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

use yscv_tensor::Tensor;

use crate::Graph;

#[test]
fn backward_relu_masks_negative_inputs() {
    let mut graph = Graph::new();
    let x = graph.variable(Tensor::from_vec(vec![4], vec![-1.0, 2.0, -3.0, 4.0]).unwrap());
    let y = graph.relu(x).unwrap();
    let loss = graph.sum(y).unwrap();
    graph.backward(loss).unwrap();

    let x_grad = graph.grad(x).unwrap().unwrap();
    assert_eq!(x_grad.data(), &[0.0, 1.0, 0.0, 1.0]);
}

#[test]
fn backward_exp_multiplies_by_exp_output() {
    let mut graph = Graph::new();
    let x = graph.variable(Tensor::from_vec(vec![2], vec![0.0, 1.0]).unwrap());
    let y = graph.exp(x).unwrap();
    let loss = graph.sum(y).unwrap();
    graph.backward(loss).unwrap();

    // d(exp(x))/dx = exp(x)
    let x_grad = graph.grad(x).unwrap().unwrap();
    assert!((x_grad.data()[0] - 1.0).abs() < 1e-5); // exp(0)=1
    assert!((x_grad.data()[1] - std::f32::consts::E).abs() < 1e-4);
}

#[test]
fn backward_log_divides_by_input() {
    let mut graph = Graph::new();
    let x = graph.variable(Tensor::from_vec(vec![2], vec![1.0, 2.0]).unwrap());
    let y = graph.log(x).unwrap();
    let loss = graph.sum(y).unwrap();
    graph.backward(loss).unwrap();

    // d(ln(x))/dx = 1/x
    let x_grad = graph.grad(x).unwrap().unwrap();
    assert!((x_grad.data()[0] - 1.0).abs() < 1e-5);
    assert!((x_grad.data()[1] - 0.5).abs() < 1e-5);
}

#[test]
fn backward_sqrt_computes_half_reciprocal_sqrt() {
    let mut graph = Graph::new();
    let x = graph.variable(Tensor::from_vec(vec![2], vec![4.0, 9.0]).unwrap());
    let y = graph.sqrt(x).unwrap();
    let loss = graph.sum(y).unwrap();
    graph.backward(loss).unwrap();

    // d(sqrt(x))/dx = 0.5 / sqrt(x)
    let x_grad = graph.grad(x).unwrap().unwrap();
    assert!((x_grad.data()[0] - 0.25).abs() < 1e-5); // 0.5/2
    assert!((x_grad.data()[1] - 1.0 / 6.0).abs() < 1e-5); // 0.5/3
}

#[test]
fn backward_sigmoid_computes_s_times_1_minus_s() {
    let mut graph = Graph::new();
    let x = graph.variable(Tensor::from_vec(vec![2], vec![0.0, 2.0]).unwrap());
    let y = graph.sigmoid(x).unwrap();
    let loss = graph.sum(y).unwrap();
    graph.backward(loss).unwrap();

    let y_val = graph.value(y).unwrap();
    let x_grad = graph.grad(x).unwrap().unwrap();
    for i in 0..2 {
        let s = y_val.data()[i];
        let expected = s * (1.0 - s);
        assert!(
            (x_grad.data()[i] - expected).abs() < 1e-5,
            "sigmoid grad mismatch at {i}"
        );
    }
}

#[test]
fn backward_tanh_computes_1_minus_tanh_sq() {
    let mut graph = Graph::new();
    let x = graph.variable(Tensor::from_vec(vec![2], vec![0.0, 1.0]).unwrap());
    let y = graph.tanh(x).unwrap();
    let loss = graph.sum(y).unwrap();
    graph.backward(loss).unwrap();

    let y_val = graph.value(y).unwrap();
    let x_grad = graph.grad(x).unwrap().unwrap();
    for i in 0..2 {
        let t = y_val.data()[i];
        let expected = 1.0 - t * t;
        assert!(
            (x_grad.data()[i] - expected).abs() < 1e-5,
            "tanh grad mismatch at {i}"
        );
    }
}

#[test]
fn leaky_relu_forward_and_backward() {
    let mut g = Graph::new();
    let x = g.variable(Tensor::from_vec(vec![3], vec![-2.0, 0.0, 3.0]).unwrap());
    let y = g.leaky_relu(x, 0.1).unwrap();
    let loss = g.sum(y).unwrap();
    g.backward(loss).unwrap();
    assert!((g.value(y).unwrap().data()[0] - (-0.2)).abs() < 1e-6);
    assert!((g.value(y).unwrap().data()[2] - 3.0).abs() < 1e-6);
    let grad = g.grad(x).unwrap().unwrap();
    assert!((grad.data()[0] - 0.1).abs() < 1e-6);
    assert!((grad.data()[2] - 1.0).abs() < 1e-6);
}

