-
Notifications
You must be signed in to change notification settings - Fork 26
/
googlenet.py
225 lines (175 loc) · 8.36 KB
/
googlenet.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
import torch.nn as nn
import torch.nn.functional as F
import torch
from torch.autograd import Function, Variable
class ConstantPadNd(Function):
@staticmethod
def forward(ctx, input, pad, value=0):
ctx.pad = pad
ctx.value = value
ctx.input_size = input.size()
ctx.l_inp = len(input.size())
ctx.pad_tup = tuple([(a, b) for a, b in zip(pad[:-1:2], pad[1::2])][::-1])
ctx.l_pad = len(ctx.pad_tup)
ctx.l_diff = ctx.l_inp - ctx.l_pad
assert ctx.l_inp >= ctx.l_pad
new_dim = tuple([sum((d,) + ctx.pad_tup[i]) for i, d in enumerate(input.size()[-ctx.l_pad:])])
assert all([d > 0 for d in new_dim]), 'input is too small'
# crop input if necessary
output = input.new(input.size()[:(ctx.l_diff)] + new_dim).fill_(ctx.value)
c_input = input
for i, p in zip(range(ctx.l_inp)[-ctx.l_pad:], ctx.pad_tup):
if p[0] < 0:
c_input = c_input.narrow(i, -p[0], c_input.size(i) + p[0])
if p[1] < 0:
c_input = c_input.narrow(i, 0, c_input.size(i) + p[1])
# crop output if necessary
c_output = output
for i, p in zip(range(ctx.l_inp)[-ctx.l_pad:], ctx.pad_tup):
if p[0] > 0:
c_output = c_output.narrow(i, p[0], c_output.size(i) - p[0])
if p[1] > 0:
c_output = c_output.narrow(i, 0, c_output.size(i) - p[1])
c_output.copy_(c_input)
return output
@staticmethod
def backward(ctx, grad_output):
grad_input = Variable(grad_output.data.new(ctx.input_size).zero_())
grad_input_slices = [slice(0, x,) for x in ctx.input_size]
def narrow_slice(dim, start, length):
grad_input_slices[dim] = (slice(grad_input_slices[dim].start + start,
grad_input_slices[dim].start + start + length))
def slice_length(dim):
return grad_input_slices[dim].stop - grad_input_slices[dim].start
# crop grad_input if necessary
for i, p in zip(range(ctx.l_inp)[-ctx.l_pad:], ctx.pad_tup):
if p[0] < 0:
narrow_slice(i, -p[0], slice_length(i) + p[0])
if p[1] < 0:
narrow_slice(i, 0, slice_length(i) + p[1])
# crop grad_output if necessary
cg_output = grad_output
for i_s, p in zip(range(ctx.l_inp)[-ctx.l_pad:], ctx.pad_tup):
if p[0] > 0:
cg_output = cg_output.narrow(i_s, p[0], cg_output.size(i_s) - p[0])
if p[1] > 0:
cg_output = cg_output.narrow(i_s, 0, cg_output.size(i_s) - p[1])
gis = tuple(grad_input_slices)
grad_input[gis] = cg_output
return grad_input, None, None
def pad(input, pad, mode='constant', value=0):
assert len(pad) % 2 == 0, 'Padding length must be divisible by 2'
assert len(pad) // 2 <= input.dim(), 'Padding length too large'
if mode == 'constant':
return ConstantPadNd.apply(input, pad, value)
def local_response_norm(input, size, alpha=1e-4, beta=0.75, k=1):
r"""Applies local response normalization over an input signal composed of
several input planes, where channels occupy the second dimension.
Applies normalization across channels.
See :class:`~torch.nn.LocalResponseNorm` for details.
