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
223
224
225
226
227
228
229
230
231

use yscv_tensor::Tensor;

use crate::Graph;

#[test]
fn backward_conv2d_nhwc_computes_weight_and_input_grads() {
    let mut graph = Graph::new();
    // input: [1, 3, 3, 1], weight: [2, 2, 1, 1], no bias, stride=1
    let input_data: Vec<f32> = (1..=9).map(|v| v as f32).collect();
    let input = graph.variable(Tensor::from_vec(vec![1, 3, 3, 1], input_data).unwrap());
    let weight =
        graph.variable(Tensor::from_vec(vec![2, 2, 1, 1], vec![1.0, 0.0, 0.0, -1.0]).unwrap());

    let out = graph.conv2d_nhwc(input, weight, None, 1, 1).unwrap();
    // out shape: [1, 2, 2, 1]
    assert_eq!(graph.value(out).unwrap().shape(), &[1, 2, 2, 1]);

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

    let w_grad = graph.grad(weight).unwrap().unwrap();
    assert_eq!(w_grad.shape(), &[2, 2, 1, 1]);

    let i_grad = graph.grad(input).unwrap().unwrap();
    assert_eq!(i_grad.shape(), &[1, 3, 3, 1]);

    // Numerical gradient check for weight:
    // grad_w[kh,kw,0,0] = sum_{oh,ow} input[oh+kh, ow+kw, 0]
    // For kh=0,kw=0: input at (0,0),(0,1),(1,0),(1,1) = 1+2+4+5 = 12
    // For kh=0,kw=1: (0,1),(0,2),(1,1),(1,2) = 2+3+5+6 = 16
    // For kh=1,kw=0: (1,0),(1,1),(2,0),(2,1) = 4+5+7+8 = 24
    // For kh=1,kw=1: (1,1),(1,2),(2,1),(2,2) = 5+6+8+9 = 28
    let expected_wg = [12.0, 16.0, 24.0, 28.0];
    for (i, &e) in expected_wg.iter().enumerate() {
        assert!(
            (w_grad.data()[i] - e).abs() < 1e-4,
            "weight grad mismatch at {i}: got {}, expected {e}",
            w_grad.data()[i]
        );
    }
}

#[test]
fn backward_conv2d_nhwc_with_bias_produces_bias_grad() {
    let mut graph = Graph::new();
    let input = graph.variable(Tensor::filled(vec![1, 3, 3, 2], 1.0).unwrap());
    let weight = graph.variable(Tensor::filled(vec![2, 2, 2, 1], 0.5).unwrap());
    let bias = graph.variable(Tensor::from_vec(vec![1], vec![0.1]).unwrap());

    let out = graph.conv2d_nhwc(input, weight, Some(bias), 1, 1).unwrap();
    assert_eq!(graph.value(out).unwrap().shape(), &[1, 2, 2, 1]);

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

    // bias grad = number of output spatial elements = 2*2 = 4
    let b_grad = graph.grad(bias).unwrap().unwrap();
    assert_eq!(b_grad.shape(), &[1]);
    assert!((b_grad.data()[0] - 4.0).abs() < 1e-4);
}

#[test]
fn backward_conv2d_nhwc_stride2_shape_and_grad() {
    let mut graph = Graph::new();
    let input = graph.variable(Tensor::filled(vec![1, 4, 4, 1], 1.0).unwrap());
    let weight = graph.variable(Tensor::filled(vec![2, 2, 1, 1], 1.0).unwrap());

    let out = graph.conv2d_nhwc(input, weight, None, 2, 2).unwrap();
    // OH = (4-2)/2+1 = 2, OW = 2
    assert_eq!(graph.value(out).unwrap().shape(), &[1, 2, 2, 1]);

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

    let i_grad = graph.grad(input).unwrap().unwrap();
    assert_eq!(i_grad.shape(), &[1, 4, 4, 1]);
    // Each input pixel contributes to at most one output with stride=2 and kernel=2
    for &g in i_grad.data() {
        assert!((0.0..=1.0 + 1e-6).contains(&g));
    }
}

