robust-gan/model.py at master · watml/robust-gan · GitHub

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
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F


class Generator(nn.Module):

    def __init__(self, p, initializer=None):
        super().__init__()
        self.p = p
        if initializer is None:
            self.eta = nn.Parameter(torch.randn(self.p) / self.p)
        else:
            self.eta = nn.Parameter(initializer)

    def forward(self, x):
        x = x + self.eta
        return x

    def get_numpy_eta(self):
        return self.eta.detach().cpu().numpy()


class Discriminator(nn.Module):

    def __init__(self, input_dim, lst_num_hidden, lst_activation, kappa=None,
                 l1_constrain_type=None):
        super().__init__()

        assert len(lst_activation) == len(lst_num_hidden)

        self.lst_activation = lst_activation
        self.kappa = kappa
        self.layers = nn.ModuleList()

        lst_num_nodes = [input_dim, ] + lst_num_hidden + [1, ]

        for i in range(len(lst_num_nodes) - 2):
            # fc = torch.nn.utils.spectral_norm(
            #     nn.Linear(lst_num_nodes[i], lst_num_nodes[i + 1]))
            # fc = L2RowConstrainedLinear(lst_num_nodes[i], lst_num_nodes[i + 1])
            fc = nn.Linear(lst_num_nodes[i], lst_num_nodes[i + 1])
            self.layers.append(fc)
        assert len(lst_activation) == len(self.layers)

        if self.kappa is None:
            self.last_linear_layer = nn.Linear(
                lst_num_nodes[-2], lst_num_nodes[-1], bias=False)
        else:
            assert (l1_constrain_type is not None)
            self.last_linear_layer = L1ConstrainedLinear(
                lst_num_nodes[-2], lst_num_nodes[-1], bias=False,
                l1_constrain_type=l1_constrain_type)

        # assert len(lst_num_hidden) == 1
        # self.bn = nn.BatchNorm1d(lst_num_hidden[0])

    def forward(self, x):
        for i in range(len(self.layers)):
            x = self.layers[i](x)
            x = self.lst_activation[i](x)

        # x = self.bn(x)

        if self.kappa is not None:
            x = self.last_linear_layer(x, self.kappa)
        else:
            x = self.last_linear_layer(x)

        return x
        # return torch.tanh(x)


class L1ConstrainedLinear(nn.Linear):

    def __init__(self, in_features, out_features, bias=True,
                 l1_constrain_type="proj"):
        super().__init__(in_features, out_features, bias)
        self.l1_constrain_type = l1_constrain_type

    def forward(self, input, kappa):
        # XXX: currently constraining the norm of the entire weight, maybe
        #   should just constrain each row
        if self.l1_constrain_type == "reparam":
            ww = _l1_scale(self.weight, kappa)
        elif self.l1_constrain_type == "scale":
            self.weight.data = _l1_scale(self.weight.data, kappa)
            ww = self.weight
        elif self.l1_constrain_type == "proj":
            self.weight.data = _l1_proj(self.weight.data, kappa)
            ww = self.weight
        else:
            raise ValueError(self.l1_constrain_type)

        return F.linear(input, ww, self.bias)


def _l1_scale(x, kappa):
    l1norm = x.abs().sum()
    # TODO: could be simplified with max
    if l1norm > kappa:
        return kappa * (x / l1norm)
    else:
        return x


def _l1_proj(x, kappa):
    from advertorch.utils import batch_l1_proj
    return batch_l1_proj(x[None, :], kappa)[0]


class L2RowConstrainedLinear(nn.Linear):

    def __init__(self, in_features, out_features, bias=True,
                 l2_constrain_type="reparam"):
        super().__init__(in_features, out_features, bias)
        self.l2_constrain_type = l2_constrain_type

    def forward(self, input, rownorm=1.):
        if self.l2_constrain_type == "reparam":
            ww = _batch_l2_scale(self.weight, rownorm)
        elif self.l2_constrain_type in ["scale", "proj"]:
            self.weight.data = _batch_l2_scale(self.weight.data, rownorm)
            ww = self.weight
        else:
            raise ValueError(self.l2_constrain_type)

        return F.linear(input, ww, self.bias)


def _batch_l2_scale(x, rownorm):
    from advertorch.utils import clamp_by_pnorm
    return clamp_by_pnorm(x, 2., rownorm)