-
Notifications
You must be signed in to change notification settings - Fork 0
/
model.py
239 lines (212 loc) · 8.38 KB
/
model.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
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
from collections import OrderedDict
import torch
import torch.nn as nn
from torch.distributions.multinomial import Multinomial
class SimpleModel(nn.Module):
"""SimpleModel represents a lightweight model for checking codes.
This model is quite simple to check codes quickly.
Attributes
----------
self.num_classes int : number of classes of dataset.
self.layers nn.ModuleDict : ModuleDict of models.
"""
def __init__(self, num_classes):
"""
Parameters
----------
num_classes int : number of classes of dataset.
"""
super(SimpleModel, self).__init__()
self.num_classes = num_classes
self.layers = nn.ModuleList([
nn.Conv2d(3, 9, 3, padding=1, stride=1),
nn.GroupNorm(3, 9),
nn.ELU(),
nn.Conv2d(9, self.num_classes, 1, padding=0, stride=2),
nn.AdaptiveAvgPool2d(1)
])
def forward(self, x):
"""
Parameters
----------
x torch.Tensor : input tensor whose shape is [b, c, h, w].
Returns
-------
torch.squeeze(x) torch.Tensor : logit tensor which will be input of softmax.
"""
for layer in self.layers:
x = layer(x)
return torch.reshape(x, x.shape[:2]) # [b, num_classes]
class Model(nn.Module):
"""Model represents a model mainly used in experiments.
Attributes
----------
self.num_classes int : number of classes of dataset.
self.layers nn.ModuleDict : ModuleDict of models.
"""
def __init__(self, num_classes):
"""
Parameters
----------
num_classes int : number of classes of dataset.
"""
super(Model, self).__init__()
self.num_classes = num_classes
self.layers = nn.ModuleDict(OrderedDict([
# CONV-GN-ELU
("conv1", nn.Conv2d(3, 96, 3, padding=1, stride=1)),
("GN1", nn.GroupNorm(3, 96)),
("ELU1", nn.ELU()),
# CONV-GN-ELU * 2 + Dropout
("conv2", nn.Conv2d(96, 96, 3, padding=1, stride=1)),
("GN2", nn.GroupNorm(3, 96)),
("ELU2", nn.ELU()),
("conv3", nn.Conv2d(96, 96, 3, padding=1, stride=2)),
("GN3", nn.GroupNorm(3, 96)),
("ELU3", nn.ELU()),
("DO1", nn.Dropout(0.5)),
# CONV-GN-ELU * 3 + Dropout
("conv4", nn.Conv2d(96, 192, 3, padding=1, stride=1)),
("GN4", nn.GroupNorm(6, 192)),
("ELU4", nn.ELU()),
("conv5", nn.Conv2d(192, 192, 3, padding=1, stride=1)),
("GN5", nn.GroupNorm(6, 192)),
("ELU5", nn.ELU()),
("conv6", nn.Conv2d(192, 192, 3, padding=1, stride=2)),
("GN6", nn.GroupNorm(6, 192)),
("ELU6", nn.ELU()),
("DO2", nn.Dropout(0.5)),
# CONV-GN-ELU * 2 + CONV + GAP
("conv7", nn.Conv2d(192, 192, 3, padding=1, stride=1)),
("GN7", nn.GroupNorm(6, 192)),
("ELU7", nn.ELU()),
("conv8", nn.Conv2d(192, 192, 1, padding=0, stride=1)),
("GN8", nn.GroupNorm(6, 192)),
("ELU8", nn.ELU()),
("conv9", nn.Conv2d(192, self.num_classes, 1, padding=0, stride=2)),
("pool", nn.AdaptiveAvgPool2d(1))
]))
def forward(self, x):
"""
Parameters
----------
x torch.Tensor : input tensor whose shape is [b, c, h, w].
Returns
-------
torch.squeeze(x) torch.Tensor : logit tensor which will be input of softmax.
"""
for layer in self.layers.values():
x = layer(x)
return torch.reshape(x, x.shape[:2]) # [b, num_classes]
class StochasticActivationPruning(nn.Module):
"""SimpleModel represents a nn.Module of Stochastic Activation Pruning.
The original paper is https://arxiv.org/abs/1803.01442.
Attributes
----------
self.ratio float : ratio of pruning which can be larger than 1.0.
self.is_valid bool : if this flag is True, inject SAP.
"""
def __init__(self, ratio=1.0, is_valid=False):
"""
Parameters
----------
ratio float : ratio of pruning which can be larger than 1.0.
is_valid bool : if this flag is True, inject SAP.
