-
Notifications
You must be signed in to change notification settings - Fork 4
/
mixnet.py
377 lines (313 loc) · 16.5 KB
/
mixnet.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
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
# ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ #
# LightNet++: Boosted Light-weighted Networks for Real-time Semantic Segmentation
# ---------------------------------------------------------------------------------------------------------------- #
# PyTorch implementation for MixNet
# class:
# > Swish
# > SEBlock
# > GPConv
# > MDConv
# > MixDepthBlock
# > MixNet(S, M, L)
# ---------------------------------------------------------------------------------------------------------------- #
# Author: Huijun Liu M.Sc.
# Date: 15.02.2020
# ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ #
from torch.nn import functional as F
from collections import OrderedDict
from torch import nn
import torch
import math
# ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ #
# Swish: Swish Activation Function
# ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ #
class Swish(nn.Module):
def __init__(self, inplace=True):
super(Swish, self).__init__()
self.inplace = inplace
def forward(self, x):
return x.mul_(x.sigmoid()) if self.inplace else x.mul(x.sigmoid())
# ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ #
# SEBlock: Squeeze & Excitation (SCSE)
# namely, Channel-wise Attention
# ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ #
class SEBlock(nn.Module):
def __init__(self, in_planes, reduced_dim, act_type="swish"):
super(SEBlock, self).__init__()
self.channel_se = nn.Sequential(OrderedDict([
("linear1", nn.Conv2d(in_planes, reduced_dim, kernel_size=1, stride=1, padding=0, bias=True)),
("act", Swish(inplace=True) if act_type == "swish" else nn.ReLU(inplace=True)),
("linear2", nn.Conv2d(reduced_dim, in_planes, kernel_size=1, stride=1, padding=0, bias=True))
]))
def forward(self, x):
x_se = torch.sigmoid(self.channel_se(F.adaptive_avg_pool2d(x, output_size=(1, 1))))
return torch.mul(x, x_se)
class ConvBlock(nn.Module):
def __init__(self, in_planes, out_planes, kernel_size, stride=1,
groups=1, dilate=1, act_type="swish"):
super(ConvBlock, self).__init__()
assert stride in [1, 2]
dilate = 1 if stride > 1 else dilate
padding = ((kernel_size - 1) // 2) * dilate
self.conv_block = nn.Sequential(OrderedDict([
("conv", nn.Conv2d(in_channels=in_planes, out_channels=out_planes,
kernel_size=kernel_size, stride=stride, padding=padding,
dilation=dilate, groups=groups, bias=False)),
("norm", nn.BatchNorm2d(num_features=out_planes,
eps=1e-3, momentum=0.01)),
("act", Swish(inplace=True) if act_type == "swish" else nn.ReLU(inplace=True))
]))
def forward(self, x):
return self.conv_block(x)
# ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ #
# GPConv: Grouped Point-wise Convolution for MixDepthBlock
# ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ #
class GPConv(nn.Module):
def __init__(self, in_planes, out_planes, kernel_sizes):
super(GPConv, self).__init__()
self.num_groups = len(kernel_sizes)
assert in_planes % self.num_groups == 0
sub_in_dim = in_planes // self.num_groups
sub_out_dim = out_planes // self.num_groups
self.group_point_wise = nn.ModuleList()
