-
Notifications
You must be signed in to change notification settings - Fork 1
/
fresnet_srm.py
464 lines (405 loc) · 21.1 KB
/
fresnet_srm.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
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
# Licensed to the Apache Software Foundation (ASF) under one
# or more contributor license agreements. See the NOTICE file
# distributed with this work for additional information
# regarding copyright ownership. The ASF licenses this file
# to you under the Apache License, Version 2.0 (the
# "License"); you may not use this file except in compliance
# with the License. You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing,
# software distributed under the License is distributed on an
# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
# KIND, either express or implied. See the License for the
# specific language governing permissions and limitations
# under the License.
'''
Adapted from https://github.com/tornadomeet/ResNet/blob/master/symbol_resnet.py
Original author Wei Wu
Implemented the following paper:
Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun. "Identity Mappings in Deep Residual Networks"
'''
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import sys
import os
import mxnet as mx
import numpy as np
import symbol_utils
import memonger
import sklearn
sys.path.append(os.path.join(os.path.dirname(__file__), '..'))
from config import config
names =[0]
def Conv(**kwargs):
#name = kwargs.get('name')
#_weight = mx.symbol.Variable(name+'_weight')
#_bias = mx.symbol.Variable(name+'_bias', lr_mult=2.0, wd_mult=0.0)
#body = mx.sym.Convolution(weight = _weight, bias = _bias, **kwargs)
body = mx.sym.Convolution(**kwargs)
return body
def Act(data, act_type, name):
if act_type=='prelu':
body = mx.sym.LeakyReLU(data = data, act_type='prelu', name = name)
else:
body = mx.symbol.Activation(data=data, act_type=act_type, name=name)
return body
def residual_unit_v3(data, num_filter, stride, dim_match, name, bottle_neck, **kwargs):
### units = [3, 4, 14, 3],res50,use to sge reshape transport is 3 4 14 3
"""Return ResNet Unit symbol for building ResNet
Parameters
----------
data : str
Input data
num_filter : int
Number of output channels
bnf : int
Bottle neck channels factor with regard to num_filter
stride : tuple
Stride used in convolution
dim_match : Boolean
True means channel number between input and output is the same, otherwise means differ
name : str
Base name of the operators
workspace : int
Workspace used in convolution operator
"""
use_se = kwargs.get('version_se', 1)
bn_mom = kwargs.get('bn_mom', 0.9)
workspace = kwargs.get('workspace', 256)
memonger = kwargs.get('memonger', False)
act_type = kwargs.get('version_act', 'prelu')
use_srm = kwargs.get('version_srm', 1) #default 1 ,the input is 1,me add
#print('in unit3')
if bottle_neck:
bn1 = mx.sym.BatchNorm(data=data, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn1')
conv1 = Conv(data=bn1, num_filter=int(num_filter*0.25), kernel=(1,1), stride=(1,1), pad=(0,0),
no_bias=True, workspace=workspace, name=name + '_conv1')
bn2 = mx.sym.BatchNorm(data=conv1, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn2')
act1 = Act(data=bn2, act_type=act_type, name=name + '_relu1')
conv2 = Conv(data=act1, num_filter=int(num_filter*0.25), kernel=(3,3), stride=(1,1), pad=(1,1),
no_bias=True, workspace=workspace, name=name + '_conv2')
bn3 = mx.sym.BatchNorm(data=conv2, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn3')
act2 = Act(data=bn3, act_type=act_type, name=name + '_relu2')
conv3 = Conv(data=act2, num_filter=num_filter, kernel=(1,1), stride=stride, pad=(0,0), no_bias=True,
