-
Notifications
You must be signed in to change notification settings - Fork 562
/
squeezenext.py
500 lines (434 loc) · 15 KB
/
squeezenext.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
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
"""
SqueezeNext for ImageNet-1K, implemented in TensorFlow.
Original paper: 'SqueezeNext: Hardware-Aware Neural Network Design,' https://arxiv.org/abs/1803.10615.
"""
__all__ = ['SqueezeNext', 'sqnxt23_w1', 'sqnxt23_w3d2', 'sqnxt23_w2', 'sqnxt23v5_w1', 'sqnxt23v5_w3d2', 'sqnxt23v5_w2']
import os
import tensorflow as tf
from .common import maxpool2d, conv_block, conv1x1_block, conv7x7_block, is_channels_first, flatten
def sqnxt_unit(x,
in_channels,
out_channels,
strides,
training,
data_format,
name="sqnxt_unit"):
"""
SqueezeNext unit.
Parameters:
----------
x : Tensor
Input tensor.
in_channels : int
Number of input channels.
out_channels : int
Number of output channels.
strides : int or tuple/list of 2 int
Strides of the convolution.
training : bool, or a TensorFlow boolean scalar tensor
Whether to return the output in training mode or in inference mode.
data_format : str
The ordering of the dimensions in tensors.
name : str, default 'sqnxt_unit'
Block name.
Returns:
-------
Tensor
Resulted tensor.
"""
if strides == 2:
reduction_den = 1
resize_identity = True
elif in_channels > out_channels:
reduction_den = 4
resize_identity = True
else:
reduction_den = 2
resize_identity = False
if resize_identity:
identity = conv1x1_block(
x=x,
in_channels=in_channels,
out_channels=out_channels,
strides=strides,
use_bias=True,
training=training,
data_format=data_format,
name=name + "/identity_conv")
else:
identity = x
x = conv1x1_block(
x=x,
in_channels=in_channels,
out_channels=(in_channels // reduction_den),
strides=strides,
use_bias=True,
training=training,
data_format=data_format,
name=name + "/conv1")
x = conv1x1_block(
x=x,
in_channels=(in_channels // reduction_den),
out_channels=(in_channels // (2 * reduction_den)),
use_bias=True,
training=training,
data_format=data_format,
name=name + "/conv2")
x = conv_block(
x=x,
in_channels=(in_channels // (2 * reduction_den)),
out_channels=(in_channels // reduction_den),
kernel_size=(1, 3),
strides=1,
padding=(0, 1),
use_bias=True,
training=training,
data_format=data_format,
name=name + "/conv3")
x = conv_block(
x=x,
in_channels=(in_channels // reduction_den),
out_channels=(in_channels // reduction_den),
kernel_size=(3, 1),
strides=1,
padding=(1, 0),
use_bias=True,
training=training,
data_format=data_format,
name=name + "/conv4")
x = conv1x1_block(
x=x,
in_channels=(in_channels // reduction_den),
out_channels=out_channels,
use_bias=True,
training=training,
data_format=data_format,
name=name + "/conv5")
x = x + identity
x = tf.nn.relu(x, name=name + "/final_activ")
return x
def sqnxt_init_block(x,
in_channels,
out_channels,
training,
data_format,
name="sqnxt_init_block"):
"""
ResNet specific initial block.
Parameters:
----------
x : Tensor
Input tensor.
in_channels : int
Number of input channels.
out_channels : int
Number of output channels.
training : bool, or a TensorFlow boolean scalar tensor
Whether to return the output in training mode or in inference mode.
data_format : str
The ordering of the dimensions in tensors.
name : str, default 'sqnxt_init_block'
Block name.
Returns:
-------
Tensor
Resulted tensor.
"""
x = conv7x7_block(
x=x,
in_channels=in_channels,
out_channels=out_channels,
strides=2,
padding=1,
use_bias=True,
training=training,
data_format=data_format,
name=name + "/conv")
x = maxpool2d(
x=x,
pool_size=3,
strides=2,
ceil_mode=True,
data_format=data_format,
name=name + "/pool")
return x
class SqueezeNext(object):
"""
SqueezeNext model from 'SqueezeNext: Hardware-Aware Neural Network Design,' https://arxiv.org/abs/1803.10615.
Parameters:
----------
channels : list of list of int
Number of output channels for each unit.
init_block_channels : int
Number of output channels for the initial unit.
final_block_channels : int
Number of output channels for the final block of the feature extractor.
in_channels : int, default 3
Number of input channels.
in_size : tuple of two ints, default (224, 224)
Spatial size of the expected input image.
classes : int, default 1000
Number of classification classes.
data_format : str, default 'channels_last'
The ordering of the dimensions in tensors.
