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
|
import logging
logging.basicConfig()
log = logging.getLogger("AnyExp")
log.setLevel(logging.DEBUG)
# For more depths, add the block config here
BLOCK_CONFIG = {
18: (2, 2, 2, 2),
34: (3, 4, 6, 3),
50: (3, 4, 6, 3),
101: (3, 4, 23, 3),
152: (3, 8, 36, 3),
200: (3, 32, 36, 3),
264: (3, 64, 36, 3),
284: (3, 32, 64, 3),
}
def gen_forward_pass_builder_fun(self, model, dataset, is_train):
split = 'train' if is_train else 'test'
opts = self.opts
def model_creator(model, loss_scale):
model, softmax, loss = resnet_imagenet_create_model(
model=model,
data='data',
labels='label',
split=split,
opts=opts,
dataset=dataset,
)
return [loss]
return model_creator
def resnet_imagenet_create_model(model, data, labels, split, opts, dataset):
model_helper = ResNetModelHelper(model, split, opts)
opts_depth = opts['model_param']['num_layer']
engine = opts['model_param']['engine']
log.info(' | ResNet-{} Imagenet'.format(opts_depth))
assert opts_depth in BLOCK_CONFIG.keys(), \
'Block config is not defined for specified model depth. Please check.'
(n1, n2, n3, n4) = BLOCK_CONFIG[opts_depth]
num_features = 2048
residual_block = model_helper.bottleneck_block
if opts_depth in [18, 34]:
num_features = 512
residual_block = model_helper.basic_block
num_classes = 1000
conv_blob = model.Conv(
data, 'conv1', 3, 64, 7, stride=2, pad=3, weight_init=('MSRAFill', {}),
bias_init=('ConstantFill', {'value': 0.}), no_bias=0, engine=engine
)
test_mode = False
if split in ['test', 'val']:
test_mode = True
bn_blob = model.SpatialBN(
conv_blob, 'res_conv1_bn', 64,
# does not appear to affect test_loss performance
# epsilon=1e-3,
epsilon=opts['model_param']['bn_epsilon'],
# momentum=0.1,
momentum=opts['model_param']['bn_momentum'],
is_test=test_mode,
)
relu_blob = model.Relu(bn_blob, bn_blob)
max_pool = model.MaxPool(relu_blob, 'pool1', kernel=3, stride=2, pad=1)
# TODO: This can be further optimized by passing dim_in, dim_out = features,
# dim_out = features * 4
if opts_depth in [50, 101, 152, 200, 264, 284]:
blob_in, dim_in = model_helper.residual_layer(
residual_block, max_pool, 64, 256, stride=1, num_blocks=n1,
prefix='res2', dim_inner=64
)
blob_in, dim_in = model_helper.residual_layer(
residual_block, blob_in, dim_in, 512, stride=2, num_blocks=n2,
prefix='res3', dim_inner=128
)
blob_in, dim_in = model_helper.residual_layer(
residual_block, blob_in, dim_in, 1024, stride=2, num_blocks=n3,
prefix='res4', dim_inner=256
)
blob_in, dim_in = model_helper.residual_layer(
residual_block, blob_in, dim_in, 2048, stride=2, num_blocks=n4,
prefix='res5', dim_inner=512
)
elif opts_depth in [18, 34]:
blob_in, dim_in = model_helper.residual_layer(
residual_block, max_pool, 64, 64, stride=1, num_blocks=n1,
prefix='res2',
)
blob_in, dim_in = model_helper.residual_layer(
residual_block, blob_in, dim_in, 128, stride=2, num_blocks=n2,
prefix='res3',
)
blob_in, dim_in = model_helper.residual_layer(
residual_block, blob_in, dim_in, 256, stride=2, num_blocks=n3,
prefix='res4',
)
blob_in, dim_in = model_helper.residual_layer(