workspace=workspace, name=name + '_conv3')
bn4 = mx.sym.BatchNorm(data=conv3, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn4')
if use_se:
#se begin
body = mx.sym.Pooling(data=bn4, global_pool=True, kernel=(7, 7), pool_type='avg', name=name+'_se_pool1')
body = Conv(data=body, num_filter=num_filter//16, kernel=(1,1), stride=(1,1), pad=(0,0),
name=name+"_se_conv1", workspace=workspace)
body = Act(data=body, act_type=act_type, name=name+'_se_relu1')
body = Conv(data=body, num_filter=num_filter, kernel=(1,1), stride=(1,1), pad=(0,0),
name=name+"_se_conv2", workspace=workspace)
body = mx.symbol.Activation(data=body, act_type='sigmoid', name=name+"_se_sigmoid")
bn4 = mx.symbol.broadcast_mul(bn4, body)
#se end
if dim_match:
shortcut = data
else:
conv1sc = Conv(data=data, num_filter=num_filter, kernel=(1,1), stride=stride, no_bias=True,
workspace=workspace, name=name+'_conv1sc')
shortcut = mx.sym.BatchNorm(data=conv1sc, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_sc')
if memonger:
shortcut._set_attr(mirror_stage='True')
return bn4 + shortcut
else:
bn1 = mx.sym.BatchNorm(data=data, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn1')
conv1 = Conv(data=bn1, num_filter=num_filter, kernel=(3,3), stride=(1,1), pad=(1,1),
no_bias=True, workspace=workspace, name=name + '_conv1')
bn2 = mx.sym.BatchNorm(data=conv1, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn2')
act1 = Act(data=bn2, act_type=act_type, name=name + '_relu1')
conv2 = Conv(data=act1, num_filter=num_filter, kernel=(3,3), stride=stride, pad=(1,1),
no_bias=True, workspace=workspace, name=name + '_conv2')
bn3 = mx.sym.BatchNorm(data=conv2, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn3')
if use_se:
# print ("^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^")
#se begin
body = mx.sym.Pooling(data=bn3, global_pool=True, kernel=(7, 7), pool_type='avg', name=name+'_se_pool1')
body = Conv(data=body, num_filter=num_filter//16, kernel=(1,1), stride=(1,1), pad=(0,0),
name=name+"_se_conv1", workspace=workspace)
body = Act(data=body, act_type=act_type, name=name+'_se_relu1')
body = Conv(data=body, num_filter=num_filter, kernel=(1,1), stride=(1,1), pad=(0,0),
name=name+"_se_conv2", workspace=workspace)
body = mx.symbol.Activation(data=body, act_type='sigmoid', name=name+"_se_sigmoid")
bn3 = mx.symbol.broadcast_mul(bn3, body)
# data_shape = (32, 3, 112, 112)
# print(body.infer_shape(data=data_shape)[1][0])
#se end
if use_srm:
#print (name)
groups=64
batch = config.per_batch_size
data_shape = (batch, 3, 112, 112)
shape_size = bn3.infer_shape(data=data_shape)[1][0] #[bchw]
channel_mean = mx.sym.mean(mx.sym.reshape(bn3, shape=(0,0,-1)),axis=2,keepdims=True)
sub = mx.sym.broadcast_sub(mx.sym.reshape(bn3, shape=(0,0,-1)),channel_mean)
var =mx.sym.mean((sub**2),axis=2,keepdims=True) + 1e-5
channel_std = mx.symbol.sqrt(var)
srm_weight = mx.symbol.Variable(name+"_srm_weight"+name, shape=(shape_size[1],2),lr_mult=1.0,wd_mult=1.0,init=mx.init.Normal(0.001))#(c,2)
t = mx.symbol.concat(channel_mean,channel_std,dim=2) # B x C x 2
cfc = mx.sym.broadcast_mul(t,srm_weight) #bc2
z=mx.sym.sum(cfc,axis=2) #bc
### expand bc11
z=mx.sym.expand_dims(data=z,axis=-1)
z=mx.sym.expand_dims(data=z,axis=-1)
bn_srm=mx.sym.BatchNorm(data=z, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn_srm')
g=mx.symbol.Activation(data=bn_srm, act_type='sigmoid', name=name+"_srm_sigmoid")
# bchw *bc11
bn3 = mx.symbol.broadcast_mul(bn3, g)
#print(bn3.infer_shape(data=data_shape)[1][0])
###########################################################
if dim_match:
shortcut = data
else:
