File size: 5,759 Bytes
f0c7f08
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import torch
import torch.nn as nn
import torch.nn.functional as F


def weights_init(m):
    classname = m.__class__.__name__
    if classname.find('Conv') != -1:
        try:
            nn.init.xavier_uniform_(m.weight.data)
            m.bias.data.fill_(0)
        except AttributeError:
            print("Skipping initialization of ", classname)


class GatedActivation(nn.Module):
    def __init__(self):
        super().__init__()

    def forward(self, x):
        x, y = x.chunk(2, dim=1)
        return F.tanh(x) * F.sigmoid(y)


class GatedMaskedConv2d(nn.Module):
    def __init__(self, mask_type, dim, kernel, residual=True, n_classes=10, bh_model=False):
        super().__init__()
        assert kernel % 2 == 1, print("Kernel size must be odd")
        self.mask_type = mask_type
        self.residual = residual
        self.bh_model = bh_model

        self.class_cond_embedding = nn.Embedding(
            n_classes, 2 * dim
        )

        kernel_shp = (kernel // 2 + 1, 3 if self.bh_model else 1)  # (ceil(n/2), n)
        padding_shp = (kernel // 2, 1 if self.bh_model else 0)
        self.vert_stack = nn.Conv2d(
            dim, dim * 2,
            kernel_shp, 1, padding_shp
        )

        self.vert_to_horiz = nn.Conv2d(2 * dim, 2 * dim, 1)

        kernel_shp = (1, 2)
        padding_shp = (0, 1)
        self.horiz_stack = nn.Conv2d(
            dim, dim * 2,
            kernel_shp, 1, padding_shp
        )

        self.horiz_resid = nn.Conv2d(dim, dim, 1)

        self.gate = GatedActivation()

    def make_causal(self):
        self.vert_stack.weight.data[:, :, -1].zero_()  # Mask final row
        self.horiz_stack.weight.data[:, :, :, -1].zero_()  # Mask final column

    def forward(self, x_v, x_h, h):
        if self.mask_type == 'A':
            self.make_causal()

        h = self.class_cond_embedding(h)
        h_vert = self.vert_stack(x_v)
        h_vert = h_vert[:, :, :x_v.size(-2), :]
        out_v = self.gate(h_vert + h[:, :, None, None])

        if self.bh_model:
            h_horiz = self.horiz_stack(x_h)
            h_horiz = h_horiz[:, :, :, :x_h.size(-1)]
            v2h = self.vert_to_horiz(h_vert)

            out = self.gate(v2h + h_horiz + h[:, :, None, None])
            if self.residual:
                out_h = self.horiz_resid(out) + x_h
            else:
                out_h = self.horiz_resid(out)
        else:
            if self.residual:
                out_v = self.horiz_resid(out_v) + x_v
            else:
                out_v = self.horiz_resid(out_v)
            out_h = out_v

        return out_v, out_h


class GatedPixelCNN(nn.Module):
    def __init__(self, input_dim=256, dim=64, n_layers=15, n_classes=10, audio=False, bh_model=False):
        super().__init__()
        self.dim = dim
        self.audio = audio
        self.bh_model = bh_model

        if self.audio:
            self.embedding_aud = nn.Conv2d(256, dim, 1, 1, padding=0)
            self.fusion_v = nn.Conv2d(dim * 2, dim, 1, 1, padding=0)
            self.fusion_h = nn.Conv2d(dim * 2, dim, 1, 1, padding=0)

        # Create embedding layer to embed input
        self.embedding = nn.Embedding(input_dim, dim)

        # Building the PixelCNN layer by layer
        self.layers = nn.ModuleList()

        # Initial block with Mask-A convolution
        # Rest with Mask-B convolutions
        for i in range(n_layers):
            mask_type = 'A' if i == 0 else 'B'
            kernel = 7 if i == 0 else 3
            residual = False if i == 0 else True

            self.layers.append(
                GatedMaskedConv2d(mask_type, dim, kernel, residual, n_classes, bh_model)
            )

        # Add the output layer
        self.output_conv = nn.Sequential(
            nn.Conv2d(dim, 512, 1),
            nn.ReLU(True),
            nn.Conv2d(512, input_dim, 1)
        )

        self.apply(weights_init)

        self.dp = nn.Dropout(0.1)

    def forward(self, x, label, aud=None):
        shp = x.size() + (-1,)
        x = self.embedding(x.view(-1)).view(shp)  # (B, H, W, C)
        x = x.permute(0, 3, 1, 2)  # (B, C, W, W)

        x_v, x_h = (x, x)
        for i, layer in enumerate(self.layers):
            if i == 1 and self.audio is True:
                aud = self.embedding_aud(aud)
                a = torch.ones(aud.shape[-2]).to(aud.device)
                a = self.dp(a)
                aud = (aud.transpose(-1, -2) * a).transpose(-1, -2)
                x_v = self.fusion_v(torch.cat([x_v, aud], dim=1))
                if self.bh_model:
                    x_h = self.fusion_h(torch.cat([x_h, aud], dim=1))
            x_v, x_h = layer(x_v, x_h, label)

        if self.bh_model:
            return self.output_conv(x_h)
        else:
            return self.output_conv(x_v)

    def generate(self, label, shape=(8, 8), batch_size=64, aud_feat=None, pre_latents=None, pre_audio=None):
        param = next(self.parameters())
        x = torch.zeros(
            (batch_size, *shape),
            dtype=torch.int64, device=param.device
        )
        if pre_latents is not None:
            x = torch.cat([pre_latents, x], dim=1)
            aud_feat = torch.cat([pre_audio, aud_feat], dim=2)
            h0 = pre_latents.shape[1]
            h = h0 + shape[0]
        else:
            h0 = 0
            h = shape[0]

        for i in range(h0, h):
            for j in range(shape[1]):
                if self.audio:
                    logits = self.forward(x, label, aud_feat)
                else:
                    logits = self.forward(x, label)
                probs = F.softmax(logits[:, :, i, j], -1)
                x.data[:, i, j].copy_(
                    probs.multinomial(1).squeeze().data
                )
        return x[:, h0:h]