for _ in kernel_sizes:
self.group_point_wise.append(nn.Conv2d(sub_in_dim, sub_out_dim,
kernel_size=1, stride=1, padding=0,
groups=1, dilation=1, bias=False))
def forward(self, x):
if self.num_groups == 1:
return self.group_point_wise[0](x)
chunks = torch.chunk(x, chunks=self.num_groups, dim=1)
mix = [self.group_point_wise[stream](chunks[stream]) for stream in range(self.num_groups)]
return torch.cat(mix, dim=1)
# ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ #
# MDConv: Mixed Depth-wise Convolution for MixDepthBlock
# ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ #
class MDConv(nn.Module):
def __init__(self, in_planes, kernel_sizes, stride=1, dilate=1):
super(MDConv, self).__init__()
self.num_groups = len(kernel_sizes)
assert in_planes % self.num_groups == 0
sub_hidden_dim = in_planes // self.num_groups
assert stride in [1, 2]
dilate = 1 if stride > 1 else dilate
self.mixed_depth_wise = nn.ModuleList()
for kernel_size in kernel_sizes:
padding = ((kernel_size - 1) // 2) * dilate
self.mixed_depth_wise.append(nn.Conv2d(sub_hidden_dim, sub_hidden_dim,
kernel_size=kernel_size, stride=stride, padding=padding,
groups=sub_hidden_dim, dilation=dilate, bias=False))
def forward(self, x):
if self.num_groups == 1:
return self.mixed_depth_wise[0](x)
chunks = torch.chunk(x, chunks=self.num_groups, dim=1)
mix = [self.mixed_depth_wise[stream](chunks[stream]) for stream in range(self.num_groups)]
return torch.cat(mix, dim=1)
# ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ #
# MixDepthBlock: MixDepthBlock for MixNet
# ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ #
class MixDepthBlock(nn.Module):
def __init__(self, in_planes, out_planes,
expand_ratio, exp_kernel_sizes, kernel_sizes, poi_kernel_sizes, stride, dilate,
reduction_ratio=4, dropout_rate=0.2, act_type="swish"):
super(MixDepthBlock, self).__init__()
self.dropout_rate = dropout_rate
self.expand_ratio = expand_ratio
self.groups = len(kernel_sizes)
self.use_se = (reduction_ratio is not None) and (reduction_ratio > 1)
self.use_residual = in_planes == out_planes and stride == 1
assert stride in [1, 2]
dilate = 1 if stride > 1 else dilate
hidden_dim = in_planes * expand_ratio
# step 1. Expansion phase/Point-wise convolution
if expand_ratio != 1:
self.expansion = nn.Sequential(OrderedDict([
("conv", GPConv(in_planes, hidden_dim, kernel_sizes=exp_kernel_sizes)),
("norm", nn.BatchNorm2d(hidden_dim, eps=1e-3, momentum=0.01)),
("act", Swish(inplace=True) if act_type == "swish" else nn.ReLU(inplace=True))
]))
# step 2. Depth-wise convolution phase
self.depth_wise = nn.Sequential(OrderedDict([
("conv", MDConv(hidden_dim, kernel_sizes=kernel_sizes, stride=stride, dilate=dilate)),
("norm", nn.BatchNorm2d(hidden_dim, eps=1e-3, momentum=0.01)),
("act", Swish(inplace=True) if act_type == "swish" else nn.ReLU(inplace=True))
]))
# step 3. Squeeze and Excitation
if self.use_se:
reduced_dim = max(1, int(in_planes / reduction_ratio))
self.se_block = SEBlock(hidden_dim, reduced_dim, act_type=act_type)
# step 4. Point-wise convolution phase
self.point_wise = nn.Sequential(OrderedDict([
("conv", GPConv(hidden_dim, out_planes, kernel_sizes=poi_kernel_sizes)),
("norm", nn.BatchNorm2d(out_planes, eps=1e-3, momentum=0.01))
]))