"""
super(StochasticActivationPruning, self).__init__()
self.ratio = ratio
self.is_valid = is_valid
def forward(self, inputs):
"""
If self.training or not self.is_valid, just return inputs.
If self.is_valid apply SAP to inputs and return the result tensor.
Parameters
----------
inputs torch.Tensor : input tensor whose shape is [b, c, h, w].
Returns
-------
outputs torch.Tensor : just return inputs or stochastically pruned inputs.
"""
if self.training or not self.is_valid:
return inputs
else:
b, c, h, w = inputs.shape
inputs_1d = inputs.reshape([b, c * h * w]) # [b, c * h * w]
outputs = torch.zeros_like(inputs_1d) # outputs with 0 initilization
inputs_1d_sum = torch.sum(torch.abs(inputs_1d), dim=-1, keepdim=True)
inputs_1d_prob = torch.abs(inputs_1d) / inputs_1d_sum
repeat_num = int(c * h * w * self.ratio)
idx = Multinomial(repeat_num, inputs_1d_prob).sample()
outputs[idx.nonzero(as_tuple=True)] = inputs_1d[idx.nonzero(as_tuple=True)]
outputs = outputs / (1 - (1 - inputs_1d_prob) ** repeat_num + 1e-12)
outputs = outputs.reshape([b, c, h, w]) # [b, c, h, w]
return outputs
class ModelSAP(nn.Module):
"""Model represents a model mainly used in experiments.
Attributes
----------
self.num_classes int : number of classes of dataset.
self.layers nn.ModuleDict : ModuleDict of models.
"""
def __init__(self, num_classes):
"""
Parameters
----------
num_classes int : number of classes of dataset.
"""
super(ModelSAP, self).__init__()
self.num_classes = num_classes
self.layers = nn.ModuleDict(OrderedDict([
# CONV-GN-ELU
("conv1", nn.Conv2d(3, 96, 3, padding=1, stride=1)),
("GN1", nn.GroupNorm(3, 96)),
("ELU1", nn.ELU()),
("SAP1", StochasticActivationPruning()),
# CONV-GN-ELU * 2 + Dropout
("conv2", nn.Conv2d(96, 96, 3, padding=1, stride=1)),
("GN2", nn.GroupNorm(3, 96)),
("ELU2", nn.ELU()),
("SAP2", StochasticActivationPruning()),
("conv3", nn.Conv2d(96, 96, 3, padding=1, stride=2)),
("GN3", nn.GroupNorm(3, 96)),
("ELU3", nn.ELU()),
("SAP3", StochasticActivationPruning()),
("DO1", nn.Dropout(0.5)),
# CONV-GN-ELU * 3 + Dropout
("conv4", nn.Conv2d(96, 192, 3, padding=1, stride=1)),
("GN4", nn.GroupNorm(6, 192)),
("ELU4", nn.ELU()),
("SAP4", StochasticActivationPruning()),
("conv5", nn.Conv2d(192, 192, 3, padding=1, stride=1)),
("GN5", nn.GroupNorm(6, 192)),
("ELU5", nn.ELU()),
("SAP5", StochasticActivationPruning()),
("conv6", nn.Conv2d(192, 192, 3, padding=1, stride=2)),
("GN6", nn.GroupNorm(6, 192)),
("ELU6", nn.ELU()),
("SAP6", StochasticActivationPruning()),
("DO2", nn.Dropout(0.5)),
# CONV-GN-ELU * 2 + CONV + GAP
("conv7", nn.Conv2d(192, 192, 3, padding=1, stride=1)),
("GN7", nn.GroupNorm(6, 192)),
("ELU7", nn.ELU()),
("SAP7", StochasticActivationPruning(is_valid=True)),
("conv8", nn.Conv2d(192, 192, 1, padding=0, stride=1)),
("GN8", nn.GroupNorm(6, 192)),
("ELU8", nn.ELU()),
("SAP8", StochasticActivationPruning(is_valid=True)),
("conv9", nn.Conv2d(192, self.num_classes, 1, padding=0, stride=2)),
("pool", nn.AdaptiveAvgPool2d(1))
]))
def forward(self, x):
"""
Parameters
----------
x torch.Tensor : input tensor whose shape is [b, c, h, w].
Returns
-------
torch.squeeze(x) torch.Tensor : logit tensor which will be input of softmax.
"""
for layer in self.layers.values():
x = layer(x)
return torch.reshape(x, x.shape[:2]) # [b, num_classes]