#[test]
fn backward_conv2d_nhwc_numerical_gradient_check() {
    let eps = 1e-3;
    let input_data: Vec<f32> = vec![
        0.5, 1.2, 0.3, 0.8, 1.0, 0.7, 0.9, 1.5, 0.2, 0.6, 1.1, 0.4, 0.8, 0.3, 0.7, 1.0,
    ];
    let weight_data: Vec<f32> = vec![0.1, -0.2, 0.3, -0.4];

    // Compute analytical gradients
    let mut graph = Graph::new();
    let input = graph.variable(Tensor::from_vec(vec![1, 4, 4, 1], input_data.clone()).unwrap());
    let weight = graph.variable(Tensor::from_vec(vec![2, 2, 1, 1], weight_data.clone()).unwrap());
    let out = graph.conv2d_nhwc(input, weight, None, 1, 1).unwrap();
    let loss = graph.sum(out).unwrap();
    graph.backward(loss).unwrap();
    let analytic_wg = graph.grad(weight).unwrap().unwrap().data().to_vec();
    let analytic_ig = graph.grad(input).unwrap().unwrap().data().to_vec();

    // Numerical gradient check for weight
    for wi in 0..4 {
        let mut wp = weight_data.clone();
        wp[wi] += eps;
        let mut wm = weight_data.clone();
        wm[wi] -= eps;

        let mut gp = Graph::new();
        let ip = gp.variable(Tensor::from_vec(vec![1, 4, 4, 1], input_data.clone()).unwrap());
        let wp_node = gp.variable(Tensor::from_vec(vec![2, 2, 1, 1], wp).unwrap());
        let op = gp.conv2d_nhwc(ip, wp_node, None, 1, 1).unwrap();
        let lp = gp.sum(op).unwrap();
        let loss_p = gp.value(lp).unwrap().data()[0];

        let mut gm = Graph::new();
        let im = gm.variable(Tensor::from_vec(vec![1, 4, 4, 1], input_data.clone()).unwrap());
        let wm_node = gm.variable(Tensor::from_vec(vec![2, 2, 1, 1], wm).unwrap());
        let om = gm.conv2d_nhwc(im, wm_node, None, 1, 1).unwrap();
        let lm = gm.sum(om).unwrap();
        let loss_m = gm.value(lm).unwrap().data()[0];

        let numerical = (loss_p - loss_m) / (2.0 * eps);
        assert!(
            (analytic_wg[wi] - numerical).abs() < 1e-2,
            "weight numerical grad mismatch at {wi}: analytic={}, numerical={numerical}",
            analytic_wg[wi]
        );
    }

    // Numerical gradient check for input (spot check first 4 elements)
    for ii in 0..4 {
        let mut ip_data = input_data.clone();
        ip_data[ii] += eps;
        let mut im_data = input_data.clone();
        im_data[ii] -= eps;

        let mut gp = Graph::new();
        let ip = gp.variable(Tensor::from_vec(vec![1, 4, 4, 1], ip_data).unwrap());
        let wp = gp.variable(Tensor::from_vec(vec![2, 2, 1, 1], weight_data.clone()).unwrap());
        let op = gp.conv2d_nhwc(ip, wp, None, 1, 1).unwrap();
        let lp = gp.sum(op).unwrap();
        let loss_p = gp.value(lp).unwrap().data()[0];

        let mut gm = Graph::new();
        let im = gm.variable(Tensor::from_vec(vec![1, 4, 4, 1], im_data).unwrap());
        let wm = gm.variable(Tensor::from_vec(vec![2, 2, 1, 1], weight_data.clone()).unwrap());
        let om = gm.conv2d_nhwc(im, wm, None, 1, 1).unwrap();
        let lm = gm.sum(om).unwrap();
        let loss_m = gm.value(lm).unwrap().data()[0];

        let numerical = (loss_p - loss_m) / (2.0 * eps);
        assert!(
            (analytic_ig[ii] - numerical).abs() < 1e-2,
            "input numerical grad mismatch at {ii}: analytic={}, numerical={numerical}",
            analytic_ig[ii]
        );
    }
}

