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
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

use num_traits::Float;
use rand_distr::uniform::SampleUniform;

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

use super::*;

pub mod builder {
    #[derive(Debug)]
    pub struct Conv2D<
        const IN_CHAN: usize,
        const OUT_CHAN: usize,
        const KERNEL_SIZE: usize,
        const STRIDE: usize = 1,
        const PADDING: usize = 0,
        const DILATION: usize = 1,
        const GROUPS: usize = 1,
    >;
}

impl<
        const I: usize,
        const O: usize,
        const K: usize,
        const S: usize,
        const P: usize,
        const L: usize,
        const G: usize,
        E,
        D,
    > BuildOnDevice<D, E> for builder::Conv2D<I, O, K, S, P, L, G>
where
    E: Dtype,
    D: Device<E>,
    Const<{ I / G }>: Sized,
    Conv2D<I, O, K, S, P, L, G, E, D>: BuildModule<D, E>,
{
    type Built = Conv2D<I, O, K, S, P, L, G, E, D>;
    fn try_build_on_device(device: &D) -> Result<Self::Built, <D>::Err> {
        Self::Built::try_build(device)
    }
}

/// **Requires Nightly** Performs *unbiased* 2d convolutions on 3d and 4d images.
///
/// **Pytorch Equivalent**: `torch.nn.Conv2d(..., bias=False)`
///
/// To create a biased conv, combine with [crate::nn::modules::Bias2D]:
/// ```ignore
/// # use dfdx::prelude::*;
/// type BiasedConv = (Conv2D<3, 5, 4>, Bias2D<5>);
/// ```
///
/// Generics:
/// - `IN_CHAN`: The number of input channels in an image.
/// - `OUT_CHAN`: The number of channels in the output of the layer.
/// - `KERNEL_SIZE`: The size of the kernel applied to both width and height of the images.
/// - `STRIDE`: How far to move the kernel each step. Defaults to `1`
/// - `PADDING`: How much zero padding to add around the images. Defaults to `0`.
/// - `DILATION`: Controls the spacing between kernel points. Defaults to `1`.
/// - `GROUPS`: Controls the connections between inputs and outputs.
///     `IN_CHAN` and `OUT_CHAN` must both be divisible by `GROUPS`. For example,
///
/// See [conv animations](https://github.com/vdumoulin/conv_arithmetic/blob/master/README.md) for helpful
/// visualization of all of these parameters.

#[derive(Debug, Clone)]
pub struct Conv2D<
    const IN_CHAN: usize,
    const OUT_CHAN: usize,
    const KERNEL_SIZE: usize,
    const STRIDE: usize,
    const PADDING: usize,
    const DILATION: usize,
    const GROUPS: usize,
    E: Dtype,
    D: Storage<E>,
> where
    Const<{ IN_CHAN / GROUPS }>: Sized,
{
    pub weight: Tensor<Rank4<OUT_CHAN, { IN_CHAN / GROUPS }, KERNEL_SIZE, KERNEL_SIZE>, E, D>,
}

impl<
        const I: usize,
        const O: usize,
        const K: usize,
        const S: usize,
        const P: usize,
        const L: usize,
        const G: usize,
        E,
        D,
    > TensorCollection<E, D> for Conv2D<I, O, K, S, P, L, G, E, D>
where
    Const<{ I / G }>: Sized,
    E: Dtype + Float + SampleUniform,
    D: Device<E>,
{
    type To<E2: Dtype, D2: Device<E2>> = Conv2D<I, O, K, S, P, L, G, E2, D2>;

    fn iter_tensors<V: ModuleVisitor<Self, E, D>>(
        visitor: &mut V,
    ) -> Result<Option<Self::To<V::E2, V::D2>>, V::Err> {
        visitor.visit_fields(
            Self::tensor(
                "weight",
                |s| &s.weight,
                |s| &mut s.weight,
                TensorOptions::reset_with(|t| {
                    let scale = E::from_f64(G as f64 / (I * K * K) as f64).unwrap();
                    let b = scale.sqrt();
                    t.try_fill_with_distr(rand_distr::Uniform::new(-b, b))
                }),
            ),
            |weight| Conv2D { weight },
        )
    }
}

