yscv-autograd 0.1.9

Dynamic computation graph with tape-based reverse-mode autodiff
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

use yscv_tensor::Tensor;

use crate::Graph;

#[test]
fn test_scatter_forward() {
    let mut g = Graph::new();
    // input: 4x3 matrix
    let input = g.variable(
        Tensor::from_vec(
            vec![4, 3],
            vec![
                1.0, 2.0, 3.0, // row 0
                4.0, 5.0, 6.0, // row 1
                7.0, 8.0, 9.0, // row 2
                10.0, 11.0, 12.0, // row 3
            ],
        )
        .unwrap(),
    );
    // indices: scatter into rows 1 and 3
    let indices = g.constant(Tensor::from_vec(vec![2], vec![1.0, 3.0]).unwrap());
    // src: 2x3 replacement rows
    let src = g.variable(
        Tensor::from_vec(vec![2, 3], vec![100.0, 200.0, 300.0, 400.0, 500.0, 600.0]).unwrap(),
    );

    let result = g.scatter(input, indices, src).unwrap();
    let val = g.value(result).unwrap();
    assert_eq!(val.shape(), &[4, 3]);
    assert_eq!(
        val.data(),
        &[
            1.0, 2.0, 3.0, // row 0 unchanged
            100.0, 200.0, 300.0, // row 1 replaced
            7.0, 8.0, 9.0, // row 2 unchanged
            400.0, 500.0, 600.0, // row 3 replaced
        ]
    );
}

#[test]
fn test_embedding_lookup_forward() {
    let mut g = Graph::new();
    // weight: 4x3 embedding matrix (vocab=4, dim=3)
    let weight = g.variable(
        Tensor::from_vec(
            vec![4, 3],
            vec![
                0.1, 0.2, 0.3, // row 0
                0.4, 0.5, 0.6, // row 1
                0.7, 0.8, 0.9, // row 2
                1.0, 1.1, 1.2, // row 3
            ],
        )
        .unwrap(),
    );
    let indices = g.constant(Tensor::from_vec(vec![3], vec![2.0, 0.0, 3.0]).unwrap());

    let result = g.embedding_lookup(weight, indices).unwrap();
    let val = g.value(result).unwrap();
    assert_eq!(val.shape(), &[3, 3]);
    let data = val.data();
    // row 0 of output = weight[2]
    assert!((data[0] - 0.7).abs() < 1e-6);
    assert!((data[1] - 0.8).abs() < 1e-6);
    assert!((data[2] - 0.9).abs() < 1e-6);
    // row 1 of output = weight[0]
    assert!((data[3] - 0.1).abs() < 1e-6);
    assert!((data[4] - 0.2).abs() < 1e-6);
    assert!((data[5] - 0.3).abs() < 1e-6);
    // row 2 of output = weight[3]
    assert!((data[6] - 1.0).abs() < 1e-6);
    assert!((data[7] - 1.1).abs() < 1e-6);
    assert!((data[8] - 1.2).abs() < 1e-6);
}

#[test]
fn test_embedding_lookup_backward() {
    let mut g = Graph::new();
    // weight: 4x3 embedding matrix
    let weight = g.variable(
        Tensor::from_vec(
            vec![4, 3],
            vec![0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 1.0, 1.1, 1.2],
        )
        .unwrap(),
    );
    let indices = g.constant(Tensor::from_vec(vec![2], vec![1.0, 3.0]).unwrap());

    let lookup = g.embedding_lookup(weight, indices).unwrap();
    let loss = g.sum(lookup).unwrap();
    g.backward(loss).unwrap();

    let grad = g.grad(weight).unwrap().unwrap();
    assert_eq!(grad.shape(), &[4, 3]);
    let gd = grad.data();
    // row 0: not looked up => grad = 0
    assert!((gd[0]).abs() < 1e-6);
    assert!((gd[1]).abs() < 1e-6);
    assert!((gd[2]).abs() < 1e-6);
    // row 1: looked up once => grad = 1 (from sum backward)
    assert!((gd[3] - 1.0).abs() < 1e-6);
    assert!((gd[4] - 1.0).abs() < 1e-6);
    assert!((gd[5] - 1.0).abs() < 1e-6);
    // row 2: not looked up => grad = 0
    assert!((gd[6]).abs() < 1e-6);
    assert!((gd[7]).abs() < 1e-6);
    assert!((gd[8]).abs() < 1e-6);
    // row 3: looked up once => grad = 1
    assert!((gd[9] - 1.0).abs() < 1e-6);
    assert!((gd[10] - 1.0).abs() < 1e-6);
    assert!((gd[11] - 1.0).abs() < 1e-6);
}

#[test]
fn test_embedding_training() {
    let mut g = Graph::new();
    // vocab=5, dim=3
    let weight = g.variable(
        Tensor::from_vec(
            vec![5, 3],
            vec![
                1.0, 2.0, 3.0, // row 0
                4.0, 5.0, 6.0, // row 1
                7.0, 8.0, 9.0, // row 2
                10.0, 11.0, 12.0, // row 3
                13.0, 14.0, 15.0, // row 4
            ],
        )
        .unwrap(),
    );
    // Look up indices [0, 2]
    let indices = g.constant(Tensor::from_vec(vec![2], vec![0.0, 2.0]).unwrap());

    let lookup = g.embedding_lookup(weight, indices).unwrap();
    // loss = sum of all looked-up values
    let loss = g.sum(lookup).unwrap();
    g.backward(loss).unwrap();

    let grad = g.grad(weight).unwrap().unwrap();
    assert_eq!(grad.shape(), &[5, 3]);
    let gd = grad.data();

    // row 0: looked up => grad = 1
    assert!((gd[0] - 1.0).abs() < 1e-6);
    assert!((gd[1] - 1.0).abs() < 1e-6);
    assert!((gd[2] - 1.0).abs() < 1e-6);
    // row 1: not looked up => grad = 0
    assert!((gd[3]).abs() < 1e-6);
    assert!((gd[4]).abs() < 1e-6);
    assert!((gd[5]).abs() < 1e-6);
    // row 2: looked up => grad = 1
    assert!((gd[6] - 1.0).abs() < 1e-6);
    assert!((gd[7] - 1.0).abs() < 1e-6);
    assert!((gd[8] - 1.0).abs() < 1e-6);
    // row 3: not looked up => grad = 0
    assert!((gd[9]).abs() < 1e-6);
    assert!((gd[10]).abs() < 1e-6);
    assert!((gd[11]).abs() < 1e-6);
    // row 4: not looked up => grad = 0
    assert!((gd[12]).abs() < 1e-6);
    assert!((gd[13]).abs() < 1e-6);
    assert!((gd[14]).abs() < 1e-6);
}