-
Notifications
You must be signed in to change notification settings - Fork 2
/
efficientnet.py
494 lines (379 loc) · 12.8 KB
/
efficientnet.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
# ================================================================
# MIT License
# Copyright (c) 2021 edwardyehuang (https://github.com/edwardyehuang)
# ================================================================
import tensorflow as tf
import copy
import math
from distutils.version import LooseVersion
if LooseVersion(tf.version.VERSION) < LooseVersion("2.7.0"):
from tensorflow.python.keras.applications import imagenet_utils
elif LooseVersion(tf.version.VERSION) < LooseVersion("2.13.0"):
from keras.applications import imagenet_utils
else:
from keras.src.applications import imagenet_utils
from iseg.layers.normalizations import normalization
from iseg.utils.keras3_utils import Keras3_Model_Wrapper
DEFAULT_BLOCKS_ARGS = [
{
"kernel_size": 3,
"repeats": 1,
"filters_in": 32,
"filters_out": 16,
"expand_ratio": 1,
"id_skip": True,
"strides": 1,
"se_ratio": 0.25,
},
{
"kernel_size": 3,
"repeats": 2,
"filters_in": 16,
"filters_out": 24,
"expand_ratio": 6,
"id_skip": True,
"strides": 2,
"se_ratio": 0.25,
},
{
"kernel_size": 5,
"repeats": 2,
"filters_in": 24,
"filters_out": 40,
"expand_ratio": 6,
"id_skip": True,
"strides": 2,
"se_ratio": 0.25,
},
{
"kernel_size": 3,
"repeats": 3,
"filters_in": 40,
"filters_out": 80,
"expand_ratio": 6,
"id_skip": True,
"strides": 2,
"se_ratio": 0.25,
},
{
"kernel_size": 5,
"repeats": 3,
"filters_in": 80,
"filters_out": 112,
"expand_ratio": 6,
"id_skip": True,
"strides": 1,
"se_ratio": 0.25,
},
{
"kernel_size": 5,
"repeats": 4,
"filters_in": 112,
"filters_out": 192,
"expand_ratio": 6,
"id_skip": True,
"strides": 2,
"se_ratio": 0.25,
},
{
"kernel_size": 3,
"repeats": 1,
"filters_in": 192,
"filters_out": 320,
"expand_ratio": 6,
"id_skip": True,
"strides": 1,
"se_ratio": 0.25,
},
]
CONV_KERNEL_INITIALIZER = {
"class_name": "VarianceScaling",
"config": {"scale": 2.0, "mode": "fan_out", "distribution": "truncated_normal"},
}
def round_filters(filters, coefficient, divisor=8):
"""Round number of filters based on depth multiplier."""
filters *= coefficient
new_filters = max(divisor, int(filters + divisor / 2) // divisor * divisor)
# Make sure that round down does not go down by more than 10%.
if new_filters < 0.9 * filters:
new_filters += divisor
return int(new_filters)
def round_repeats(repeats, depth_coefficient):
"""Round number of repeats based on depth multiplier."""
