CoCalc -- inception_blocks

Real-time collaboration for Jupyter Notebooks, Linux Terminals, LaTeX, VS Code, R IDE, and more,
all in one place. Commercial Alternative to JupyterHub.

GitHub Repository: y33-j3T/Coursera-Deep-Learning
Path: blob/master/Convolutional Neural Networks/week4/Face Recognition/inception_blocks_v2.py
Views: ¹³³⁷⁷
1
import tensorflow as tf
2
import numpy as np
3
import os
4
from numpy import genfromtxt
5
from keras import backend as K
6
from keras.layers import Conv2D, ZeroPadding2D, Activation, Input, concatenate
7
from keras.models import Model
8
from keras.layers.normalization import BatchNormalization
9
from keras.layers.pooling import MaxPooling2D, AveragePooling2D
10
import fr_utils
11
from keras.layers.core import Lambda, Flatten, Dense
12

13
def inception_block_1a(X):
14
    """
15
    Implementation of an inception block
16
    """
17
    
18
    X_3x3 = Conv2D(96, (1, 1), data_format='channels_first', name ='inception_3a_3x3_conv1')(X)
19
    X_3x3 = BatchNormalization(axis=1, epsilon=0.00001, name = 'inception_3a_3x3_bn1')(X_3x3)
20
    X_3x3 = Activation('relu')(X_3x3)
21
    X_3x3 = ZeroPadding2D(padding=(1, 1), data_format='channels_first')(X_3x3)
22
    X_3x3 = Conv2D(128, (3, 3), data_format='channels_first', name='inception_3a_3x3_conv2')(X_3x3)
23
    X_3x3 = BatchNormalization(axis=1, epsilon=0.00001, name='inception_3a_3x3_bn2')(X_3x3)
24
    X_3x3 = Activation('relu')(X_3x3)
25
    
26
    X_5x5 = Conv2D(16, (1, 1), data_format='channels_first', name='inception_3a_5x5_conv1')(X)
27
    X_5x5 = BatchNormalization(axis=1, epsilon=0.00001, name='inception_3a_5x5_bn1')(X_5x5)
28
    X_5x5 = Activation('relu')(X_5x5)
29
    X_5x5 = ZeroPadding2D(padding=(2, 2), data_format='channels_first')(X_5x5)
30
    X_5x5 = Conv2D(32, (5, 5), data_format='channels_first', name='inception_3a_5x5_conv2')(X_5x5)
31
    X_5x5 = BatchNormalization(axis=1, epsilon=0.00001, name='inception_3a_5x5_bn2')(X_5x5)
32
    X_5x5 = Activation('relu')(X_5x5)
33

34
    X_pool = MaxPooling2D(pool_size=3, strides=2, data_format='channels_first')(X)
35
    X_pool = Conv2D(32, (1, 1), data_format='channels_first', name='inception_3a_pool_conv')(X_pool)
36
    X_pool = BatchNormalization(axis=1, epsilon=0.00001, name='inception_3a_pool_bn')(X_pool)
37
    X_pool = Activation('relu')(X_pool)
38
    X_pool = ZeroPadding2D(padding=((3, 4), (3, 4)), data_format='channels_first')(X_pool)
39

40
    X_1x1 = Conv2D(64, (1, 1), data_format='channels_first', name='inception_3a_1x1_conv')(X)
41
    X_1x1 = BatchNormalization(axis=1, epsilon=0.00001, name='inception_3a_1x1_bn')(X_1x1)
42
    X_1x1 = Activation('relu')(X_1x1)
43
        
44
    # CONCAT
45
    inception = concatenate([X_3x3, X_5x5, X_pool, X_1x1], axis=1)
46

47
    return inception
48

49
def inception_block_1b(X):
50
    X_3x3 = Conv2D(96, (1, 1), data_format='channels_first', name='inception_3b_3x3_conv1')(X)
51
    X_3x3 = BatchNormalization(axis=1, epsilon=0.00001, name='inception_3b_3x3_bn1')(X_3x3)
52
    X_3x3 = Activation('relu')(X_3x3)
53
    X_3x3 = ZeroPadding2D(padding=(1, 1), data_format='channels_first')(X_3x3)
54
    X_3x3 = Conv2D(128, (3, 3), data_format='channels_first', name='inception_3b_3x3_conv2')(X_3x3)
55
    X_3x3 = BatchNormalization(axis=1, epsilon=0.00001, name='inception_3b_3x3_bn2')(X_3x3)
56
    X_3x3 = Activation('relu')(X_3x3)
57

58
    X_5x5 = Conv2D(32, (1, 1), data_format='channels_first', name='inception_3b_5x5_conv1')(X)
59
    X_5x5 = BatchNormalization(axis=1, epsilon=0.00001, name='inception_3b_5x5_bn1')(X_5x5)
60
    X_5x5 = Activation('relu')(X_5x5)
61
    X_5x5 = ZeroPadding2D(padding=(2, 2), data_format='channels_first')(X_5x5)
62
    X_5x5 = Conv2D(64, (5, 5), data_format='channels_first', name='inception_3b_5x5_conv2')(X_5x5)
63
    X_5x5 = BatchNormalization(axis=1, epsilon=0.00001, name='inception_3b_5x5_bn2')(X_5x5)
64
    X_5x5 = Activation('relu')(X_5x5)
65

