-
Notifications
You must be signed in to change notification settings - Fork 1
/
MBP_EncoderDecoder.py
199 lines (152 loc) · 5.95 KB
/
MBP_EncoderDecoder.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
import torch
import torch.nn as nn
from torch.nn import init
import torch.nn.functional as F
from torch.autograd import Variable
import numpy as np
class ImageEncoderResidualBlock(nn.Module):
def __init__(self, depth, bottleneck_depth, stride):
super(ImageEncoderResidualBlock,self).__init__()
self.stride = stride
self.conv1 = nn.Conv2d( depth, depth, (1, 1))
self.conv2 = nn.Conv2d( depth, bottleneck_depth, (3, 3),padding=(1,1),stride=stride)
self.conv3 = nn.Conv2d( bottleneck_depth, depth, (1, 1))
if(self.stride>1):
self.conv_res = nn.Conv2d(depth, depth, (1,1),stride = (stride,stride))
for i in range(1,len(list(self.modules()))):
list(self.modules())[i].weight.data.normal_(0.0,0.01)
def forward(self, ins):
v = ins
#print("v1",v.shape)
v = F.relu(self.conv1(v))
#print("v2",v.shape)
v = F.relu(self.conv2(v))
#print("v3",v.shape)
v = F.relu(self.conv3(v))
#print("v4",v.shape)
#print("stride",self.stride)
res = ins if self.stride <= 1 else self.conv_res(ins)
#print("res",res.shape)
v = v + res
return v
class ImageEncoder(nn.Module):
def __init__(self, num_resblocks=6, res_depth=64, bottleneck_depth=32, output_dim=500, strides=(2,1,2,1,2,1), input_shape=(320,224),res_input_shape = (32,32),linear_in=1024):
super(ImageEncoder,self).__init__() #sonic image size (320 x 224 x 3)
self.linear_in = linear_in
self.num_resblocks = num_resblocks
rescale = (input_shape[0]//res_input_shape[0], input_shape[1]//res_input_shape[1])
print("rescale",rescale)
self.conv_initial = torch.nn.Conv2d(3, res_depth, kernel_size=rescale, stride=rescale)
self.resblocks = torch.nn.ModuleList()
for i in range(num_resblocks):
self.resblocks.append(ImageEncoderResidualBlock(res_depth,bottleneck_depth,strides[i]))
self.linear = torch.nn.Linear(linear_in, output_dim)
def forward(self, v):
#print("v-1",v.shape)
v = F.relu(self.conv_initial(v))
#print("v0",v.shape)
for i in range(self.num_resblocks):
v = self.resblocks[i](v)
v = v.view([-1,self.linear_in])
v = F.tanh(self.linear(v))
return v
class ImageDecoderResidualBlock(nn.Module):
def __init__(self, depth, bottleneck_depth, stride):
super(ImageDecoderResidualBlock,self).__init__()
#only works with strides of 1 or 2
self.stride = stride
self.conv1 = nn.ConvTranspose2d(depth, bottleneck_depth, (1, 1))
if(stride>1):
self.conv2 = nn.ConvTranspose2d(bottleneck_depth, depth, (3, 3),padding=(0,0),stride=2)
else:
self.conv2 = nn.ConvTranspose2d(bottleneck_depth, depth, (3, 3),padding=(1,1),stride=1)
self.conv3 = nn.ConvTranspose2d( depth, depth, (1, 1))
if(self.stride>1):
self.conv_res = nn.ConvTranspose2d(depth, depth, (stride,stride),stride = stride)
for i in range(1,len(list(self.modules()))):
list(self.modules())[i].weight.data.normal_(0.0,0.01)
def forward(self, ins):
v = ins
#print("v1",v.shape)
v = F.relu(self.conv1(v))
