-
Notifications
You must be signed in to change notification settings - Fork 1
/
model.py
299 lines (246 loc) · 10.9 KB
/
model.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
import tensorflow as tf
import tensorflow.keras.backend as K
from tensorflow.keras.layers import Layer
from tensorflow.keras.layers import Conv2D
from tensorflow.keras.layers import MaxPooling2D
from tensorflow.keras.layers import BatchNormalization
from tensorflow.keras.layers.experimental import SyncBatchNormalization
from tensorflow.keras.regularizers import l2
from tensorflow.keras.losses import cosine_similarity
from tensorflow.keras.models import Model
DEFAULT_ARGS = {
"use_bias": False,
"kernel_regularizer": l2(1e-5)}
BatchNorm_DICT = {
"bn": BatchNormalization,
"syncbn": SyncBatchNormalization}
class ResBlock(Model):
def __init__(self, filters, kernel_size=3, stride=1, conv_shortcut=True, bn='bn', name=None, **kwargs):
super(ResBlock, self).__init__(name=name, **kwargs)
self.filters = filters
self.kernel_size = kernel_size
self.stride = stride
self.conv_shortcut = conv_shortcut
self.bn = bn
def build(self, input_shape):
self.conv1 = Conv2D(self.filters, 1, strides=self.stride, name=self.name+'_1_conv', **DEFAULT_ARGS)
self.bn1 = BatchNorm_DICT[self.bn](axis=-1, epsilon=1.001e-5, name=self.name+'_1_bn')
self.conv2 = Conv2D(self.filters, self.kernel_size, padding='SAME', name=self.name+'_2_conv', **DEFAULT_ARGS)
self.bn2 = BatchNorm_DICT[self.bn](axis=-1, epsilon=1.001e-5, name=self.name+'_2_bn')
self.conv3 = Conv2D(4 * self.filters, 1, name=self.name+'_3_conv', **DEFAULT_ARGS)
self.bn3 = BatchNorm_DICT[self.bn](axis=-1, epsilon=1.001e-5, name=self.name+'_3_bn')
if self.conv_shortcut:
self.shortcut_conv = Conv2D(4 * self.filters, 1, strides=self.stride, name=self.name+'_0_conv', **DEFAULT_ARGS)
self.shortcut_bn = BatchNorm_DICT[self.bn](axis=-1, epsilon=1.001e-5, name=self.name+'_0_bn')
def call(self, inputs, training=None):
if self.conv_shortcut:
shortcut = self.shortcut_conv(inputs)
shortcut = self.shortcut_bn(shortcut, training)
else:
shortcut = inputs
x = tf.nn.relu(self.bn1(self.conv1(inputs), training))
x = tf.nn.relu(self.bn2(self.conv2(x), training))
x = self.bn3(self.conv3(x), training)
x = shortcut + x
x = tf.nn.relu(x)
return x
def get_config(self):
config = super(ResBlock, self).get_config()
config.update({
"filters": self.filters,
"kernel_size": self.kernel_size,
"stride": self.stride,
"conv_shortcut": self.conv_shortcut,
"bn": self.bn})
return config
class ResStack(Model):
def __init__(self, filters, blocks, stride1=2, bn='bn', name=None, **kwargs):
super(ResStack, self).__init__(name=name, **kwargs)
self.filters = filters
self.blocks = blocks
self.stride1 = stride1
self.bn = bn
def build(self, input_shape):
self.stacks = [ResBlock(self.filters, stride=self.stride1, bn=self.bn, name=self.name+'_block1')]
for i in range(2, self.blocks+1):
self.stacks.append(ResBlock(self.filters, conv_shortcut=False, bn=self.bn, name=self.name+'_block'+str(i)))
def call(self, inputs, training=None):
x = inputs
for stack in self.stacks:
x = stack(x, training)
return x
def get_config(self):
config = super(ResStack, self).get_config()
config.update({
"filters": self.filters,
"blocks": self.blocks,
"stride1": self.stride1,
"bn": self.bn})
return config
class ResNet50(Model):
def __init__(self, bn='bn', **kwargs):
super(ResNet50, self).__init__(**kwargs)
self.bn = bn
def build(self, input_shape):