66
    X_pool = AveragePooling2D(pool_size=(3, 3), strides=(3, 3), data_format='channels_first')(X)
67
    X_pool = Conv2D(64, (1, 1), data_format='channels_first', name='inception_3b_pool_conv')(X_pool)
68
    X_pool = BatchNormalization(axis=1, epsilon=0.00001, name='inception_3b_pool_bn')(X_pool)
69
    X_pool = Activation('relu')(X_pool)
70
    X_pool = ZeroPadding2D(padding=(4, 4), data_format='channels_first')(X_pool)
71

72
    X_1x1 = Conv2D(64, (1, 1), data_format='channels_first', name='inception_3b_1x1_conv')(X)
73
    X_1x1 = BatchNormalization(axis=1, epsilon=0.00001, name='inception_3b_1x1_bn')(X_1x1)
74
    X_1x1 = Activation('relu')(X_1x1)
75

76
    inception = concatenate([X_3x3, X_5x5, X_pool, X_1x1], axis=1)
77

78
    return inception
79

80
def inception_block_1c(X):
81
    X_3x3 = fr_utils.conv2d_bn(X,
82
                           layer='inception_3c_3x3',
83
                           cv1_out=128,
84
                           cv1_filter=(1, 1),
85
                           cv2_out=256,
86
                           cv2_filter=(3, 3),
87
                           cv2_strides=(2, 2),
88
                           padding=(1, 1))
89

90
    X_5x5 = fr_utils.conv2d_bn(X,
91
                           layer='inception_3c_5x5',
92
                           cv1_out=32,
93
                           cv1_filter=(1, 1),
94
                           cv2_out=64,
95
                           cv2_filter=(5, 5),
96
                           cv2_strides=(2, 2),
97
                           padding=(2, 2))
98

99
    X_pool = MaxPooling2D(pool_size=3, strides=2, data_format='channels_first')(X)
100
    X_pool = ZeroPadding2D(padding=((0, 1), (0, 1)), data_format='channels_first')(X_pool)
101

102
    inception = concatenate([X_3x3, X_5x5, X_pool], axis=1)
103

104
    return inception
105

106
def inception_block_2a(X):
107
    X_3x3 = fr_utils.conv2d_bn(X,
108
                           layer='inception_4a_3x3',
109
                           cv1_out=96,
110
                           cv1_filter=(1, 1),
111
                           cv2_out=192,
112
                           cv2_filter=(3, 3),
113
                           cv2_strides=(1, 1),
114
                           padding=(1, 1))
115
    X_5x5 = fr_utils.conv2d_bn(X,
116
                           layer='inception_4a_5x5',
117
                           cv1_out=32,
118
                           cv1_filter=(1, 1),
119
                           cv2_out=64,
120
                           cv2_filter=(5, 5),
121
                           cv2_strides=(1, 1),
122
                           padding=(2, 2))
123

124
    X_pool = AveragePooling2D(pool_size=(3, 3), strides=(3, 3), data_format='channels_first')(X)
125
    X_pool = fr_utils.conv2d_bn(X_pool,
126
                           layer='inception_4a_pool',
127
                           cv1_out=128,
128
                           cv1_filter=(1, 1),
129
                           padding=(2, 2))
130
    X_1x1 = fr_utils.conv2d_bn(X,
131
                           layer='inception_4a_1x1',
132
                           cv1_out=256,
133
                           cv1_filter=(1, 1))
134
    inception = concatenate([X_3x3, X_5x5, X_pool, X_1x1], axis=1)
135

136
    return inception
137

138
def inception_block_2b(X):
139
    #inception4e
140
    X_3x3 = fr_utils.conv2d_bn(X,
141
                           layer='inception_4e_3x3',
142
                           cv1_out=160,
143
                           cv1_filter=(1, 1),
144
                           cv2_out=256,
145
                           cv2_filter=(3, 3),
146
                           cv2_strides=(2, 2),
147
                           padding=(1, 1))
148
    X_5x5 = fr_utils.conv2d_bn(X,
149
                           layer='inception_4e_5x5',
150
                           cv1_out=64,
151
                           cv1_filter=(1, 1),
152
                           cv2_out=128,
153
                           cv2_filter=(5, 5),
154
                           cv2_strides=(2, 2),
155
                           padding=(2, 2))
156
    
157
    X_pool = MaxPooling2D(pool_size=3, strides=2, data_format='channels_first')(X)
158
    X_pool = ZeroPadding2D(padding=((0, 1), (0, 1)), data_format='channels_first')(X_pool)
159