def forward(self, x):
res = x
# step 1. Expansion phase/Point-wise convolution
if self.expand_ratio != 1:
x = self.expansion(x)
# step 2. Depth-wise convolution phase
x = self.depth_wise(x)
# step 3. Squeeze and Excitation
if self.use_se:
x = self.se_block(x)
# step 4. Point-wise convolution phase
x = self.point_wise(x)
# step 5. Skip connection and drop connect
if self.use_residual:
if self.training and (self.dropout_rate is not None):
x = F.dropout2d(input=x, p=self.dropout_rate,
training=self.training, inplace=True)
x = x + res
return x
class MixNet(nn.Module):
def __init__(self, arch="s", num_classes=1000):
super(MixNet, self).__init__()
params = {
's': (16, [
# t, c, n, k, ek, pk, s, d, a, se
[1, 16, 1, [3], [1], [1], 1, 1, "relu", None],
[6, 24, 1, [3], [1, 1], [1, 1], 2, 1, "relu", None],
[3, 24, 1, [3], [1, 1], [1, 1], 1, 1, "relu", None],
[6, 40, 1, [3, 5, 7], [1], [1], 2, 1, "swish", 2],
[6, 40, 3, [3, 5], [1, 1], [1, 1], 1, 1, "swish", 2],
[6, 80, 1, [3, 5, 7], [1], [1, 1], 2, 1, "swish", 4],
[6, 80, 2, [3, 5], [1], [1, 1], 1, 1, "swish", 4],
[6, 120, 1, [3, 5, 7], [1, 1], [1, 1], 1, 1, "swish", 2],
[3, 120, 2, [3, 5, 7, 9], [1, 1], [1, 1], 1, 1, "swish", 2],
[6, 200, 1, [3, 5, 7, 9, 11], [1], [1], 2, 1, "swish", 2],
[6, 200, 2, [3, 5, 7, 9], [1], [1, 1], 1, 1, "swish", 2]
], 1.0, 1.0, 0.2),
'm': (24, [
# t, c, n, k, ek, pk, s, d, a, se
[1, 24, 1, [3], [1], [1], 1, 1, "relu", None],
[6, 32, 1, [3, 5, 7], [1, 1], [1, 1], 2, 1, "relu", None],
[3, 32, 1, [3], [1, 1], [1, 1], 1, 1, "relu", None],
[6, 40, 1, [3, 5, 7, 9], [1], [1], 2, 1, "swish", 2],
[6, 40, 3, [3, 5], [1, 1], [1, 1], 1, 1, "swish", 2],
[6, 80, 1, [3, 5, 7], [1], [1], 2, 1, "swish", 4],
[6, 80, 3, [3, 5, 7, 9], [1, 1], [1, 1], 1, 1, "swish", 4],
[6, 120, 1, [3], [1], [1], 1, 1, "swish", 2],
[3, 120, 3, [3, 5, 7, 9], [1, 1], [1, 1], 1, 1, "swish", 2],
[6, 200, 1, [3, 5, 7, 9], [1], [1], 2, 1, "swish", 2],
[6, 200, 3, [3, 5, 7, 9], [1], [1, 1], 1, 1, "swish", 2]
], 1.0, 1.0, 0.25),
'l': (24, [
# t, c, n, k, ek, pk, s, d, a, se
[1, 24, 1, [3], [1], [1], 1, 1, "relu", None],
[6, 32, 1, [3, 5, 7], [1, 1], [1, 1], 2, 1, "relu", None],
[3, 32, 1, [3], [1, 1], [1, 1], 1, 1, "relu", None],
[6, 40, 1, [3, 5, 7, 9], [1], [1], 2, 1, "swish", 2],
[6, 40, 3, [3, 5], [1, 1], [1, 1], 1, 1, "swish", 2],
[6, 80, 1, [3, 5, 7], [1], [1], 2, 1, "swish", 4],
[6, 80, 3, [3, 5, 7, 9], [1, 1], [1, 1], 1, 1, "swish", 4],
[6, 120, 1, [3], [1], [1], 1, 1, "swish", 2],
[3, 120, 3, [3, 5, 7, 9], [1, 1], [1, 1], 1, 1, "swish", 2],
[6, 200, 1, [3, 5, 7, 9], [1], [1], 2, 1, "swish", 2],
[6, 200, 3, [3, 5, 7, 9], [1], [1, 1], 1, 1, "swish", 2]
], 1.3, 1.0, 0.25),
}
stem_planes, settings, width_multi, depth_multi, self.dropout_rate = params[arch]
out_channels = self._round_filters(stem_planes, width_multi)
self.mod1 = ConvBlock(3, out_channels, kernel_size=3, stride=2,
groups=1, dilate=1, act_type="relu")
in_channels = out_channels
drop_rate = self.dropout_rate
mod_id = 0
for t, c, n, k, ek, pk, s, d, a, se in settings:
out_channels = self._round_filters(c, width_multi)
repeats = self._round_repeats(n, depth_multi)
if self.dropout_rate:
drop_rate = self.dropout_rate * float(mod_id+1) / len(settings)
# Create blocks for module
blocks = []
