dfdx 0.13.0

Ergonomic auto differentiation in Rust, with pytorch like apis.
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

import torchvision
import torch
import numpy as np
with torch.no_grad():
    model = torchvision.models.resnet18(pretrained=True)
    model.eval()
    np.savez("resnet18.npz", **{
        # head
        "0.0.weight": model.conv1.weight.numpy(),
        "0.1.scale": model.bn1.weight.numpy(),
        "0.1.bias": model.bn1.bias.numpy(),
        "0.1.running_mean": model.bn1.running_mean.numpy(),
        "0.1.running_var": model.bn1.running_var.numpy(),

        # layer 1
        "1.0.0.0.weight": model.layer1[0].conv1.weight.numpy(),
        "1.0.0.1.scale": model.layer1[0].bn1.weight.numpy(),
        "1.0.0.1.bias": model.layer1[0].bn1.bias.numpy(),
        "1.0.0.1.running_mean": model.layer1[0].bn1.running_mean.numpy(),
        "1.0.0.1.running_var": model.layer1[0].bn1.running_var.numpy(),
        "1.0.0.3.weight": model.layer1[0].conv2.weight.numpy(),
        "1.0.0.4.scale": model.layer1[0].bn2.weight.numpy(),
        "1.0.0.4.bias": model.layer1[0].bn2.bias.numpy(),
        "1.0.0.4.running_mean": model.layer1[0].bn2.running_mean.numpy(),
        "1.0.0.4.running_var": model.layer1[0].bn2.running_var.numpy(),
        "1.2.0.0.weight": model.layer1[1].conv1.weight.numpy(),
        "1.2.0.1.scale": model.layer1[1].bn1.weight.numpy(),
        "1.2.0.1.bias": model.layer1[1].bn1.bias.numpy(),
        "1.2.0.1.running_mean": model.layer1[1].bn1.running_mean.numpy(),
        "1.2.0.1.running_var": model.layer1[1].bn1.running_var.numpy(),
        "1.2.0.3.weight": model.layer1[1].conv2.weight.numpy(),
        "1.2.0.4.scale": model.layer1[1].bn2.weight.numpy(),
        "1.2.0.4.bias": model.layer1[1].bn2.bias.numpy(),
        "1.2.0.4.running_mean": model.layer1[1].bn2.running_mean.numpy(),
        "1.2.0.4.running_var": model.layer1[1].bn2.running_var.numpy(),

        # layer2
        "2.0.f.0.weight": model.layer2[0].conv1.weight.numpy(),
        "2.0.f.1.scale": model.layer2[0].bn1.weight.numpy(),
        "2.0.f.1.bias": model.layer2[0].bn1.bias.numpy(),
        "2.0.f.1.running_mean": model.layer2[0].bn1.running_mean.numpy(),
        "2.0.f.1.running_var": model.layer2[0].bn1.running_var.numpy(),
        "2.0.f.3.weight": model.layer2[0].conv2.weight.numpy(),
        "2.0.f.4.scale": model.layer2[0].bn2.weight.numpy(),
        "2.0.f.4.bias": model.layer2[0].bn2.bias.numpy(),
        "2.0.f.4.running_mean": model.layer2[0].bn2.running_mean.numpy(),
        "2.0.f.4.running_var": model.layer2[0].bn2.running_var.numpy(),
        "2.0.r.0.weight": model.layer2[0].downsample[0].weight.numpy(),
        "2.0.r.1.scale": model.layer2[0].downsample[1].weight.numpy(),
        "2.0.r.1.bias": model.layer2[0].downsample[1].bias.numpy(),
        "2.0.r.1.running_mean": model.layer2[0].downsample[1].running_mean.numpy(),
        "2.0.r.1.running_var": model.layer2[0].downsample[1].running_var.numpy(),
        "2.2.0.0.weight": model.layer2[1].conv1.weight.numpy(),
        "2.2.0.1.scale": model.layer2[1].bn1.weight.numpy(),
        "2.2.0.1.bias": model.layer2[1].bn1.bias.numpy(),
        "2.2.0.1.running_mean": model.layer2[1].bn1.running_mean.numpy(),
        "2.2.0.1.running_var": model.layer2[1].bn1.running_var.numpy(),
        "2.2.0.3.weight": model.layer2[1].conv2.weight.numpy(),
        "2.2.0.4.scale": model.layer2[1].bn2.weight.numpy(),
        "2.2.0.4.bias": model.layer2[1].bn2.bias.numpy(),
        "2.2.0.4.running_mean": model.layer2[1].bn2.running_mean.numpy(),
        "2.2.0.4.running_var": model.layer2[1].bn2.running_var.numpy(),