self.conv1 = Conv2D(64, 7, strides=2, padding='SAME', name='conv1_conv', **DEFAULT_ARGS)
self.bn1 = BatchNorm_DICT[self.bn](axis=-1, epsilon=1.001e-5, name='conv1_bn')
self.pool1 = MaxPooling2D(3, strides=2, padding='SAME', name='poo1_pool')
self.resblock1 = ResStack(64, 3, stride1=1, bn=self.bn, name='conv2')
self.resblock2 = ResStack(128, 4, bn=self.bn, name='conv3')
self.resblock3 = ResStack(256, 6, bn=self.bn, name='conv4')
self.resblock4 = ResStack(512, 3, bn=self.bn, name='conv5')
def call(self, inputs, training=None):
x = tf.nn.relu(self.bn1(self.conv1(inputs), training))
x = self.pool1(x)
x = self.resblock1(x, training)
x = self.resblock2(x, training)
x = self.resblock3(x, training)
x = self.resblock4(x, training)
return x
def get_config(self):
config = super(ResNet50, self).get_config()
config.update({"bn": self.bn})
return config
class Projection(Model):
def __init__(self, channel, bn='bn', **kwargs):
super(Projection, self).__init__(**kwargs)
self.channel = channel
self.bn = bn
def build(self, input_shape):
self.backbone = ResNet50(name='resnet50')
self.proj_conv1 = Conv2D(input_shape[-1], 1, name='proj_conv1', **DEFAULT_ARGS)
self.proj_bn1 = BatchNorm_DICT[self.bn](axis=-1, epsilon=1.001e-5, name='proj_bn1')
self.proj_conv2 = Conv2D(self.channel, 1, name='proj_conv2', **DEFAULT_ARGS)
def call(self, inputs, training=None):
x = self.backbone(inputs, training)
x = self.proj_conv1(x)
x = self.proj_bn1(x, training)
x = tf.nn.relu(x)
x = self.proj_conv2(x)
return x
def get_config(self):
config = super(Projection, self).get_config()
config.update({"channel": self.channel, "bn": self.bn})
return config
class PPM(Model):
def __init__(self, num_layers, gamma=2., bn='bn', name=None, **kwargs):
super(PPM, self).__init__(name=name, **kwargs)
self.num_layers = num_layers
self.gamma = gamma
self.bn = bn
def build(self, input_shape):
self.transforms = []
if self.num_layers > 0:
self.transforms.append(('conv', Conv2D(input_shape[-1], 1, name='ppm_conv1', **DEFAULT_ARGS)))
if self.num_layers > 1:
self.transforms.append(('bn', BatchNorm_DICT[self.bn](name=f'ppm_bn2')))
self.transforms.append(('conv', Conv2D(input_shape[-1], 1, name='ppm_conv2', **DEFAULT_ARGS)))
def call(self, inputs, training=None):
shape = K.int_shape(inputs)
_, h, w, c = shape
t = inputs
for transform in self.transforms:
if 'bn' == transform[0]:
t = tf.nn.relu(transform[1](t, training))
else:
t = transform[1](t)
t = tf.reshape(t, (-1, h*w, c))
x = tf.reshape(inputs, (-1, h*w, c))
sim = -cosine_similarity(x[:,:,None,:], x[:,None,:,:], axis=-1) # (-1, h*w, h*w)
sim = tf.nn.relu(sim)
sim = tf.pow(sim, self.gamma)
y = tf.matmul(sim, t) # (-1, h*w, c)
y = tf.reshape(y, (-1, h, w, c))
return y
def get_config(self):
config = super(PPM, self).get_config()
config.update({
"num_layers" : self.num_layers,
"gamma" : self.gamma,
"bn": self.bn})
return config
class PixPro(Model):
def __init__(
self,
logger,
norm='bn',
img_size=224,
channel=256,
gamma=2.,
num_layers=1,
snapshot=None,
**kwargs):
super(PixPro, self).__init__(**kwargs)
self.norm = norm
self.channel = channel
self.gamma = gamma
self.num_layers = num_layers
logger.info('Set regular encoder')
self.encoder_regular = Projection(self.channel, bn=self.norm, name='encoder_regular')
logger.info('Set momentum encoder')
self.encoder_momentum = Projection(self.channel, bn=self.norm, name='encoder_momentum')
logger.info('Set propagation encoder')
self.ppm = PPM(self.num_layers, gamma=self.gamma, bn=self.norm, name='ppm')
