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
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

use yscv_tensor::Tensor;

use crate::Graph;

#[test]
fn layer_norm_forward_and_backward() {
    let mut g = Graph::new();
    // Input: [2, 3] (two rows, each normalized over 3 features)
    let x = g.variable(Tensor::from_vec(vec![2, 3], vec![1.0, 2.0, 3.0, 4.0, 5.0, 6.0]).unwrap());
    let gamma = g.variable(Tensor::from_vec(vec![3], vec![1.0, 1.0, 1.0]).unwrap());
    let beta = g.variable(Tensor::from_vec(vec![3], vec![0.0, 0.0, 0.0]).unwrap());

    let out = g.layer_norm(x, gamma, beta, 1e-5).unwrap();
    assert_eq!(g.value(out).unwrap().shape(), &[2, 3]);

    // Each row should be normalized: mean ~ 0, var ~ 1
    let out_data = g.value(out).unwrap().data().to_vec();
    for row in 0..2 {
        let base = row * 3;
        let row_mean: f32 = out_data[base..base + 3].iter().sum::<f32>() / 3.0;
        assert!(row_mean.abs() < 1e-5, "row {row} mean = {row_mean}");
    }

    let loss = g.sum(out).unwrap();
    g.backward(loss).unwrap();

    // grad_beta = sum of upstream (all 1s) over 2 rows => [2, 2, 2]
    let beta_grad = g.grad(beta).unwrap().unwrap();
    assert_eq!(beta_grad.shape(), &[3]);
    for &v in beta_grad.data() {
        assert!((v - 2.0).abs() < 1e-4, "beta grad {v} not ~ 2.0");
    }

    let gamma_grad = g.grad(gamma).unwrap().unwrap();
    assert_eq!(gamma_grad.shape(), &[3]);

    let input_grad = g.grad(x).unwrap().unwrap();
    assert_eq!(input_grad.shape(), &[2, 3]);
    // Each row's input gradient should sum to ~0 (property of layer norm backward)
    for row in 0..2 {
        let base = row * 3;
        let row_sum: f32 = input_grad.data()[base..base + 3].iter().sum();
        assert!(
            row_sum.abs() < 1e-4,
            "row {row} input grad sum = {row_sum}, expected ~0"
        );
    }
}

#[test]
fn layer_norm_backward_finite_differences() {
    let eps = 1e-4_f32;
    let ln_eps = 1e-5_f32;
    let input_data = vec![1.5, -0.5, 2.0, 0.5];

    // Compute analytical gradients
    let mut g = Graph::new();
    let x = g.variable(Tensor::from_vec(vec![2, 2], input_data.clone()).unwrap());
    let gamma = g.variable(Tensor::from_vec(vec![2], vec![1.0, 1.0]).unwrap());
    let beta = g.variable(Tensor::from_vec(vec![2], vec![0.0, 0.0]).unwrap());
    let out = g.layer_norm(x, gamma, beta, ln_eps).unwrap();
    let loss = g.sum(out).unwrap();
    g.backward(loss).unwrap();
    let analytical = g.grad(x).unwrap().unwrap().data().to_vec();

    // Compute numerical gradients via finite differences
    for idx in 0..input_data.len() {
        let mut data_plus = input_data.clone();
        data_plus[idx] += eps;
        let mut g2 = Graph::new();
        let xp = g2.variable(Tensor::from_vec(vec![2, 2], data_plus).unwrap());
        let gp = g2.variable(Tensor::from_vec(vec![2], vec![1.0, 1.0]).unwrap());
        let bp = g2.variable(Tensor::from_vec(vec![2], vec![0.0, 0.0]).unwrap());
        let op = g2.layer_norm(xp, gp, bp, ln_eps).unwrap();
        let lp = g2.sum(op).unwrap();
        let loss_plus = g2.value(lp).unwrap().data()[0];

        let mut data_minus = input_data.clone();
        data_minus[idx] -= eps;
        let mut g3 = Graph::new();
        let xm = g3.variable(Tensor::from_vec(vec![2, 2], data_minus).unwrap());
        let gm = g3.variable(Tensor::from_vec(vec![2], vec![1.0, 1.0]).unwrap());
        let bm = g3.variable(Tensor::from_vec(vec![2], vec![0.0, 0.0]).unwrap());
        let om = g3.layer_norm(xm, gm, bm, ln_eps).unwrap();
        let lm = g3.sum(om).unwrap();
        let loss_minus = g3.value(lm).unwrap().data()[0];

        let numerical = (loss_plus - loss_minus) / (2.0 * eps);
        assert!(
            (analytical[idx] - numerical).abs() < 1e-3,
            "layer_norm grad mismatch at {idx}: analytical={}, numerical={}",
            analytical[idx],
            numerical
        );
    }
}

#[test]
fn group_norm_forward_and_backward() {
    let mut g = Graph::new();
    // [1, 2, 2, 4] input, 2 groups of 2 channels
    let x = g.variable(
        Tensor::from_vec(
            vec![1, 2, 2, 4],
            vec![
                1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0, 9.0, 10.0, 11.0, 12.0, 13.0, 14.0, 15.0,
                16.0,
            ],
        )
        .unwrap(),
    );
    let gamma = g.variable(Tensor::from_vec(vec![4], vec![1.0, 1.0, 1.0, 1.0]).unwrap());
    let beta = g.variable(Tensor::from_vec(vec![4], vec![0.0, 0.0, 0.0, 0.0]).unwrap());

    let out = g.group_norm(x, gamma, beta, 2, 1e-5).unwrap();
    assert_eq!(g.value(out).unwrap().shape(), &[1, 2, 2, 4]);

    let loss = g.sum(out).unwrap();
    g.backward(loss).unwrap();

    let beta_grad = g.grad(beta).unwrap().unwrap();
    assert_eq!(beta_grad.shape(), &[4]);
    // grad_beta[c] = number of spatial positions = 4 (2x2)
    for &v in beta_grad.data() {
        assert!((v - 4.0).abs() < 1e-4, "beta grad {v} not ~ 4.0");
    }

    let gamma_grad = g.grad(gamma).unwrap().unwrap();
    assert_eq!(gamma_grad.shape(), &[4]);

    let input_grad = g.grad(x).unwrap().unwrap();
    assert_eq!(input_grad.shape(), &[1, 2, 2, 4]);
}