return int(math.ceil(depth_coefficient * repeats))
class Block(Keras3_Model_Wrapper):
def __init__(
self,
activation=tf.nn.swish,
drop_rate=0,
filters_in=32,
filters_out=16,
kernel_size=3,
strides=1,
expand_ratio=1,
se_ratio=0,
id_skip=True,
name=None,
**kwargs
):
super(Block, self).__init__(name=name, **kwargs)
self.strides = strides
self.kernel_size = kernel_size
self.activation = activation if activation is not None else tf.nn.swish
self.filters_in = filters_in
self.filters_out = filters_out
self.drop_rate = drop_rate
self.id_skip = id_skip
self.output_endpoint = strides > 1
filters = filters_in * expand_ratio
if expand_ratio != 1:
self.expand_conv = tf.keras.layers.Conv2D(
filters=filters,
kernel_size=1,
padding="same",
use_bias=False,
kernel_initializer=CONV_KERNEL_INITIALIZER,
name=name + "expand_conv",
)
self.expand_conv_bn = normalization(name=name + "expand_bn")
else:
self.expand_conv = None
conv_pad = "valid" if self.strides == 2 else "same"
self.dwconv = tf.keras.layers.DepthwiseConv2D(
kernel_size=kernel_size,
strides=strides,
padding=conv_pad,
use_bias=False,
depthwise_initializer=CONV_KERNEL_INITIALIZER,
name=name + "dwconv",
)
self.dwconv_bn = normalization(name=name + "bn")
if 0 < se_ratio <= 1:
filters_se = max(1, int(filters_in * se_ratio))
self.se_reduce = tf.keras.layers.Conv2D(
filters=filters_se,
kernel_size=1,
padding="same",
kernel_initializer=CONV_KERNEL_INITIALIZER,
name=name + "se_reduce",
)
self.se_expand = tf.keras.layers.Conv2D(
filters=filters,
kernel_size=1,
padding="same",
kernel_initializer=CONV_KERNEL_INITIALIZER,
name=name + "se_expand",
)
else:
self.se_reduce = None
self.project_conv = tf.keras.layers.Conv2D(
filters=filters_out,
kernel_size=1,
padding="same",
use_bias=False,
kernel_initializer=CONV_KERNEL_INITIALIZER,
name=name + "project_conv",
)
self.project_bn = normalization(name=name + "project_bn")
if drop_rate > 0:
self.dropout = tf.keras.layers.Dropout(drop_rate, noise_shape=(None, 1, 1, 1))
def build(self, input_shape):
super().build(input_shape)
def call(self, inputs, training=None):
x = inputs
current_strides = self.dwconv.strides[0]
if self.expand_conv is not None:
x = self.expand_conv(x)
x = self.expand_conv_bn(x, training=training)
x = self.activation(x)
if current_strides == 2:
padding = imagenet_utils.correct_pad(x, self.kernel_size)
x = tf.keras.backend.spatial_2d_padding(x, padding)
x = self.dwconv(x)
x = self.dwconv_bn(x, training=training)
x = self.activation(x)
if self.se_reduce is not None:
se = tf.reduce_mean(x, (1, 2), keepdims=True, name=self.name + "se_squeeze")
se = self.se_reduce(se)
se = self.activation(se)
se = self.se_expand(se)
se = tf.nn.sigmoid(se)
x = tf.multiply(x, se, name=self.name + "se_excite")
x = self.project_conv(x)
x = self.project_bn(x, training=training)
if self.id_skip and current_strides == 1 and self.filters_in == self.filters_out:
if self.drop_rate > 0:
x = self.dropout(x, training=training)
x += tf.cast(inputs, x.dtype)
return x
class EfficientNet(Keras3_Model_Wrapper):
def __init__(
self,
width_confficient,
depth_confficient,
drop_connect_rate=0.2,
depth_divisor=8,
activation=tf.nn.swish,
blocks_args="default",
return_endpoints=False,
name="efficientnet",
**kwargs
):
super(EfficientNet, self).__init__(name=name)
self.return_endpoints = return_endpoints
if blocks_args == "default":
blocks_args = DEFAULT_BLOCKS_ARGS
blocks_args = copy.deepcopy(blocks_args)
self.activation = activation if activation is not None else tf.nn.swish
self.stem_conv = tf.keras.layers.Conv2D(
filters=round_filters(32, width_confficient, depth_divisor),
kernel_size=3,
strides=2,
padding="valid",
use_bias=False,
kernel_initializer=CONV_KERNEL_INITIALIZER,
name="stem_conv",
)
self.steam_conv_bn = normalization(name="stem_bn")
b = 0
self.blocks = []