# build!
self.encoder_regular(tf.zeros((1, img_size, img_size, 3), dtype=tf.float32), training=False)
self.encoder_momentum(tf.zeros((1, img_size, img_size, 3), dtype=tf.float32), training=False)
self.ppm(tf.zeros((1, img_size//(2**5), img_size//(2**5), self.channel), dtype=tf.float32), training=False)
if snapshot:
self.load_weights(snapshot)
logger.info('Load weights at {}'.format(snapshot))
else:
for layer in self.encoder_regular.layers:
self.encoder_momentum.get_layer(name=layer.name).set_weights(layer.get_weights())
self.encoder_momentum.trainable = False
def call(self, inputs, training):
x = self.encoder_regular(inputs, training)
x = self.ppm(x, training)
return x
def compile(
self,
optimizer,
batch_size,
num_workers,
run_eagerly=None):
super(PixPro, self).compile(
optimizer=optimizer, run_eagerly=run_eagerly)
self._batch_size = batch_size
self._num_workers = num_workers
def consistency_loss(self, y, x, mask):
'''
y (b, h, w, c) : feature map of the propagation encoder
x (b, h, w, c) : featrue map of the momentum encoder
mask (b, h*w, h*w) : mask applied to x
'''
_, h, w, c = x.shape
y = tf.reshape(y, (-1, h*w, c))
x = tf.reshape(x, (-1, h*w, c))
cos = -cosine_similarity(y[:,:,None,:], tf.stop_gradient(x)[:,None,:,:], axis=-1) # (-1, h*w, h*w)
cos *= mask # (-1, h*w, h*w)
cos = tf.reduce_sum(cos, axis=(1,2)) # (-1, 1)
mask_cnt = tf.math.count_nonzero(mask, axis=(1,2), dtype=tf.float32) # (-1, 1)
cos = tf.math.divide_no_nan(cos, mask_cnt) # (-1, 1)
cos = tf.reduce_mean(cos) # (,)
return cos
def train_step(self, data):
view1 = data['view1']
view2 = data['view2']
view1_mask = data['view1_mask']
view2_mask = data['view2_mask']
views = tf.concat((view1, view2), axis=0)
x = self.encoder_momentum(views, training=False) # (None, 7, 7, 256)
x_i, x_j = tf.split(x, num_or_size_splits=2, axis=0)
with tf.GradientTape() as tape:
y = self(views, training=True)
y_i, y_j = tf.split(y, num_or_size_splits=2, axis=0)
cos_ij = self.consistency_loss(y_i, x_j, view2_mask)
cos_ji = self.consistency_loss(y_j, x_i, view1_mask)
loss_cos = -cos_ij-cos_ji
loss_decay = sum(self.losses)
loss = loss_cos + loss_decay
total_loss = loss / self._num_workers
trainable_vars = self.trainable_variables
grads = tape.gradient(total_loss, trainable_vars)
self.optimizer.apply_gradients(zip(grads, trainable_vars))
results = {'loss_cos': loss_cos, 'loss_ij': cos_ij, 'loss_ji': cos_ji, 'loss': loss, 'weight_decay': loss_decay}
return results