"""
dim = input.dim()
if dim < 3:
raise ValueError('Expected 3D or higher dimensionality \
input (got {} dimensions)'.format(dim))
div = input.mul(input).unsqueeze(1)
if dim == 3:
div = pad(div, (0, 0, size // 2, (size - 1) // 2))
div = F.avg_pool2d(div, (size, 1), stride=1).squeeze(1)
else:
sizes = input.size()
div = div.view(sizes[0], 1, sizes[1], sizes[2], -1)
div = pad(div, (0, 0, 0, 0, size // 2, (size - 1) // 2))
div = F.avg_pool3d(div, (size, 1, 1), stride=1).squeeze(1)
div = div.view(sizes)
div = div.mul(alpha).add(k).pow(beta)
return input / div
class Inception(nn.Module):
def __init__(self, in_channel, br_1x1, br_3x3_reduce, br_3x3,
br_5x5_reduce, br_5x5, pool_proj):
super(Inception, self).__init__()
self.add_module('1x1', nn.Conv2d(in_channel, br_1x1, kernel_size=1))
self.add_module('relu_1x1', nn.ReLU())
self.add_module('3x3_reduce', nn.Conv2d(in_channel, br_3x3_reduce, kernel_size=1))
self.add_module('relu_3x3_reduce', nn.ReLU())
self.add_module('3x3', nn.Conv2d(br_3x3_reduce, br_3x3, kernel_size=3, padding=1))
self.add_module('relu_3x3', nn.ReLU())
self.add_module('5x5_reduce', nn.Conv2d(in_channel, br_5x5_reduce, kernel_size=1))
self.add_module('relu_5x5_reduce', nn.ReLU())
self.add_module('5x5', nn.Conv2d(br_5x5_reduce, br_5x5, kernel_size=5, padding=2))
self.add_module('relu_5x5', nn.ReLU())
self.add_module('pool', nn.MaxPool2d(kernel_size=3, stride=1, padding=1))
self.add_module('pool_proj', nn.Conv2d(in_channel, pool_proj, kernel_size=1))
self.add_module('relu_pool_proj', nn.ReLU())
def forward(self, x):
x1 = getattr(self, '1x1')(x)
x1 = getattr(self, 'relu_1x1')(x1)
x2 = getattr(self, '3x3_reduce')(x)
x2 = getattr(self, 'relu_3x3_reduce')(x2)
x2 = getattr(self, '3x3')(x2)
x2 = getattr(self, 'relu_3x3')(x2)
x3 = getattr(self, '5x5_reduce')(x)
x3 = getattr(self, 'relu_5x5_reduce')(x3)
x3 = getattr(self, '5x5')(x3)
x3 = getattr(self, 'relu_5x5')(x3)
x4 = getattr(self, 'pool')(x)
x4 = getattr(self, 'pool_proj')(x4)
x4 = getattr(self, 'relu_pool_proj')(x4)
return torch.cat([x1, x2, x3, x4], dim=1)
class GoogleNet(nn.Module):
def __init__(self):
super(GoogleNet, self).__init__()
self.add_module('conv1.7x7_s2', nn.Conv2d(3, 64, kernel_size=7, stride=2, padding=3))
self.add_module('conv1.relu_7x7', nn.ReLU())
self.add_module('pool1', nn.MaxPool2d(3, stride=2, ceil_mode=True))
self.add_module('conv2.3x3_reduce', nn.Conv2d(64, 64, kernel_size=1))
self.add_module('conv2.relu_3x3_reduce', nn.ReLU())
self.add_module('conv2.3x3', nn.Conv2d(64, 192, kernel_size=3, padding=1))
self.add_module('conv2.relu_3x3', nn.ReLU())
self.add_module('pool2', nn.MaxPool2d(3, stride=2, ceil_mode=True))
self.inception_3a = Inception(192, 64, 96, 128, 16, 32, 32)
self.inception_3b = Inception(256, 128, 128, 192, 32, 96, 64)
self.add_module('pool3', nn.MaxPool2d(3, stride=2, ceil_mode=True))
self.inception_4a = Inception(480, 192, 96, 208, 16, 48, 64)
self.inception_4b = Inception(512, 160, 112, 224, 24, 64, 64)
self.inception_4c = Inception(512, 128, 128, 256, 24, 64, 64)
self.inception_4d = Inception(512, 112, 144, 288, 32, 64, 64)
self.inception_4e = Inception(528, 256, 160, 320, 32, 128, 128)
self.add_module('pool4', nn.MaxPool2d(3, stride=2, ceil_mode=True))
self.inception_5a = Inception(832, 256, 160, 320, 32, 128, 128)
self.inception_5b = Inception(832, 384, 192, 384, 48, 128, 128)
self.add_module('pool5', nn.AvgPool2d(7, stride=1))
self.add_module('drop', nn.Dropout2d(p=0.4))
self.add_module('loss3.classifier', nn.Linear(1024, 1000))
def forward(self, x):
x = getattr(self, 'conv1.7x7_s2')(x)
x = getattr(self, 'conv1.relu_7x7')(x)
x = self.pool1(x)
x = local_response_norm(x, 5)
x = getattr(self, 'conv2.3x3_reduce')(x)
x = getattr(self, 'conv2.relu_3x3_reduce')(x)
x = getattr(self, 'conv2.3x3')(x)
x = getattr(self, 'conv2.relu_3x3')(x)
x = local_response_norm(x, 5)
x = self.pool2(x)
x = self.inception_3a(x)
x = self.inception_3b(x)
x = self.pool3(x)
x = self.inception_4a(x)
x = self.inception_4b(x)
x = self.inception_4c(x)
x = self.inception_4d(x)
x = self.inception_4e(x)
x = self.pool4(x)
x = self.inception_5a(x)
x = self.inception_5b(x)
x = self.pool5(x)
x = self.drop(x)
x = x.view(-1, 1024)
x = getattr(self, 'loss3.classifier')(x)
return x
def get_googlenet(pretrain=False, pth_path='./weights/googlenet.pth'):
model = GoogleNet()
if pretrain:
model.load_state_dict(torch.load(pth_path))
return model