        # layer3
        "3.0.f.0.weight": model.layer3[0].conv1.weight.numpy(),
        "3.0.f.1.scale": model.layer3[0].bn1.weight.numpy(),
        "3.0.f.1.bias": model.layer3[0].bn1.bias.numpy(),
        "3.0.f.1.running_mean": model.layer3[0].bn1.running_mean.numpy(),
        "3.0.f.1.running_var": model.layer3[0].bn1.running_var.numpy(),
        "3.0.f.3.weight": model.layer3[0].conv2.weight.numpy(),
        "3.0.f.4.scale": model.layer3[0].bn2.weight.numpy(),
        "3.0.f.4.bias": model.layer3[0].bn2.bias.numpy(),
        "3.0.f.4.running_mean": model.layer3[0].bn2.running_mean.numpy(),
        "3.0.f.4.running_var": model.layer3[0].bn2.running_var.numpy(),
        "3.0.r.0.weight": model.layer3[0].downsample[0].weight.numpy(),
        "3.0.r.1.scale": model.layer3[0].downsample[1].weight.numpy(),
        "3.0.r.1.bias": model.layer3[0].downsample[1].bias.numpy(),
        "3.0.r.1.running_mean": model.layer3[0].downsample[1].running_mean.numpy(),
        "3.0.r.1.running_var": model.layer3[0].downsample[1].running_var.numpy(),
        "3.2.0.0.weight": model.layer3[1].conv1.weight.numpy(),
        "3.2.0.1.scale": model.layer3[1].bn1.weight.numpy(),
        "3.2.0.1.bias": model.layer3[1].bn1.bias.numpy(),
        "3.2.0.1.running_mean": model.layer3[1].bn1.running_mean.numpy(),
        "3.2.0.1.running_var": model.layer3[1].bn1.running_var.numpy(),
        "3.2.0.3.weight": model.layer3[1].conv2.weight.numpy(),
        "3.2.0.4.scale": model.layer3[1].bn2.weight.numpy(),
        "3.2.0.4.bias": model.layer3[1].bn2.bias.numpy(),
        "3.2.0.4.running_mean": model.layer3[1].bn2.running_mean.numpy(),
        "3.2.0.4.running_var": model.layer3[1].bn2.running_var.numpy(),

        # layer4
        "4.0.f.0.weight": model.layer4[0].conv1.weight.numpy(),
        "4.0.f.1.scale": model.layer4[0].bn1.weight.numpy(),
        "4.0.f.1.bias": model.layer4[0].bn1.bias.numpy(),
        "4.0.f.1.running_mean": model.layer4[0].bn1.running_mean.numpy(),
        "4.0.f.1.running_var": model.layer4[0].bn1.running_var.numpy(),
        "4.0.f.3.weight": model.layer4[0].conv2.weight.numpy(),
        "4.0.f.4.scale": model.layer4[0].bn2.weight.numpy(),
        "4.0.f.4.bias": model.layer4[0].bn2.bias.numpy(),
        "4.0.f.4.running_mean": model.layer4[0].bn2.running_mean.numpy(),
        "4.0.f.4.running_var": model.layer4[0].bn2.running_var.numpy(),
        "4.0.r.0.weight": model.layer4[0].downsample[0].weight.numpy(),
        "4.0.r.1.scale": model.layer4[0].downsample[1].weight.numpy(),
        "4.0.r.1.bias": model.layer4[0].downsample[1].bias.numpy(),
        "4.0.r.1.running_mean": model.layer4[0].downsample[1].running_mean.numpy(),
        "4.0.r.1.running_var": model.layer4[0].downsample[1].running_var.numpy(),
        "4.2.0.0.weight": model.layer4[1].conv1.weight.numpy(),
        "4.2.0.1.scale": model.layer4[1].bn1.weight.numpy(),
        "4.2.0.1.bias": model.layer4[1].bn1.bias.numpy(),
        "4.2.0.1.running_mean": model.layer4[1].bn1.running_mean.numpy(),
        "4.2.0.1.running_var": model.layer4[1].bn1.running_var.numpy(),
        "4.2.0.3.weight": model.layer4[1].conv2.weight.numpy(),
        "4.2.0.4.scale": model.layer4[1].bn2.weight.numpy(),
        "4.2.0.4.bias": model.layer4[1].bn2.bias.numpy(),
        "4.2.0.4.running_mean": model.layer4[1].bn2.running_mean.numpy(),
        "4.2.0.4.running_var": model.layer4[1].bn2.running_var.numpy(),

        # tail
        "5.1.weight": model.fc.weight.numpy(),
        "5.1.bias": model.fc.bias.numpy(),
    })

    x = torch.randn(10, 3, 224, 224)
    y = model(x)

    np.save("resnet18_x.npy", x.numpy())
    np.save("resnet18_y.npy", y.numpy())