residual_block, blob_in, dim_in, 512, stride=2, num_blocks=n4,
prefix='res5',
)
pool_blob = model.AveragePool(blob_in, 'pool5', kernel=7, stride=1)
loss_scale = 1. / opts['distributed']['num_xpus'] / \
opts['distributed']['num_shards']
loss = None
fc_blob = model.FC(
pool_blob, 'pred', num_features, num_classes,
# does not appear to affect test_loss performance
# weight_init=('GaussianFill', {'std': opts.fc_init_std}),
# bias_init=('ConstantFill', {'value': 0.})
weight_init=None,
bias_init=None)
softmax, loss = model.SoftmaxWithLoss(
[fc_blob, labels],
['softmax', 'loss'],
scale=loss_scale)
model.Accuracy(['softmax', labels], 'accuracy')
return model, softmax, loss
class ResNetModelHelper():
def __init__(self, model, split, opts):
self.model = model
self.split = split
self.opts = opts
self.engine = opts['model_param']['engine']
# shortcut type B
def add_shortcut(self, blob_in, dim_in, dim_out, stride, prefix):
if dim_in == dim_out:
return blob_in
conv_blob = self.model.Conv(
blob_in, prefix, dim_in, dim_out, kernel=1,
stride=stride,
weight_init=("MSRAFill", {}),
bias_init=('ConstantFill', {'value': 0.}), no_bias=1, engine=self.engine
)
test_mode = False
if self.split in ['test', 'val']:
test_mode = True
bn_blob = self.model.SpatialBN(
conv_blob, prefix + "_bn", dim_out,
# epsilon=1e-3,
# momentum=0.1,
epsilon=self.opts['model_param']['bn_epsilon'],
momentum=self.opts['model_param']['bn_momentum'],
is_test=test_mode,
)
return bn_blob
def conv_bn(
self, blob_in, dim_in, dim_out, kernel, stride, prefix, group=1, pad=1,
):
conv_blob = self.model.Conv(
blob_in, prefix, dim_in, dim_out, kernel, stride=stride,
pad=pad, group=group,
weight_init=("MSRAFill", {}),
bias_init=('ConstantFill', {'value': 0.}), no_bias=1, engine=self.engine
)
test_mode = False
if self.split in ['test', 'val']:
test_mode = True
bn_blob = self.model.SpatialBN(
conv_blob, prefix + "_bn", dim_out,
epsilon=self.opts['model_param']['bn_epsilon'],
momentum=self.opts['model_param']['bn_momentum'],
is_test=test_mode,
)
return bn_blob
def conv_bn_relu(
self, blob_in, dim_in, dim_out, kernel, stride, prefix, pad=1, group=1,
):
bn_blob = self.conv_bn(
blob_in, dim_in, dim_out, kernel, stride, prefix, group=group,
pad=pad
)
return self.model.Relu(bn_blob, bn_blob)
# 3(a)this block uses multi-way group conv implementation that splits blobs
def multiway_bottleneck_block(
self, blob_in, dim_in, dim_out, stride, prefix, dim_inner, group
):
blob_out = self.conv_bn_relu(
blob_in, dim_in, dim_inner, 1, 1, prefix + "_branch2a", pad=0,
)
conv_blob = self.model.GroupConv_Deprecated(
blob_out, prefix + "_branch2b", dim_inner, dim_inner, kernel=3,
stride=stride, pad=1, group=group, weight_init=("MSRAFill", {}),
bias_init=('ConstantFill', {'value': 0.}), no_bias=1, engine=self.engine
)
test_mode = False
if self.split in ['test', 'val']:
test_mode = True
bn_blob = self.model.SpatialBN(
conv_blob, prefix + "_branch2b_bn", dim_out,
epsilon=self.opts['model_param']['bn_epsilon'],
momentum=self.opts['model_param']['bn_momentum'], is_test=test_mode,
)
relu_blob = self.model.Relu(bn_blob, bn_blob)