#[test]
fn gelu_forward_and_backward() {
    let mut g = Graph::new();
    let x = g.variable(Tensor::from_vec(vec![3], vec![-1.0, 0.0, 1.0]).unwrap());
    let y = g.gelu(x).unwrap();
    let loss = g.sum(y).unwrap();
    g.backward(loss).unwrap();

    // GELU(0) = 0
    let y_val = g.value(y).unwrap();
    assert!((y_val.data()[1] - 0.0).abs() < 1e-6);

    // Check gradient via finite differences
    let grad = g.grad(x).unwrap().unwrap();
    let eps = 1e-4_f32;
    for i in 0..3 {
        let xi = [-1.0_f32, 0.0, 1.0][i];
        let gelu_plus = {
            let xp = xi + eps;
            xp * (1.0 / (1.0 + (-1.702 * xp).exp()))
        };
        let gelu_minus = {
            let xm = xi - eps;
            xm * (1.0 / (1.0 + (-1.702 * xm).exp()))
        };
        let numerical = (gelu_plus - gelu_minus) / (2.0 * eps);
        assert!(
            (grad.data()[i] - numerical).abs() < 1e-3,
            "gelu grad mismatch at {i}: got {} expected {}",
            grad.data()[i],
            numerical
        );
    }
}

#[test]
fn silu_forward_and_backward() {
    let mut g = Graph::new();
    let x = g.variable(Tensor::from_vec(vec![3], vec![-1.0, 0.0, 1.0]).unwrap());
    let y = g.silu(x).unwrap();
    let loss = g.sum(y).unwrap();
    g.backward(loss).unwrap();

    // SiLU(0) = 0
    let y_val = g.value(y).unwrap();
    assert!((y_val.data()[1] - 0.0).abs() < 1e-6);

    // Check gradient: SiLU'(x) = sigmoid(x) + x * sigmoid(x) * (1 - sigmoid(x))
    let grad = g.grad(x).unwrap().unwrap();
    for i in 0..3 {
        let xi = [-1.0_f32, 0.0, 1.0][i];
        let s = 1.0 / (1.0 + (-xi).exp());
        let expected = s + xi * s * (1.0 - s);
        assert!(
            (grad.data()[i] - expected).abs() < 1e-5,
            "silu grad mismatch at {i}: got {} expected {}",
            grad.data()[i],
            expected
        );
    }
}

#[test]
fn mish_forward_and_backward() {
    let mut g = Graph::new();
    let x = g.variable(Tensor::from_vec(vec![3], vec![-1.0, 0.0, 1.0]).unwrap());
    let y = g.mish(x).unwrap();
    let loss = g.sum(y).unwrap();
    g.backward(loss).unwrap();

    // Mish(0) = 0 * tanh(ln(2)) = 0
    let y_val = g.value(y).unwrap();
    assert!((y_val.data()[1] - 0.0).abs() < 1e-6);

    // Check gradient via finite differences
    let grad = g.grad(x).unwrap().unwrap();
    let eps = 1e-4_f32;
    for i in 0..3 {
        let xi = [-1.0_f32, 0.0, 1.0][i];
        let mish_val = |x: f32| -> f32 {
            let sp = (1.0 + x.exp()).ln();
            x * sp.tanh()
        };
        let numerical = (mish_val(xi + eps) - mish_val(xi - eps)) / (2.0 * eps);
        assert!(
            (grad.data()[i] - numerical).abs() < 1e-3,
            "mish grad mismatch at {i}: got {} expected {}",
            grad.data()[i],
            numerical
        );
    }
}

#[test]
fn softmax_forward_and_backward() {
    let mut g = Graph::new();
    let x = g.variable(Tensor::from_vec(vec![1, 3], vec![1.0, 2.0, 3.0]).unwrap());
    let sm = g.softmax(x).unwrap();
    let loss = g.sum(sm).unwrap();
    g.backward(loss).unwrap();
    // softmax sums to 1 per row
    let sm_val = g.value(sm).unwrap();
    let sum: f32 = sm_val.data().iter().sum();
    assert!((sum - 1.0).abs() < 1e-5);
    // gradients should exist
    let grad = g.grad(x).unwrap().unwrap();
    assert_eq!(grad.shape(), &[1, 3]);
}