for block_id in range(repeats):
stride = s if block_id == 0 else 1
dilate = d if stride == 1 else 1
blocks.append(("block%d" % (block_id + 1), MixDepthBlock(in_channels, out_channels,
expand_ratio=t, exp_kernel_sizes=ek,
kernel_sizes=k, poi_kernel_sizes=pk,
stride=stride, dilate=dilate,
reduction_ratio=se,
dropout_rate=drop_rate,
act_type=a)))
in_channels = out_channels
self.add_module("mod%d" % (mod_id + 2), nn.Sequential(OrderedDict(blocks)))
mod_id += 1
self.last_channels = 1536
self.last_feat = ConvBlock(in_channels, self.last_channels,
kernel_size=1, stride=1,
groups=1, dilate=1, act_type="relu")
self.classifier = nn.Linear(self.last_channels, num_classes)
self._initialize_weights()
def _initialize_weights(self):
# weight initialization
for m in self.modules():
if isinstance(m, nn.Conv2d):
nn.init.kaiming_normal_(m.weight, mode='fan_out')
if m.bias is not None:
nn.init.zeros_(m.bias)
elif isinstance(m, nn.BatchNorm2d):
nn.init.ones_(m.weight)
nn.init.zeros_(m.bias)
elif isinstance(m, nn.Linear):
fan_out = m.weight.size(0)
init_range = 1.0 / math.sqrt(fan_out)
nn.init.uniform_(m.weight, -init_range, init_range)
if m.bias is not None:
nn.init.zeros_(m.bias)
@staticmethod
def _make_divisible(value, divisor=8):
new_value = max(divisor, int(value + divisor / 2) // divisor * divisor)
if new_value < 0.9 * value:
new_value += divisor
return new_value
def _round_filters(self, filters, width_multi):
if width_multi == 1.0:
return filters
return int(self._make_divisible(filters * width_multi))
@staticmethod
def _round_repeats(repeats, depth_multi):
if depth_multi == 1.0:
return repeats
return int(math.ceil(depth_multi * repeats))
def forward(self, x):
x = self.mod2(self.mod1(x)) # (N, C, H/2, W/2)
x = self.mod4(self.mod3(x)) # (N, C, H/4, W/4)
x = self.mod6(self.mod5(x)) # (N, C, H/8, W/8)
x = self.mod10(self.mod9(self.mod8(self.mod7(x)))) # (N, C, H/16, W/16)
x = self.mod12(self.mod11(x)) # (N, C, H/32, W/32)
x = self.last_feat(x)
x = F.adaptive_avg_pool2d(x, (1, 1)).view(-1, self.last_channels)
if self.training and (self.dropout_rate is not None):
x = F.dropout(input=x, p=self.dropout_rate,
training=self.training, inplace=True)
x = self.classifier(x)
return x
if __name__ == "__main__":
import os
import time
from torchstat import stat
from pytorch_memlab import MemReporter
os.environ["CUDA_VISIBLE_DEVICES"] = "0"
arch = "l"
img_preparam = {"s": (224, 0.875), "m": (224, 0.875), "l": (224, 0.875)}
net_h = img_preparam[arch][0]
model = MixNet(arch=arch, num_classes=1000)
optimizer = torch.optim.SGD(model.parameters(), lr=1e-1,
momentum=0.90, weight_decay=1.0e-4, nesterov=True)
# stat(model, (3, net_h, net_h))
model = model.cuda().train()
loss_func = nn.CrossEntropyLoss().cuda()
dummy_in = torch.randn(2, 3, net_h, net_h).cuda().requires_grad_()
dummy_target = torch.ones(2).cuda().long().cuda()
reporter = MemReporter(model)
optimizer.zero_grad()
dummy_out = model(dummy_in)
loss = loss_func(dummy_out, dummy_target)
print('========================================== before backward ===========================================')
reporter.report()
loss.backward()
optimizer.step()
print('========================================== after backward =============================================')
reporter.report()