conv1sc = Conv(data=data, num_filter=num_filter, kernel=(1,1), stride=stride, no_bias=True,
workspace=workspace, name=name+'_conv1sc')
shortcut = mx.sym.BatchNorm(data=conv1sc, fix_gamma=False, momentum=bn_mom, eps=2e-5, name=name + '_sc')
if memonger:
shortcut._set_attr(mirror_stage='True')
return bn3 + shortcut
def residual_unit_v1_L(data, num_filter, stride, dim_match, name, bottle_neck, **kwargs):
"""Return ResNet Unit symbol for building ResNet
Parameters
----------
data : str
Input data
num_filter : int
Number of output channels
bnf : int
Bottle neck channels factor with regard to num_filter
stride : tuple
Stride used in convolution
dim_match : Boolean
True means channel number between input and output is the same, otherwise means differ
name : str
Base name of the operators
workspace : int
Workspace used in convolution operator
"""
use_se = kwargs.get('version_se', 1)
bn_mom = kwargs.get('bn_mom', 0.9)
workspace = kwargs.get('workspace', 256)
memonger = kwargs.get('memonger', False)
act_type = kwargs.get('version_act', 'prelu')
#print('in unit1')
if bottle_neck:##False is 3*3 3*3 not 1*1
conv1 = Conv(data=data, num_filter=int(num_filter*0.25), kernel=(1,1), stride=(1,1), pad=(0,0),
no_bias=True, workspace=workspace, name=name + '_conv1')
bn1 = mx.sym.BatchNorm(data=conv1, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn1')
act1 = Act(data=bn1, act_type=act_type, name=name + '_relu1')
conv2 = Conv(data=act1, num_filter=int(num_filter*0.25), kernel=(3,3), stride=(1,1), pad=(1,1),
no_bias=True, workspace=workspace, name=name + '_conv2')
bn2 = mx.sym.BatchNorm(data=conv2, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn2')
act2 = Act(data=bn2, act_type=act_type, name=name + '_relu2')
conv3 = Conv(data=act2, num_filter=num_filter, kernel=(1,1), stride=stride, pad=(0,0), no_bias=True,
workspace=workspace, name=name + '_conv3')
bn3 = mx.sym.BatchNorm(data=conv3, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn3')
if use_se:
#se begin
body = mx.sym.Pooling(data=bn3, global_pool=True, kernel=(7, 7), pool_type='avg', name=name+'_se_pool1')
body = Conv(data=body, num_filter=num_filter//16, kernel=(1,1), stride=(1,1), pad=(0,0),
name=name+"_se_conv1", workspace=workspace)
body = Act(data=body, act_type=act_type, name=name+'_se_relu1')
body = Conv(data=body, num_filter=num_filter, kernel=(1,1), stride=(1,1), pad=(0,0),
name=name+"_se_conv2", workspace=workspace)
body = mx.symbol.Activation(data=body, act_type='sigmoid', name=name+"_se_sigmoid")
bn3 = mx.symbol.broadcast_mul(bn3, body)
#se end
if dim_match:
shortcut = data
else:
conv1sc = Conv(data=data, num_filter=num_filter, kernel=(1,1), stride=stride, no_bias=True,
workspace=workspace, name=name+'_conv1sc')
shortcut = mx.sym.BatchNorm(data=conv1sc, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_sc')
if memonger:
shortcut._set_attr(mirror_stage='True')
return Act(data=bn3 + shortcut, act_type=act_type, name=name + '_relu3')
else:
conv1 = Conv(data=data, num_filter=num_filter, kernel=(3,3), stride=(1,1), pad=(1,1),
no_bias=True, workspace=workspace, name=name + '_conv1')
bn1 = mx.sym.BatchNorm(data=conv1, fix_gamma=False, momentum=bn_mom, eps=2e-5, name=name + '_bn1')
act1 = Act(data=bn1, act_type=act_type, name=name + '_relu1')
conv2 = Conv(data=act1, num_filter=num_filter, kernel=(3,3), stride=stride, pad=(1,1),
no_bias=True, workspace=workspace, name=name + '_conv2')
bn2 = mx.sym.BatchNorm(data=conv2, fix_gamma=False, momentum=bn_mom, eps=2e-5, name=name + '_bn2')
if use_se: #config config.net_se = 0 not to use se
#se begin
body = mx.sym.Pooling(data=bn2, global_pool=True, kernel=(7, 7), pool_type='avg', name=name+'_se_pool1')