"""
def __init__(self,
channels,
init_block_channels,
final_block_channels,
in_channels=3,
in_size=(224, 224),
classes=1000,
data_format="channels_last",
**kwargs):
super(SqueezeNext, self).__init__(**kwargs)
assert (data_format in ["channels_last", "channels_first"])
self.channels = channels
self.init_block_channels = init_block_channels
self.final_block_channels = final_block_channels
self.in_channels = in_channels
self.in_size = in_size
self.classes = classes
self.data_format = data_format
def __call__(self,
x,
training=False):
"""
Build a model graph.
Parameters:
----------
x : Tensor
Input tensor.
training : bool, or a TensorFlow boolean scalar tensor, default False
Whether to return the output in training mode or in inference mode.
Returns:
-------
Tensor
Resulted tensor.
"""
in_channels = self.in_channels
x = sqnxt_init_block(
x=x,
in_channels=in_channels,
out_channels=self.init_block_channels,
training=training,
data_format=self.data_format,
name="features/init_block")
in_channels = self.init_block_channels
for i, channels_per_stage in enumerate(self.channels):
for j, out_channels in enumerate(channels_per_stage):
strides = 2 if (j == 0) and (i != 0) else 1
x = sqnxt_unit(
x=x,
in_channels=in_channels,
out_channels=out_channels,
strides=strides,
training=training,
data_format=self.data_format,
name="features/stage{}/unit{}".format(i + 1, j + 1))
in_channels = out_channels
x = conv1x1_block(
x=x,
in_channels=in_channels,
out_channels=self.final_block_channels,
use_bias=True,
training=training,
data_format=self.data_format,
name="features/final_block")
x = tf.keras.layers.AveragePooling2D(
pool_size=7,
strides=1,
data_format=self.data_format,
name="features/final_pool")(x)
# x = tf.layers.flatten(x)
x = flatten(
x=x,
data_format=self.data_format)
x = tf.keras.layers.Dense(
units=self.classes,
name="output")(x)
return x
def get_squeezenext(version,
width_scale,
model_name=None,
pretrained=False,
root=os.path.join("~", ".tensorflow", "models"),
**kwargs):
"""
Create SqueezeNext model with specific parameters.
Parameters:
----------
version : str
Version of SqueezeNet ('23' or '23v5').
width_scale : float
Scale factor for width of layers.
model_name : str or None, default None
Model name for loading pretrained model.
pretrained : bool, default False
Whether to load the pretrained weights for model.
root : str, default '~/.tensorflow/models'
Location for keeping the model parameters.
Returns:
-------
functor
Functor for model graph creation with extra fields.
"""
init_block_channels = 64
final_block_channels = 128
channels_per_layers = [32, 64, 128, 256]
if version == '23':
layers = [6, 6, 8, 1]
elif version == '23v5':
layers = [2, 4, 14, 1]
else:
raise ValueError("Unsupported SqueezeNet version {}".format(version))
channels = [[ci] * li for (ci, li) in zip(channels_per_layers, layers)]
if width_scale != 1:
channels = [[int(cij * width_scale) for cij in ci] for ci in channels]
init_block_channels = int(init_block_channels * width_scale)
final_block_channels = int(final_block_channels * width_scale)
net = SqueezeNext(
channels=channels,
init_block_channels=init_block_channels,
final_block_channels=final_block_channels,
**kwargs)
if pretrained:
if (model_name is None) or (not model_name):
raise ValueError("Parameter `model_name` should be properly initialized for loading pretrained model.")
from .model_store import download_state_dict
net.state_dict, net.file_path = download_state_dict(
model_name=model_name,
local_model_store_dir_path=root)
else:
net.state_dict = None
net.file_path = None
return net
def sqnxt23_w1(**kwargs):
"""
1.0-SqNxt-23 model from 'SqueezeNext: Hardware-Aware Neural Network Design,' https://arxiv.org/abs/1803.10615.
Parameters:
----------
pretrained : bool, default False
Whether to load the pretrained weights for model.
root : str, default '~/.tensorflow/models'
Location for keeping the model parameters.
Returns:
-------
functor
Functor for model graph creation with extra fields.