#print("v2",v.shape)
if(self.stride > 1):
v = F.relu(self.conv2(v))[:,:,:-1,:-1]
else:
v = F.relu(self.conv2(v))
#print("v3",v.shape)
v = F.relu(self.conv3(v))
#print("v4",v.shape)
res = ins if self.stride <= 1 else self.conv_res(ins)
#print("res",res.shape)
v = v + res
return v
class ImageDecoder(nn.Module):
def __init__(self, num_resblocks=6, res_depth=64, bottleneck_depth=32, input_dim=500, strides=(2,1,2,1,2,1),output_shape=(320,224),res_output_shape = (32,32),linear_out=1024):
super(ImageDecoder,self).__init__() #sonic image size (320 x 224 x 3)
self.num_resblocks = num_resblocks
self.linear_out = linear_out
self.input_dim = input_dim
self.linear = torch.nn.Linear(input_dim,linear_out)
self.res_depth = res_depth
self.resblocks = torch.nn.ModuleList()
for i in range(num_resblocks):
self.resblocks.append(ImageDecoderResidualBlock(res_depth,bottleneck_depth,strides[i]))
rescale = (output_shape[0]//res_output_shape[0],output_shape[1]//res_output_shape[1])
self.conv_initial = torch.nn.ConvTranspose2d(res_depth, 3, kernel_size=rescale, stride=rescale)
def forward(self, v):
#print("v",v.shape)
v = v.view([-1,self.input_dim])
#print("v",v.shape)
v = self.linear(v)
#print("v",v.shape)
v = v.view([-1,self.res_depth,8,8])
#print("v",v.shape)
for i in range(self.num_resblocks):
v = self.resblocks[i](v)
#v = F.relu(self.conv_initial(v))
#v = torch.clamp(v,0,255)
v = 255*F.sigmoid(self.conv_initial(v))
#v=255*F.sigmoid(v)
return v
class ImageEncoderSimplified1(nn.Module):
def __init__(self):
super(ImageEncoderSimplified1,self).__init__() #sonic image size (320 x 224 x 3)
self.conv1 = torch.nn.Conv2d(3, 8, kernel_size=(10,7), stride=(10,7))
self.conv2 = torch.nn.Conv2d(8, 8, kernel_size=3, stride=1,padding=1)
self.conv3 = torch.nn.Conv2d(8, 16, kernel_size=3, stride=2,padding=1)
self.linear = torch.nn.Linear(1024,128)
for i in range(1,len(list(self.modules()))):
list(self.modules())[i].weight.data.normal_(0.0,0.2)
def forward(self, v):
#print("v0",v.shape)
v = F.relu(self.conv1(v))
#print("v1",v.shape)
v = F.relu(self.conv2(v))+v
#print("v2",v.shape)
v = F.relu(self.conv3(v))
#print("v3",v.shape)
#v = v.view([-1])
v = F.tanh(self.linear(v))
return v
class ImageDecoderSimplified1(nn.Module):
def __init__(self):
super(ImageDecoderSimplified1,self).__init__() #sonic image size (320 x 224 x 3)
self.linear = torch.nn.Linear(128,512)
self.conv1 = torch.nn.ConvTranspose2d(8, 8, kernel_size=3, stride=2)#,padding=1)
self.conv2 = torch.nn.ConvTranspose2d(8, 8, kernel_size=3, stride=1,padding=1)
self.conv3 = torch.nn.ConvTranspose2d(8, 3, kernel_size=(10,7), stride=(10,7))
for i in range(1,len(list(self.modules()))):
list(self.modules())[i].weight.data.normal_(0.0,0.2)
def forward(self, v):
#print("v-1",v.shape)
v = F.tanh(self.linear(v))
v = v.view([-1,8,8,8])
#print("v0",v.shape)
v = F.relu(self.conv1(v))[:,:,:-1,:-1]
#print("v1",v.shape)
v = F.relu(self.conv2(v))+v
#print("v2",v.shape)
v = F.relu(self.conv3(v))
#print("v3",v.shape)
return v