body = Conv(data=body, num_filter=num_filter//16, kernel=(1,1), stride=(1,1), pad=(0,0),
name=name+"_se_conv1", workspace=workspace)
body = Act(data=body, act_type=act_type, name=name+'_se_relu1')
body = Conv(data=body, num_filter=num_filter, kernel=(1,1), stride=(1,1), pad=(0,0),
name=name+"_se_conv2", workspace=workspace)
body = mx.symbol.Activation(data=body, act_type='sigmoid', name=name+"_se_sigmoid")
bn2 = mx.symbol.broadcast_mul(bn2, body)
#se end
if dim_match:
shortcut = data
else:
conv1sc = Conv(data=data, num_filter=num_filter, kernel=(1,1), stride=stride, no_bias=True,
workspace=workspace, name=name+'_conv1sc')
shortcut = mx.sym.BatchNorm(data=conv1sc, fix_gamma=False, momentum=bn_mom, eps=2e-5, name=name + '_sc')
if memonger:
shortcut._set_attr(mirror_stage='True')
return Act(data=bn2 + shortcut, act_type=act_type, name=name + '_relu3')
def residual_unit(data, num_filter, stride, dim_match, name, bottle_neck, **kwargs):
uv = kwargs.get('version_unit', 3) #default is 3 and tansfer is no, 'version_input': config.net_input,is 3
version_input = kwargs.get('version_input', 1)
if uv==1:
if version_input==0:
return residual_unit_v1(data, num_filter, stride, dim_match, name, bottle_neck, **kwargs)
else:
return residual_unit_v1_L(data, num_filter, stride, dim_match, name, bottle_neck, **kwargs) #select this
if uv==3: ##select
return residual_unit_v3(data, num_filter, stride, dim_match, name, bottle_neck, **kwargs)
def resnet(units, num_stages, filter_list, num_classes, bottle_neck):
bn_mom = config.bn_mom
workspace = config.workspace
kwargs = {'version_se' : config.net_se,
'version_input': config.net_input,
'version_output': config.net_output,
'version_unit': config.net_unit,
'version_act': config.net_act,
'bn_mom': bn_mom,
'workspace': workspace,
'memonger': config.memonger,
'version_srm' : config.net_srm,
}
"""Return ResNet symbol of
Parameters
----------
units : list
Number of units in each stage
num_stages : int
Number of stage
filter_list : list
Channel size of each stage
num_classes : int
Ouput size of symbol
dataset : str
Dataset type, only cifar10 and imagenet supports
workspace : int
Workspace used in convolution operator
"""
version_se = kwargs.get('version_se', 1)
version_input = kwargs.get('version_input', 1)
assert version_input>=0
version_output = kwargs.get('version_output', 'E')
fc_type = version_output
version_unit = kwargs.get('version_unit', 3)
act_type = kwargs.get('version_act', 'prelu')
memonger = kwargs.get('memonger', False)
print(version_se, version_input, version_output, version_unit, act_type, memonger)
num_unit = len(units)
assert(num_unit == num_stages)
data = mx.sym.Variable(name='data')
if version_input==0:
#data = mx.sym.BatchNorm(data=data, fix_gamma=True, eps=2e-5, momentum=bn_mom, name='bn_data')
data = mx.sym.identity(data=data, name='id')
data = data-127.5
data = data*0.0078125
body = Conv(data=data, num_filter=filter_list[0], kernel=(7, 7), stride=(2,2), pad=(3, 3),
no_bias=True, name="conv0", workspace=workspace)
body = mx.sym.BatchNorm(data=body, fix_gamma=False, eps=2e-5, momentum=bn_mom, name='bn0')
body = Act(data=body, act_type=act_type, name='relu0')
#body = mx.sym.Pooling(data=body, kernel=(3, 3), stride=(2,2), pad=(1,1), pool_type='max')
elif version_input==2:
data = mx.sym.BatchNorm(data=data, fix_gamma=True, eps=2e-5, momentum=bn_mom, name='bn_data')
body = Conv(data=data, num_filter=filter_list[0], kernel=(3,3), stride=(1,1), pad=(1,1),
no_bias=True, name="conv0", workspace=workspace)
body = mx.sym.BatchNorm(data=body, fix_gamma=False, eps=2e-5, momentum=bn_mom, name='bn0')
