-
Notifications
You must be signed in to change notification settings - Fork 0
/
vqvae.py
200 lines (141 loc) · 6.72 KB
/
vqvae.py
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
import torch
import torch.nn as nn
from torch.autograd import Function
from torchsummary import summary
def weight_init(model):
classname = model.__class__.__name__
if classname.find('Conv') != -1:
nn.init.xavier_normal_(model.weight.data)
model.bias.data.fill_(0)
class VectorQuantization(Function):
@staticmethod
def forward(ctx, inputs, codebook):
"""
Parameters
----------
ctx : context. It is used to store arbitrary data that can be retrieved
during the backward pass.
inputs : Tensor
Encoded image vector(z_e_x).
codebook : Tensor
Embeddings.
Returns
-------
idx : quantized indices / latents.
"""
with torch.no_grad():
input_size = inputs.size()
emb_size = codebook.size(1)
flat_input = inputs.view(-1, emb_size)
codebook_sq = torch.sum(codebook ** 2, dim = 1)
inputs_sq = torch.sum(inputs ** 2, dim = 1, keepdim = True)
#torch.addmm: result = beta * input + alpha * (mat1[i] @ mat2[i])
l2_dis = torch.addmm(input = codebook_sq + inputs_sq,
mat1 = flat_input, mat2 = codebook.t(),
alpha = 2.0, beta = 1.0)
_, idx_flatten = torch.min(l2_dis, dim = 1)
idx = idx_flatten(*input_size[:-1])
ctx.mark_non_differentiable(idx)
return idx
@staticmethod
def backward(ctx, grad_outputs):
raise RuntimeError('Trying to call backward on graph containing `Vector Quantization`'
'which is non-differentiable. Use VQStraightThrough instead.')
class VQStraightThrough(Function):
@staticmethod
def forward(ctx, inputs, codebook):
idx = VQ(inputs, codebook)
flat_idx = idx.view(-1)
ctx.save_for_backward(flat_idx, codebook)
ctx.mark_non_differentiable(flat_idx)
codes_flatten = torch.index_select(input = codebook, dim = 0, index = flat_idx)
codes = codes_flatten.view_as(inputs)
return (codes, flat_idx)
@staticmethod
def backward(ctx, grad_outputs, grad_indices):
grad_inputs, grad_codebook = None, None
if ctx.needs_input_grad[0]:
grad_inputs = grad_outputs.clone()
if ctx.needs_input_grad[1]:
idx, codebook = ctx.saved_tensors
emb_size = codebook.size(1)
flat_grad_output = (grad_outputs.contiguous().view(-1, emb_size))
grad_codebook = torch.zeros_like(codebook)
grad_codebook.index_add_(0, idx, flat_grad_output)
return (grad_inputs, grad_codebook)
VQ = VectorQuantization.apply
VQ_ST = VQStraightThrough.apply
class VQEmbedding(nn.Module):
def __init__(self, K: int, D: int):
"""
Parameters
----------
K : int
Total number of embeddings in codebook.
D : int
Dimensions of every embedding in the codebook.
"""
super().__init__()
self.vq_embs = nn.Embedding(K, D)
self.vq_embs.weight.data.uniform_(-1.0 / K, 1.0 / K)
def forward(self, z_e_x):
z_e_x = z_e_x.permute(0,2,3,1).contiguous()
latents = VQ(z_e_x, self.vq_embs.weight)
return latents
def straight_through_forward(self, z_e_x):
zex = z_e_x.permute(0,2,3,1).contiguous()
z_q_x, idx = VQ_ST(zex, self.vq_embs.weight.detach())
z_q_x = z_q_x.permute(0,3,1,2).contiguous()
flat_zqx_tilde = torch.index_select(self.vq_embs.weight, dim = 0, index = idx)
zqx_tilde = flat_zqx_tilde.view_as(zex)
zqx_tilde = zqx_tilde.permute(0,3,1,2).contiguous()
return z_q_x, zqx_tilde
class ResidualBlock(nn.Module):
def __init__(self, channels):
super().__init__()
self.resblock = nn.Sequential(nn.ReLU(inplace = True),
nn.Conv2d(channels, channels, kernel_size = 3, stride = 1, padding = 1),
nn.BatchNorm2d(channels),
nn.ReLU(inplace = True),
nn.Conv2d(channels, channels, 1),
nn.BatchNorm2d(channels)
)
def forward(self, x):
return self.resblock(x) + x
class VQVAE(nn.Module):
def __init__(self, in_c, out_c, K = 512):
super().__init__()
self.encoder = nn.Sequential(nn.Conv2d(in_c, out_c, kernel_size = 4, stride = 2, padding = 1),
nn.BatchNorm2d(out_c),
nn.ReLU(inplace = True),
nn.Conv2d(out_c, out_c, kernel_size = 4, stride = 2, padding = 1),
ResidualBlock(out_c),
ResidualBlock(out_c)
)
self.codebook = VQEmbedding(K, D = out_c)
self.decoder = nn.Sequential(ResidualBlock(out_c),
ResidualBlock(out_c),
nn.ReLU(inplace = True),
nn.ConvTranspose2d(out_c, out_c, kernel_size = 4, stride = 2, padding = 1),
nn.BatchNorm2d(out_c),
nn.ReLU(inplace = True),
nn.ConvTranspose2d(out_c, in_c, kernel_size = 4, stride = 2, padding = 1),
nn.Tanh()
)
self.apply(weight_init)
def encode(self, x):
z_e_x = self.encoder(x)
latents = self.codebook(z_e_x)
return latents
def decode(self, latents):
z_q_x = self.codebook.vq_embs(latents).permute(0,3,1,2)
x_tilde = self.decoder(z_q_x)
return x_tilde
def forward(self, x):
z_e_x = self.encoder(x)
z_q_x_st, z_q_x = self.codebook.straight_through_forward(z_e_x)
x_tilde = self.decoder(z_q_x_st)
return z_e_x, z_q_x, x_tilde
if __name__ == '__main__':
vqvae = VQVAE(3, 256).to(torch.device('cuda'))
print(summary(vqvae, (3,32,32)))