160
    inception = concatenate([X_3x3, X_5x5, X_pool], axis=1)
161

162
    return inception
163

164
def inception_block_3a(X):
165
    X_3x3 = fr_utils.conv2d_bn(X,
166
                           layer='inception_5a_3x3',
167
                           cv1_out=96,
168
                           cv1_filter=(1, 1),
169
                           cv2_out=384,
170
                           cv2_filter=(3, 3),
171
                           cv2_strides=(1, 1),
172
                           padding=(1, 1))
173
    X_pool = AveragePooling2D(pool_size=(3, 3), strides=(3, 3), data_format='channels_first')(X)
174
    X_pool = fr_utils.conv2d_bn(X_pool,
175
                           layer='inception_5a_pool',
176
                           cv1_out=96,
177
                           cv1_filter=(1, 1),
178
                           padding=(1, 1))
179
    X_1x1 = fr_utils.conv2d_bn(X,
180
                           layer='inception_5a_1x1',
181
                           cv1_out=256,
182
                           cv1_filter=(1, 1))
183

184
    inception = concatenate([X_3x3, X_pool, X_1x1], axis=1)
185

186
    return inception
187

188
def inception_block_3b(X):
189
    X_3x3 = fr_utils.conv2d_bn(X,
190
                           layer='inception_5b_3x3',
191
                           cv1_out=96,
192
                           cv1_filter=(1, 1),
193
                           cv2_out=384,
194
                           cv2_filter=(3, 3),
195
                           cv2_strides=(1, 1),
196
                           padding=(1, 1))
197
    X_pool = MaxPooling2D(pool_size=3, strides=2, data_format='channels_first')(X)
198
    X_pool = fr_utils.conv2d_bn(X_pool,
199
                           layer='inception_5b_pool',
200
                           cv1_out=96,
201
                           cv1_filter=(1, 1))
202
    X_pool = ZeroPadding2D(padding=(1, 1), data_format='channels_first')(X_pool)
203

204
    X_1x1 = fr_utils.conv2d_bn(X,
205
                           layer='inception_5b_1x1',
206
                           cv1_out=256,
207
                           cv1_filter=(1, 1))
208
    inception = concatenate([X_3x3, X_pool, X_1x1], axis=1)
209

210
    return inception
211

212
def faceRecoModel(input_shape):
213
    """
214
    Implementation of the Inception model used for FaceNet
215
    
216
    Arguments:
217
    input_shape -- shape of the images of the dataset
218

219
    Returns:
220
    model -- a Model() instance in Keras
221
    """
222
        
223
    # Define the input as a tensor with shape input_shape
224
    X_input = Input(input_shape)
225

226
    # Zero-Padding
227
    X = ZeroPadding2D((3, 3))(X_input)
228
    
229
    # First Block
230
    X = Conv2D(64, (7, 7), strides = (2, 2), name = 'conv1')(X)
231
    X = BatchNormalization(axis = 1, name = 'bn1')(X)
232
    X = Activation('relu')(X)
233
    
234
    # Zero-Padding + MAXPOOL
235
    X = ZeroPadding2D((1, 1))(X)
236
    X = MaxPooling2D((3, 3), strides = 2)(X)
237
    
238
    # Second Block
239
    X = Conv2D(64, (1, 1), strides = (1, 1), name = 'conv2')(X)
240
    X = BatchNormalization(axis = 1, epsilon=0.00001, name = 'bn2')(X)
241
    X = Activation('relu')(X)
242
    
243
    # Zero-Padding + MAXPOOL
244
    X = ZeroPadding2D((1, 1))(X)
245

246
    # Second Block
247
    X = Conv2D(192, (3, 3), strides = (1, 1), name = 'conv3')(X)
248
    X = BatchNormalization(axis = 1, epsilon=0.00001, name = 'bn3')(X)
249
    X = Activation('relu')(X)
250
    
251
    # Zero-Padding + MAXPOOL
252
    X = ZeroPadding2D((1, 1))(X)
253
    X = MaxPooling2D(pool_size = 3, strides = 2)(X)
254
    
255
    # Inception 1: a/b/c
256
    X = inception_block_1a(X)
257
    X = inception_block_1b(X)
258
    X = inception_block_1c(X)
259
    
260
    # Inception 2: a/b
261
    X = inception_block_2a(X)
262
    X = inception_block_2b(X)
263
    
264
    # Inception 3: a/b
265
    X = inception_block_3a(X)
266
    X = inception_block_3b(X)
267
    
268
    # Top layer
269
    X = AveragePooling2D(pool_size=(3, 3), strides=(1, 1), data_format='channels_first')(X)
270
    X = Flatten()(X)
271
    X = Dense(128, name='dense_layer')(X)
272
    
273
    # L2 normalization
274
    X = Lambda(lambda  x: K.l2_normalize(x,axis=1))(X)
275

276
    # Create model instance
277
    model = Model(inputs = X_input, outputs = X, name='FaceRecoModel')
278
        
279
    return model
280
Real-time collaboration for Jupyter Notebooks, Linux Terminals, LaTeX, VS Code, R IDE, and more,
all in one place. Commercial Alternative to JupyterHub.

Product

Resources

Company

Real-time collaboration for Jupyter Notebooks, Linux Terminals, LaTeX, VS Code, R IDE, and more, all in one place. Commercial Alternative to JupyterHub.

Real-time collaboration for Jupyter Notebooks, Linux Terminals, LaTeX, VS Code, R IDE, and more,
all in one place. Commercial Alternative to JupyterHub.