impl<
        const I: usize,
        const O: usize,
        const K: usize,
        const S: usize,
        const P: usize,
        const L: usize,
        const G: usize,
        E,
        D,
        Img,
    > Module<Img> for Conv2D<I, O, K, S, P, L, G, E, D>
where
    Const<{ I / G }>: Sized,
    E: Dtype,
    D: Device<E>,
    (Img, Tensor<Rank4<O, { I / G }, K, K>, E, D>):
        TryConv2D<Const<S>, Const<P>, Const<L>, Const<G>>,
{
    type Output = <(Img, Tensor<Rank4<O, { I / G }, K, K>, E, D>) as TryConv2D<
        Const<S>,
        Const<P>,
        Const<L>,
        Const<G>,
    >>::Convolved;
    type Error = <(Img, Tensor<Rank4<O, { I / G }, K, K>, E, D>) as TryConv2D<
        Const<S>,
        Const<P>,
        Const<L>,
        Const<G>,
    >>::Error;

    fn try_forward(&self, x: Img) -> Result<Self::Output, Self::Error> {
        (x, self.weight.clone()).try_conv2d(Const, Const, Const, Const)
    }
}

impl<
        const I: usize,
        const O: usize,
        const K: usize,
        const S: usize,
        const P: usize,
        const L: usize,
        const G: usize,
        E: Dtype,
        D: Storage<E>,
    > NonMutableModule for Conv2D<I, O, K, S, P, L, G, E, D>
where
    Const<{ I / G }>: Sized,
{
}

#[cfg(test)]
mod tests {
    use crate::{
        optim::*,
        tensor::{AsArray, SampleTensor, ZerosTensor},
        tests::*,
    };

    use super::{builder::Conv2D, *};

    #[rustfmt::skip]
    #[test]
    fn test_forward_3d_sizes() {
        let dev: TestDevice = Default::default();
        let x = dev.zeros::<Rank3<3, 10, 10>>();
        let _: Tensor<Rank3<2, 8, 8>, _, _, _> = dev.build_module::<Conv2D<3, 2, 3>, TestDtype>().forward(x.clone());
        let _: Tensor<Rank3<4, 8, 8>, _, _, _> = dev.build_module::<Conv2D<3, 4, 3>, TestDtype>().forward(x.clone());
        let _: Tensor<Rank3<4, 9, 9>, _, _, _> = dev.build_module::<Conv2D<3, 4, 2>, TestDtype>().forward(x.clone());
        let _: Tensor<Rank3<4, 7, 7>, _, _, _> = dev.build_module::<Conv2D<3, 4, 4>, TestDtype>().forward(x.clone());
        let _: Tensor<Rank3<2, 4, 4>, _, _, _> = dev.build_module::<Conv2D<3, 2, 3, 2>, TestDtype>().forward(x.clone());
        let _: Tensor<Rank3<2, 3, 3>, _, _, _> = dev.build_module::<Conv2D<3, 2, 3, 3>, TestDtype>().forward(x.clone());
        let _: Tensor<Rank3<2, 10, 10>, _, _, _> = dev.build_module::<Conv2D<3, 2, 3, 1, 1>, TestDtype>().forward(x.clone());
        let _: Tensor<Rank3<2, 12, 12>, _, _, _> = dev.build_module::<Conv2D<3, 2, 3, 1, 2>, TestDtype>().forward(x.clone());
        let _: Tensor<Rank3<2, 6, 6>, _, _, _> = dev.build_module::<Conv2D<3, 2, 3, 2, 2>, TestDtype>().forward(x.clone());
    }

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

        let x = dev.zeros::<Rank3<16, 10, 10>>();

        let m = dev.build_module::<Conv2D<16, 32, 3, 1, 0, 1, 1>, TestDtype>();
        let _: Tensor<Rank4<32, 16, 3, 3>, _, _> = m.weight;
        let _: Tensor<Rank3<32, 8, 8>, _, _> = m.forward(x.clone());

        let m = dev.build_module::<Conv2D<16, 32, 3, 1, 0, 1, 2>, TestDtype>();
        let _: Tensor<Rank4<32, 8, 3, 3>, _, _> = m.weight;
        let _: Tensor<Rank3<32, 8, 8>, _, _> = m.forward(x.clone());