#[test]
fn end_to_end_mini_cnn_forward_backward() {
    let mut graph = Graph::new();

    // Build a tiny CNN: conv2d -> relu -> avgpool -> flatten -> matmul -> sum
    // Input: [1, 4, 4, 1]
    let input_data: Vec<f32> = (0..16).map(|v| v as f32 / 16.0).collect();
    let input = graph.variable(Tensor::from_vec(vec![1, 4, 4, 1], input_data).unwrap());

    // Conv: [2,2,1,2] -> 2 output channels
    let w_conv = graph.variable(
        Tensor::from_vec(
            vec![2, 2, 1, 2],
            vec![0.1, -0.1, 0.2, 0.2, -0.1, 0.3, 0.15, -0.15],
        )
        .unwrap(),
    );
    let conv_out = graph.conv2d_nhwc(input, w_conv, None, 1, 1).unwrap();
    // [1, 3, 3, 2]
    assert_eq!(graph.value(conv_out).unwrap().shape(), &[1, 3, 3, 2]);

    let relu_out = graph.relu(conv_out).unwrap();

    // AvgPool 3x3 stride 1 -> [1,1,1,2]
    let pool_out = graph.avg_pool2d_nhwc(relu_out, 3, 3, 1, 1).unwrap();
    assert_eq!(graph.value(pool_out).unwrap().shape(), &[1, 1, 1, 2]);

    let flat = graph.flatten(pool_out).unwrap();
    assert_eq!(graph.value(flat).unwrap().shape(), &[1, 2]);

    // Linear: [2, 1]
    let w_fc = graph.variable(Tensor::from_vec(vec![2, 1], vec![1.0, -1.0]).unwrap());
    let logit = graph.matmul_2d(flat, w_fc).unwrap();
    assert_eq!(graph.value(logit).unwrap().shape(), &[1, 1]);

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

    // All trainable nodes should have gradients
    assert!(graph.grad(input).unwrap().is_some());
    assert!(graph.grad(w_conv).unwrap().is_some());
    assert!(graph.grad(w_fc).unwrap().is_some());

    // Sanity: gradients should be finite
    for &g in graph.grad(w_conv).unwrap().unwrap().data() {
        assert!(g.is_finite(), "conv weight grad not finite: {g}");
    }
    for &g in graph.grad(w_fc).unwrap().unwrap().data() {
        assert!(g.is_finite(), "fc weight grad not finite: {g}");
    }
}

#[test]
fn depthwise_conv2d_nhwc_forward_and_backward() {
    let mut g = Graph::new();
    // Input: [1, 3, 3, 2] (batch=1, 3x3, 2 channels)
    let input_data: Vec<f32> = (0..18).map(|i| i as f32).collect();
    let x = g.variable(Tensor::from_vec(vec![1, 3, 3, 2], input_data).unwrap());
    // Weight: [2, 2, 2, 1] (2x2 kernel, 2 channels, depth_multiplier=1)
    let w = g.variable(Tensor::filled(vec![2, 2, 2, 1], 1.0).unwrap());
    let out = g.depthwise_conv2d_nhwc(x, w, None, 1, 1).unwrap();
    let out_shape = g.value(out).unwrap().shape().to_vec();
    // Output: [1, 2, 2, 2]
    assert_eq!(out_shape, &[1, 2, 2, 2]);
    // Verify backward runs without error
    let loss = g.sum(out).unwrap();
    g.backward(loss).unwrap();
    let grad_x = g.grad(x).unwrap().unwrap();
    let grad_w = g.grad(w).unwrap().unwrap();
    assert_eq!(grad_x.shape(), &[1, 3, 3, 2]);
    assert_eq!(grad_w.shape(), &[2, 2, 2, 1]);
}