"""
return get_squeezenext(version="23", width_scale=1.0, model_name="sqnxt23_w1", **kwargs)
def sqnxt23_w3d2(**kwargs):
"""
1.5-SqNxt-23 model from 'SqueezeNext: Hardware-Aware Neural Network Design,' https://arxiv.org/abs/1803.10615.
Parameters:
----------
pretrained : bool, default False
Whether to load the pretrained weights for model.
root : str, default '~/.tensorflow/models'
Location for keeping the model parameters.
Returns:
-------
functor
Functor for model graph creation with extra fields.
"""
return get_squeezenext(version="23", width_scale=1.5, model_name="sqnxt23_w3d2", **kwargs)
def sqnxt23_w2(**kwargs):
"""
2.0-SqNxt-23 model from 'SqueezeNext: Hardware-Aware Neural Network Design,' https://arxiv.org/abs/1803.10615.
Parameters:
----------
pretrained : bool, default False
Whether to load the pretrained weights for model.
root : str, default '~/.tensorflow/models'
Location for keeping the model parameters.
Returns:
-------
functor
Functor for model graph creation with extra fields.
"""
return get_squeezenext(version="23", width_scale=2.0, model_name="sqnxt23_w2", **kwargs)
def sqnxt23v5_w1(**kwargs):
"""
1.0-SqNxt-23v5 model from 'SqueezeNext: Hardware-Aware Neural Network Design,' https://arxiv.org/abs/1803.10615.
Parameters:
----------
pretrained : bool, default False
Whether to load the pretrained weights for model.
root : str, default '~/.tensorflow/models'
Location for keeping the model parameters.
Returns:
-------
functor
Functor for model graph creation with extra fields.
"""
return get_squeezenext(version="23v5", width_scale=1.0, model_name="sqnxt23v5_w1", **kwargs)
def sqnxt23v5_w3d2(**kwargs):
"""
1.5-SqNxt-23v5 model from 'SqueezeNext: Hardware-Aware Neural Network Design,' https://arxiv.org/abs/1803.10615.
Parameters:
----------
pretrained : bool, default False
Whether to load the pretrained weights for model.
root : str, default '~/.tensorflow/models'
Location for keeping the model parameters.
Returns:
-------
functor
Functor for model graph creation with extra fields.
"""
return get_squeezenext(version="23v5", width_scale=1.5, model_name="sqnxt23v5_w3d2", **kwargs)
def sqnxt23v5_w2(**kwargs):
"""
2.0-SqNxt-23v5 model from 'SqueezeNext: Hardware-Aware Neural Network Design,' https://arxiv.org/abs/1803.10615.
Parameters:
----------
pretrained : bool, default False
Whether to load the pretrained weights for model.
root : str, default '~/.tensorflow/models'
Location for keeping the model parameters.
Returns:
-------
functor
Functor for model graph creation with extra fields.
"""
return get_squeezenext(version="23v5", width_scale=2.0, model_name="sqnxt23v5_w2", **kwargs)
def _test():
import numpy as np
data_format = "channels_last"
pretrained = False
models = [
sqnxt23_w1,
sqnxt23_w3d2,
sqnxt23_w2,
sqnxt23v5_w1,
sqnxt23v5_w3d2,
sqnxt23v5_w2,
]
for model in models:
net = model(pretrained=pretrained, data_format=data_format)
x = tf.placeholder(
dtype=tf.float32,
shape=(None, 3, 224, 224) if is_channels_first(data_format) else (None, 224, 224, 3),
name="xx")
y_net = net(x)
weight_count = np.sum([np.prod(v.get_shape().as_list()) for v in tf.trainable_variables()])
print("m={}, {}".format(model.__name__, weight_count))
assert (model != sqnxt23_w1 or weight_count == 724056)
assert (model != sqnxt23_w3d2 or weight_count == 1511824)
assert (model != sqnxt23_w2 or weight_count == 2583752)
assert (model != sqnxt23v5_w1 or weight_count == 921816)
assert (model != sqnxt23v5_w3d2 or weight_count == 1953616)
assert (model != sqnxt23v5_w2 or weight_count == 3366344)
with tf.Session() as sess:
if pretrained:
from .model_store import init_variables_from_state_dict
init_variables_from_state_dict(sess=sess, state_dict=net.state_dict)
else:
sess.run(tf.global_variables_initializer())
x_value = np.zeros((1, 3, 224, 224) if is_channels_first(data_format) else (1, 224, 224, 3), np.float32)
y = sess.run(y_net, feed_dict={x: x_value})
assert (y.shape == (1, 1000))
tf.reset_default_graph()
if __name__ == "__main__":
_test()