topos 0.13.0

An autodiff compiler stack in Rust.
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

use crate::{Shape, Tape, Tensor};

use crate::Module;

use super::{Conv2d, conv2d};

#[test]
fn conv2d_matches_the_hand_computed_fixture() {
    let tape: Tape<f64> = Tape::new();
    let input = tape.leaf(Tensor::new(
        [1, 1, 3, 3],
        (1..=9).map(|v| v as f64).collect::<Vec<_>>(),
    ));
    // The kernel picks the window's main diagonal: `[[1, 0], [0, 1]]`.
    let weights = tape.leaf(Tensor::new([1, 1, 2, 2], [1.0, 0.0, 0.0, 1.0]));
    let bias = tape.leaf(Tensor::new([1], [10.0]));

    let output = conv2d(input, weights, bias, 1, 0);
    assert_eq!(output.shape(), Shape::new([1, 1, 2, 2]));

    let output = output.symbol();
    let network = tape.into_network();
    let run = network.forward(&network.parameters(), []);
    assert_eq!(run.of(output).to_vec(), &[16.0, 18.0, 22.0, 24.0]);
}

#[test]
fn conv2d_mixes_channels_into_filters() {
    let tape: Tape<f64> = Tape::new();
    // Two channels; 1x1 kernels make the channel mixing exact:
    // filter 0 sums the channels, filter 1 takes `2 * c0 + 3 * c1`.
    let input = tape.leaf(Tensor::new(
        [1, 2, 2, 2],
        [1.0, 2.0, 3.0, 4.0, 10.0, 20.0, 30.0, 40.0],
    ));
    let weights = tape.leaf(Tensor::new([2, 2, 1, 1], [1.0, 1.0, 2.0, 3.0]));
    let bias = tape.leaf(Tensor::filled([2], 0.0));

    let output = conv2d(input, weights, bias, 1, 0);
    assert_eq!(output.shape(), Shape::new([1, 2, 2, 2]));

    let output = output.symbol();
    let network = tape.into_network();
    let run = network.forward(&network.parameters(), []);
    assert_eq!(
        run.of(output).to_vec(),
        &[11.0, 22.0, 33.0, 44.0, 32.0, 64.0, 96.0, 128.0]
    );
}

#[test]
fn conv2d_strides_the_windows() {
    let tape: Tape<f64> = Tape::new();
    let input = tape.leaf(Tensor::new([1, 1, 1, 5], [1.0, 2.0, 3.0, 4.0, 5.0]));
    let weights = tape.leaf(Tensor::new([1, 1, 1, 2], [1.0, 1.0]));
    let bias = tape.leaf(Tensor::filled([1], 0.0));

    let output = conv2d(input, weights, bias, 2, 0);
    assert_eq!(output.shape(), Shape::new([1, 1, 1, 2]));

    let output = output.symbol();
    let network = tape.into_network();
    let run = network.forward(&network.parameters(), []);
    assert_eq!(run.of(output).to_vec(), &[3.0, 7.0]);
}

#[test]
fn padding_surrounds_the_input_with_zeros() {
    let tape: Tape<f64> = Tape::new();
    let input = tape.leaf(Tensor::new([1, 1, 2, 2], [1.0, 2.0, 3.0, 4.0]));
    // An all-ones 3x3 kernel over the zero-padded 2x2 input sums the
    // whole input at every output position.
    let weights = tape.leaf(Tensor::filled([1, 1, 3, 3], 1.0));
    let bias = tape.leaf(Tensor::filled([1], 0.0));

    let output = conv2d(input, weights, bias, 1, 1);
    assert_eq!(output.shape(), Shape::new([1, 1, 2, 2]));

    let output = output.symbol();
    let network = tape.into_network();
    let run = network.forward(&network.parameters(), []);
    assert_eq!(run.of(output).to_vec(), &[10.0; 4]);
}