body = Act(data=body, act_type=act_type, name='relu0')
else: #version_input==1 excute
data = mx.sym.identity(data=data, name='id')
data = data-127.5
data = data*0.0078125
body = data
body = Conv(data=body, num_filter=filter_list[0], kernel=(3,3), stride=(1,1), pad=(1, 1), ##same convolution311,112*112
no_bias=True, name="conv0", workspace=workspace) #this is L 7*7 to 3*3
body = mx.sym.BatchNorm(data=body, fix_gamma=False, eps=2e-5, momentum=bn_mom, name='bn0')
body = Act(data=body, act_type=act_type, name='relu0')
for i in range(num_stages): #4
#if version_input==0:
# body = residual_unit(body, filter_list[i+1], (1 if i==0 else 2, 1 if i==0 else 2), False,
# name='stage%d_unit%d' % (i + 1, 1), bottle_neck=bottle_neck, **kwargs)
#else:
# body = residual_unit(body, filter_list[i+1], (2, 2), False,
# name='stage%d_unit%d' % (i + 1, 1), bottle_neck=bottle_neck, **kwargs)
### # shortcut have conv and bn , reduce size 321 112 to 56
body = residual_unit(body, filter_list[i+1], (2, 2), False, ##filter_list = [64, 64, 128, 256, 512],False input output dinmension different
name='stage%d_unit%d' % (i + 1, 1), bottle_neck=bottle_neck, **kwargs) #
### ## # shortcut no conv and bn no reduce, 56 to 56 ,repeat 3 times
for j in range(units[i]-1): ## #units = [3, 13, 30, 3] #3 residual_unit
body = residual_unit(body, filter_list[i+1], (1,1), True, name='stage%d_unit%d' % (i+1, j+2),
bottle_neck=bottle_neck, **kwargs)
if bottle_neck: ## False
body = Conv(data=body, num_filter=512, kernel=(1,1), stride=(1,1), pad=(0,0),
no_bias=True, name="convd", workspace=workspace)
body = mx.sym.BatchNorm(data=body, fix_gamma=False, eps=2e-5, momentum=bn_mom, name='bnd')
body = Act(data=body, act_type=act_type, name='relud')
fc1 = symbol_utils.get_fc1(body, num_classes, fc_type) #c=512 dimension
return fc1
def get_symbol(): ##embedding = eval(config.net_name).get_symbol()
"""
Adapted from https://github.com/tornadomeet/ResNet/blob/master/train_resnet.py
Original author Wei Wu
"""
num_classes = config.emb_size
num_layers = config.num_layers
if num_layers >= 500:
filter_list = [64, 256, 512, 1024, 2048]
bottle_neck = True
else:
filter_list = [64, 64, 128, 256, 512] #this is block out c
bottle_neck = False
num_stages = 4
if num_layers == 18:
units = [2, 2, 2, 2]
elif num_layers == 34:
units = [3, 4, 6, 3]
elif num_layers == 49:
units = [3, 4, 14, 3]
elif num_layers == 50:
units = [3, 4, 14, 3]
elif num_layers == 74:
units = [3, 6, 24, 3]
elif num_layers == 90:
units = [3, 8, 30, 3]
elif num_layers == 98:
units = [3, 4, 38, 3]
elif num_layers == 99:
units = [3, 8, 35, 3]
elif num_layers == 100:
units = [3, 13, 30, 3]
elif num_layers == 124:
units = [3, 13, 40, 5]
elif num_layers == 160:
units = [3, 24, 49, 3]
elif num_layers == 101:
units = [3, 4, 23, 3]
elif num_layers == 152:
units = [3, 8, 36, 3]
elif num_layers == 200:
units = [3, 24, 36, 3]
elif num_layers == 269:
units = [3, 30, 48, 8]
else:
raise ValueError("no experiments done on num_layers {}, you can do it yourself".format(num_layers))
net = resnet(units = units,
num_stages = num_stages, ##4
filter_list = filter_list, #[64, 64, 128, 256, 512]
num_classes = num_classes, #config.emb_size=512
bottle_neck = bottle_neck) #False
if config.memonger: #False
dshape = (config.per_batch_size, config.image_shape[2], config.image_shape[0], config.image_shape[1])
net_mem_planned = memonger.search_plan(net, data=dshape)
old_cost = memonger.get_cost(net, data=dshape)
new_cost = memonger.get_cost(net_mem_planned, data=dshape)
print('Old feature map cost=%d MB' % old_cost)
print('New feature map cost=%d MB' % new_cost)
net = net_mem_planned
return net