bn_blob = self.conv_bn(
relu_blob, dim_inner, dim_out, 1, 1, prefix + "_branch2c", pad=0
)
if self.opts['model_param']['custom_bn_init']:
self.model.param_init_net.ConstantFill(
[bn_blob + '_s'], bn_blob + '_s',
value=self.opts['model_param']['bn_init_gamma'])
sc_blob = self.add_shortcut(
blob_in, dim_in, dim_out, stride, prefix=prefix + "_branch1"
)
sum_blob = self.model.net.Sum([bn_blob, sc_blob], prefix + "_sum")
return self.model.Relu(sum_blob, sum_blob)
# 3(c) this block uses cudnn group conv op
def group_bottleneck_block(
self, blob_in, dim_in, dim_out, stride, prefix, dim_inner, group
):
blob_out = self.conv_bn_relu(
blob_in, dim_in, dim_inner, 1, 1, prefix + "_branch2a", pad=0,
)
blob_out = self.conv_bn_relu(
blob_out, dim_inner, dim_inner, 3, stride, prefix + "_branch2b",
group=group
)
bn_blob = self.conv_bn(
blob_out, dim_inner, dim_out, 1, 1, prefix + "_branch2c", pad=0
)
if self.opts['model_param']['custom_bn_init']:
self.model.param_init_net.ConstantFill(
[bn_blob + '_s'], bn_blob + '_s',
value=self.opts['model_param']['bn_init_gamma'])
sc_blob = self.add_shortcut(
blob_in, dim_in, dim_out, stride, prefix=prefix + "_branch1"
)
sum_blob = self.model.net.Sum([bn_blob, sc_blob], prefix + "_sum")
return self.model.Relu(sum_blob, sum_blob)
# bottleneck residual layer for 50, 101, 152 layer networks
def bottleneck_block(
self, blob_in, dim_in, dim_out, stride, prefix, dim_inner, group=None
):
blob_out = self.conv_bn_relu(
blob_in, dim_in, dim_inner, 1, 1, prefix + "_branch2a", pad=0,
)
blob_out = self.conv_bn_relu(
blob_out, dim_inner, dim_inner, 3, stride, prefix + "_branch2b",
)
bn_blob = self.conv_bn(
blob_out, dim_inner, dim_out, 1, 1, prefix + "_branch2c", pad=0
)
if self.opts['model_param']['custom_bn_init']:
self.model.param_init_net.ConstantFill(
[bn_blob + '_s'], bn_blob + '_s',
value=self.opts['model_param']['bn_init_gamma'])
sc_blob = self.add_shortcut(
blob_in, dim_in, dim_out, stride, prefix=prefix + "_branch1"
)
sum_blob = self.model.net.Sum([bn_blob, sc_blob], prefix + "_sum")
return self.model.Relu(sum_blob, sum_blob)
# basic layer for the 18 and 34 layer networks and the CIFAR data netwrorks
def basic_block(
self, blob_in, dim_in, dim_out, stride, prefix, dim_inner=None,
group=None,
):
blob_out = self.conv_bn_relu(
blob_in, dim_in, dim_out, 3, stride, prefix + "_branch2a"
)
bn_blob = self.conv_bn(
blob_out, dim_out, dim_out, 3, 1, prefix + "_branch2b", pad=1
)
sc_blob = self.add_shortcut(
blob_in, dim_in, dim_out, stride, prefix=prefix + "_branch1"
)
sum_blob = self.model.net.Sum([bn_blob, sc_blob], prefix + "_sum")
return self.model.Relu(sum_blob, sum_blob)
def residual_layer(
self, block_fn, blob_in, dim_in, dim_out, stride, num_blocks, prefix,
dim_inner=None, group=None
):
# prefix is something like: res2, res3, etc.
# each res layer has num_blocks stacked
for idx in range(num_blocks):
block_prefix = "{}_{}".format(prefix, idx)
block_stride = 2 if (idx == 0 and stride == 2) else 1
blob_in = block_fn(
blob_in, dim_in, dim_out, block_stride, block_prefix, dim_inner,
group
)
dim_in = dim_out
return blob_in, dim_in
|