#[test]
fn group_norm_backward_finite_differences() {
    let eps = 1e-4_f32;
    let gn_eps = 1e-5_f32;
    let input_data: Vec<f32> = (1..=16).map(|i| i as f32 * 0.1).collect();

    // Compute analytical gradients
    let mut g = Graph::new();
    let x = g.variable(Tensor::from_vec(vec![1, 2, 2, 4], input_data.clone()).unwrap());
    let gamma = g.variable(Tensor::from_vec(vec![4], vec![1.0, 1.0, 1.0, 1.0]).unwrap());
    let beta = g.variable(Tensor::from_vec(vec![4], vec![0.0, 0.0, 0.0, 0.0]).unwrap());
    let out = g.group_norm(x, gamma, beta, 2, gn_eps).unwrap();
    let loss = g.sum(out).unwrap();
    g.backward(loss).unwrap();
    let analytical = g.grad(x).unwrap().unwrap().data().to_vec();

    // Compute numerical gradients via finite differences
    for idx in 0..input_data.len() {
        let mut data_plus = input_data.clone();
        data_plus[idx] += eps;
        let mut g2 = Graph::new();
        let xp = g2.variable(Tensor::from_vec(vec![1, 2, 2, 4], data_plus).unwrap());
        let gp = g2.variable(Tensor::from_vec(vec![4], vec![1.0, 1.0, 1.0, 1.0]).unwrap());
        let bp = g2.variable(Tensor::from_vec(vec![4], vec![0.0, 0.0, 0.0, 0.0]).unwrap());
        let op = g2.group_norm(xp, gp, bp, 2, gn_eps).unwrap();
        let lp = g2.sum(op).unwrap();
        let loss_plus = g2.value(lp).unwrap().data()[0];

        let mut data_minus = input_data.clone();
        data_minus[idx] -= eps;
        let mut g3 = Graph::new();
        let xm = g3.variable(Tensor::from_vec(vec![1, 2, 2, 4], data_minus).unwrap());
        let gm = g3.variable(Tensor::from_vec(vec![4], vec![1.0, 1.0, 1.0, 1.0]).unwrap());
        let bm = g3.variable(Tensor::from_vec(vec![4], vec![0.0, 0.0, 0.0, 0.0]).unwrap());
        let om = g3.group_norm(xm, gm, bm, 2, gn_eps).unwrap();
        let lm = g3.sum(om).unwrap();
        let loss_minus = g3.value(lm).unwrap().data()[0];

        let numerical = (loss_plus - loss_minus) / (2.0 * eps);
        assert!(
            (analytical[idx] - numerical).abs() < 1e-2,
            "group_norm grad mismatch at {idx}: analytical={}, numerical={}",
            analytical[idx],
            numerical
        );
    }
}

#[test]
fn backward_batch_norm2d_nhwc_computes_gamma_beta_input_grads() {
    let mut graph = Graph::new();
    // [1, 2, 2, 2] input, 2 channels
    let input = graph.variable(
        Tensor::from_vec(
            vec![1, 2, 2, 2],
            vec![1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0],
        )
        .unwrap(),
    );
    let gamma = graph.variable(Tensor::from_vec(vec![2], vec![1.0, 1.0]).unwrap());
    let beta = graph.variable(Tensor::from_vec(vec![2], vec![0.0, 0.0]).unwrap());
    let running_mean = graph.constant(Tensor::from_vec(vec![2], vec![4.0, 5.0]).unwrap());
    let running_var = graph.constant(Tensor::from_vec(vec![2], vec![4.0, 4.0]).unwrap());

    let out = graph
        .batch_norm2d_nhwc(input, gamma, beta, running_mean, running_var, 1e-5)
        .unwrap();
    assert_eq!(graph.value(out).unwrap().shape(), &[1, 2, 2, 2]);

    let loss = graph.sum(out).unwrap();
    graph.backward(loss).unwrap();

    let beta_grad = graph.grad(beta).unwrap().unwrap();
    assert_eq!(beta_grad.shape(), &[2]);
    // grad_beta[c] = number_of_spatial_elements = 4
    assert!((beta_grad.data()[0] - 4.0).abs() < 1e-4);
    assert!((beta_grad.data()[1] - 4.0).abs() < 1e-4);

    let gamma_grad = graph.grad(gamma).unwrap().unwrap();
    assert_eq!(gamma_grad.shape(), &[2]);

    let input_grad = graph.grad(input).unwrap().unwrap();
    assert_eq!(input_grad.shape(), &[1, 2, 2, 2]);
    // All input grads should be gamma / sqrt(var + eps) = 1 / sqrt(4 + 1e-5) ~ 0.5
    for &g in input_grad.data() {
        assert!((g - 0.5).abs() < 0.01, "input grad {g} not ~ 0.5");
    }
}