blocks_num = float(sum(round_repeats(args["repeats"], depth_confficient) for args in blocks_args))
for (i, args) in enumerate(blocks_args):
assert args["repeats"] > 0
args["filters_in"] = round_filters(args["filters_in"], width_confficient, depth_divisor)
args["filters_out"] = round_filters(args["filters_out"], width_confficient, depth_divisor)
for j in range(round_repeats(args.pop("repeats"), depth_confficient)):
if j > 0:
args["strides"] = 1
args["filters_in"] = args["filters_out"]
block = Block(
activation=activation,
drop_rate=drop_connect_rate * b / blocks_num,
name="block{}{}_".format(i + 1, chr(j + 97)),
**args
)
self.blocks.append(block)
b += 1
self.top_conv = tf.keras.layers.Conv2D(
filters=round_filters(1280, width_confficient, depth_divisor),
kernel_size=1,
padding="same",
use_bias=False,
kernel_initializer=CONV_KERNEL_INITIALIZER,
name="top_conv",
)
self.top_bn = normalization(name="top_bn")
def build(self, input_shape):
super().build(input_shape)
def call(self, inputs, training=None, **kwargs):
endpoints = []
x = inputs
x = tf.keras.backend.spatial_2d_padding(x, imagenet_utils.correct_pad(x, 3))
x = self.stem_conv(x)
x = self.steam_conv_bn(x, training=training)
x = self.activation(x)
for block in self.blocks:
if block.output_endpoint:
endpoints += [x]
x = block(x, training=training)
x = self.top_conv(x)
x = self.top_bn(x, training=training)
x = self.activation(x)
endpoints += [x]
if self.return_endpoints:
return endpoints
else:
return x
def EfficientNetB0(return_endpoints=False):
return EfficientNet(
width_confficient=1.0,
depth_confficient=1.0,
default_size=224,
drop_connect_rate=0.2,
return_endpoints=return_endpoints,
name="efficientnetb0",
)
def EfficientNetB1(return_endpoints=False):
return EfficientNet(
width_confficient=1.0,
depth_confficient=1.1,
default_size=240,
drop_connect_rate=0.2,
return_endpoints=return_endpoints,
name="efficientnetb1",
)
def EfficientNetB2(return_endpoints=False):
return EfficientNet(
width_confficient=1.1,
depth_confficient=1.2,
default_size=260,
drop_connect_rate=0.3,
return_endpoints=return_endpoints,
name="efficientnetb2",
)
def EfficientNetB3(return_endpoints=False):
return EfficientNet(
width_confficient=1.2,
depth_confficient=1.4,
default_size=300,
drop_connect_rate=0.3,
return_endpoints=return_endpoints,
name="efficientnetb3",
)
def EfficientNetB4(return_endpoints=False):
return EfficientNet(
width_confficient=1.4,
depth_confficient=1.8,
default_size=380,
drop_connect_rate=0.4,
return_endpoints=return_endpoints,
name="efficientnetb4",
)
def EfficientNetB5(return_endpoints=False):
return EfficientNet(
width_confficient=1.6,
depth_confficient=2.2,
default_size=456,
drop_connect_rate=0.4,
return_endpoints=return_endpoints,
name="efficientnetb5",
)
def EfficientNetB6(return_endpoints=False):
return EfficientNet(
width_confficient=1.8,
depth_confficient=2.6,
default_size=528,
drop_connect_rate=0.5,
return_endpoints=return_endpoints,
name="efficientnetb6",
)
def EfficientNetB7(return_endpoints=False):
return EfficientNet(
width_confficient=2.0,
depth_confficient=3.1,
default_size=600,
drop_connect_rate=0.5,
return_endpoints=return_endpoints,
name="efficientnetb7",
)
def EfficientNetL2(return_endpoints=False):
return EfficientNet(
width_confficient=4.3,
depth_confficient=5.3,
default_size=800,
drop_connect_rate=0.5,
return_endpoints=return_endpoints,
name="efficientnetl2",
)
def build_dilated_efficientnet(efficientnet, output_stride=16):
current_os = 2
current_dilation = 1
for block in efficientnet.blocks:
if current_os >= output_stride:
current_dilation *= block.dwconv.strides[0]
block.dwconv.strides = (1, 1)
block.dwconv.padding = "same"
block.dwconv.dilation_rate = (current_dilation, current_dilation)
else:
current_os *= block.dwconv.strides[0]
return efficientnet