        let m = dev.build_module::<Conv2D<16, 32, 3, 1, 0, 1, 4>, TestDtype>();
        let _: Tensor<Rank4<32, 4, 3, 3>, _, _> = m.weight;
        let _: Tensor<Rank3<32, 8, 8>, _, _> = m.forward(x.clone());

        let m = dev.build_module::<Conv2D<16, 32, 3, 1, 0, 1, 8>, TestDtype>();
        let _: Tensor<Rank4<32, 2, 3, 3>, _, _> = m.weight;
        let _: Tensor<Rank3<32, 8, 8>, _, _> = m.forward(x.clone());

        let m = dev.build_module::<Conv2D<16, 32, 3, 1, 0, 1, 16>, TestDtype>();
        let _: Tensor<Rank4<32, 1, 3, 3>, _, _> = m.weight;
        let _: Tensor<Rank3<32, 8, 8>, _, _> = m.forward(x);
    }

    #[rustfmt::skip]
    #[test]
    fn test_forward_4d_sizes() {
        let dev: TestDevice = Default::default();
        let x = dev.zeros::<Rank4<5, 3, 10, 10>>();
        let _: Tensor<Rank4<5, 2, 8, 8>, _, _, _> = dev.build_module::<Conv2D<3, 2, 3>, TestDtype>().forward(x.clone());
        let _: Tensor<Rank4<5, 4, 8, 8>, _, _, _> = dev.build_module::<Conv2D<3, 4, 3>, TestDtype>().forward(x.clone());
        let _: Tensor<Rank4<5, 4, 9, 9>, _, _, _> = dev.build_module::<Conv2D<3, 4, 2>, TestDtype>().forward(x.clone());
        let _: Tensor<Rank4<5, 4, 7, 7>, _, _, _> = dev.build_module::<Conv2D<3, 4, 4>, TestDtype>().forward(x.clone());
        let _: Tensor<Rank4<5, 2, 4, 4>, _, _, _> = dev.build_module::<Conv2D<3, 2, 3, 2>, TestDtype>().forward(x.clone());
        let _: Tensor<Rank4<5, 2, 3, 3>, _, _, _> = dev.build_module::<Conv2D<3, 2, 3, 3>, TestDtype>().forward(x.clone());
        let _: Tensor<Rank4<5, 2, 10, 10>, _, _, _> = dev.build_module::<Conv2D<3, 2, 3, 1, 1>, TestDtype>().forward(x.clone());
        let _: Tensor<Rank4<5, 2, 12, 12>, _, _, _> = dev.build_module::<Conv2D<3, 2, 3, 1, 2>, TestDtype>().forward(x.clone());
        let _: Tensor<Rank4<5, 2, 6, 6>, _, _, _> = dev.build_module::<Conv2D<3, 2, 3, 2, 2>, TestDtype>().forward(x.clone());
    }

    #[test]
    fn test_2_conv_sizes() {
        let dev = Cpu::default();
        type A = Conv2D<1, 2, 3>;
        type B = Conv2D<2, 4, 3>;
        let _: Tensor<Rank3<4, 6, 6>, _, _> = dev
            .build_module::<(A, B), TestDtype>()
            .forward(dev.zeros::<Rank3<1, 10, 10>>());
    }

    #[test]
    fn test_3_conv_sizes() {
        type A = Conv2D<1, 2, 3>;
        type B = Conv2D<2, 4, 3>;
        type C = Conv2D<4, 1, 1, 1, 1>;

        let dev = Cpu::default();
        let _: Tensor<Rank3<1, 8, 8>, _, _> = dev
            .build_module::<(A, B, C), TestDtype>()
            .forward_mut(dev.zeros::<Rank3<1, 10, 10>>());
    }

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

        let mut m = dev.build_module::<Conv2D<2, 4, 3>, TestDtype>();

        let weight_init = m.weight.clone();

        let mut opt = Sgd::new(&m, Default::default());
        let out = m.forward(dev.sample_normal::<Rank4<8, 2, 28, 28>>().leaky_trace());
        let g = out.square().mean().backward();

        assert_ne!(
            g.get(&m.weight).array(),
            [[[[TestDtype::zero(); 3]; 3]; 2]; 4]
        );

        opt.update(&mut m, &g).expect("unused params");

        assert_ne!(weight_init.array(), m.weight.array());
    }
}