();
        let r = x.leaky_trace() + 0.5;
        assert_close_to_literal!(r, [0.5, 1.5, 2.5]);
        let g = r.exp().sum().backward();
        assert_close_to_literal!(g.get(&x), [1.6487212, 4.481689, 12.182494]);
    }

    #[test]
    fn test_scalar_add_2d() {
        let dev: TestDevice = Default::default();
        let x = dev.tensor([[0.0; 2]; 3]).to_dtype::<TestDtype>();
        let r = x.leaky_trace() + 0.5;
        assert_close_to_literal!(r, [[0.5; 2]; 3]);
        let g = r.exp().sum().backward();
        assert_close_to_literal!(g.get(&x), [[1.6487212; 2]; 3]);
    }
}