#[test]
fn conv2d_gradients_flow_to_every_operand() {
    let tape: Tape<f64> = Tape::new();
    let input = tape.leaf(Tensor::new(
        [1, 1, 3, 3],
        (1..=9).map(|v| v as f64).collect::<Vec<_>>(),
    ));
    let weights = tape.leaf(Tensor::filled([1, 1, 2, 2], 1.0));
    let bias = tape.leaf(Tensor::filled([1], 0.0));

    let output = conv2d(input, weights, bias, 1, 0);
    let loss = output.sum();

    let (input, weights, bias, loss) = (
        input.symbol(),
        weights.symbol(),
        bias.symbol(),
        loss.symbol(),
    );
    let network = tape.into_network();
    let run = network.forward(&network.parameters(), []);
    let gradients = run.backward(loss);
    // Each input position is graded by how many windows cover it.
    assert_eq!(
        gradients.of(input).to_vec(),
        &[1.0, 2.0, 1.0, 2.0, 4.0, 2.0, 1.0, 2.0, 1.0]
    );
    // Each kernel position sums the input values it was multiplied by.
    assert_eq!(gradients.of(weights).to_vec(), &[12.0, 16.0, 24.0, 28.0]);
    // The bias sees the seed once per output position.
    assert_eq!(gradients.of(bias).to_vec(), &[4.0]);
}

#[test]
#[should_panic(expected = "disagree on channels")]
fn conv2d_rejects_mismatched_channels() {
    let tape: Tape<f64> = Tape::new();
    let input = tape.leaf(Tensor::filled([1, 2, 3, 3], 0.0));
    let weights = tape.leaf(Tensor::filled([1, 3, 2, 2], 0.0));
    let bias = tape.leaf(Tensor::filled([1], 0.0));
    conv2d(input, weights, bias, 1, 0);
}

#[test]
fn new_allocates_weights_and_bias() {
    let tape = Tape::new();
    let layer = Conv2d::new(
        &tape,
        Tensor::filled([4, 2, 3, 3], 0.0_f64),
        Tensor::filled([4], 0.0),
        1,
        1,
    );
    // One kernel stack and one bias vector, regardless of size.
    assert_eq!(tape.len(), 2);
    assert_eq!(layer.parameters().count(), 2);
}

#[test]
#[should_panic(expected = "must be rank 4")]
fn new_rejects_non_stack_weights() {
    let tape = Tape::new();
    Conv2d::new(
        &tape,
        Tensor::filled([4, 2, 3], 0.0_f64),
        Tensor::filled([4], 0.0),
        1,
        0,
    );
}

#[test]
#[should_panic(expected = "disagree on filters")]
fn new_rejects_mismatched_bias() {
    let tape = Tape::new();
    Conv2d::new(
        &tape,
        Tensor::filled([4, 2, 3, 3], 0.0_f64),
        Tensor::filled([3], 0.0),
        1,
        0,
    );
}

#[test]
#[should_panic(expected = "stride must be positive")]
fn new_rejects_a_zero_stride() {
    let tape = Tape::new();
    Conv2d::new(
        &tape,
        Tensor::filled([4, 2, 3, 3], 0.0_f64),
        Tensor::filled([4], 0.0),
        0,
        0,
    );
}

#[test]
fn express_records_tensor_granularity() {
    let tape = Tape::new();
    let layer = Conv2d::new(
        &tape,
        Tensor::new([1, 1, 2, 2], [1.0, 0.0, 0.0, 1.0]),
        Tensor::new([1], [10.0]),
        1,
        1,
    );
    let input = tape.leaf(Tensor::new(
        [1, 1, 3, 3],
        (1..=9).map(|v| v as f64).collect::<Vec<_>>(),
    ));
    let nodes_before = tape.len();

    let output = layer.express(input);

    // Two pads, two unfolds, the patch permute + reshape, the weight
    // permute + reshape, the product, the bias broadcast and shift, and
    // the output reshape + permute: thirteen nodes, fixed regardless of
    // any size.
    assert_eq!(tape.len(), nodes_before + 13);
    assert_eq!(output.shape(), Shape::new([1, 1, 4, 4]));
}