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: amanchadha/coursera-deep-learning-specialization
Path: blob/master/C4 - Convolutional Neural Networks/Week 4/Face Recognition/inception_blocks_v2.py
Views: ⁴⁸⁰⁴
1
import tensorflow as tf
2
import numpy as np
3
import os
4
from numpy import genfromtxt
5
from tensorflow.keras import backend as K
6
from tensorflow.keras.layers import Conv2D, ZeroPadding2D, Activation, Input, concatenate
7
from tensorflow.keras.models import Model
8
from tensorflow.keras.layers import BatchNormalization
9
from tensorflow.keras.layers import MaxPooling2D, AveragePooling2D
10
import fr_utils
11
from tensorflow.keras.layers import Lambda, Flatten, Dense
12

13
K.set_image_data_format('channels_first')
14

15
def inception_block_1a(X):
16
    """
17
    Implementation of an inception block
18
    """
19
    
20
    X_3x3 = Conv2D(96, (1, 1), data_format='channels_first', name ='inception_3a_3x3_conv1')(X)
21
    X_3x3 = BatchNormalization(axis=1, epsilon=0.00001, name = 'inception_3a_3x3_bn1')(X_3x3)
22
    X_3x3 = Activation('relu')(X_3x3)
23
    X_3x3 = ZeroPadding2D(padding=(1, 1), data_format='channels_first')(X_3x3)
24
    X_3x3 = Conv2D(128, (3, 3), data_format='channels_first', name='inception_3a_3x3_conv2')(X_3x3)
25
    X_3x3 = BatchNormalization(axis=1, epsilon=0.00001, name='inception_3a_3x3_bn2')(X_3x3)
26
    X_3x3 = Activation('relu')(X_3x3)
27
    
28
    X_5x5 = Conv2D(16, (1, 1), data_format='channels_first', name='inception_3a_5x5_conv1')(X)
29
    X_5x5 = BatchNormalization(axis=1, epsilon=0.00001, name='inception_3a_5x5_bn1')(X_5x5)
30
    X_5x5 = Activation('relu')(X_5x5)
31
    X_5x5 = ZeroPadding2D(padding=(2, 2), data_format='channels_first')(X_5x5)
32
    X_5x5 = Conv2D(32, (5, 5), data_format='channels_first', name='inception_3a_5x5_conv2')(X_5x5)
33
    X_5x5 = BatchNormalization(axis=1, epsilon=0.00001, name='inception_3a_5x5_bn2')(X_5x5)
34
    X_5x5 = Activation('relu')(X_5x5)
35

36
    X_pool = MaxPooling2D(pool_size=3, strides=2, data_format='channels_first')(X)
37
    X_pool = Conv2D(32, (1, 1), data_format='channels_first', name='inception_3a_pool_conv')(X_pool)
38
    X_pool = BatchNormalization(axis=1, epsilon=0.00001, name='inception_3a_pool_bn')(X_pool)
39
    X_pool = Activation('relu')(X_pool)
40
    X_pool = ZeroPadding2D(padding=((3, 4), (3, 4)), data_format='channels_first')(X_pool)
41

42
    X_1x1 = Conv2D(64, (1, 1), data_format='channels_first', name='inception_3a_1x1_conv')(X)
43
    X_1x1 = BatchNormalization(axis=1, epsilon=0.00001, name='inception_3a_1x1_bn')(X_1x1)
44
    X_1x1 = Activation('relu')(X_1x1)
45
        
46
    # CONCAT
47
    inception = concatenate([X_3x3, X_5x5, X_pool, X_1x1], axis=1)
48

49
    return inception
50

51
def inception_block_1b(X):
52
    X_3x3 = Conv2D(96, (1, 1), data_format='channels_first', name='inception_3b_3x3_conv1')(X)
53
    X_3x3 = BatchNormalization(axis=1, epsilon=0.00001, name='inception_3b_3x3_bn1')(X_3x3)
54
    X_3x3 = Activation('relu')(X_3x3)
55
    X_3x3 = ZeroPadding2D(padding=(1, 1), data_format='channels_first')(X_3x3)
56
    X_3x3 = Conv2D(128, (3, 3), data_format='channels_first', name='inception_3b_3x3_conv2')(X_3x3)
57
    X_3x3 = BatchNormalization(axis=1, epsilon=0.00001, name='inception_3b_3x3_bn2')(X_3x3)
58
    X_3x3 = Activation('relu')(X_3x3)
59

60
    X_5x5 = Conv2D(32, (1, 1), data_format='channels_first', name='inception_3b_5x5_conv1')(X)
61
    X_5x5 = BatchNormalization(axis=1, epsilon=0.00001, name='inception_3b_5x5_bn1')(X_5x5)
62
    X_5x5 = Activation('relu')(X_5x5)
63
    X_5x5 = ZeroPadding2D(padding=(2, 2), data_format='channels_first')(X_5x5)
64
    X_5x5 = Conv2D(64, (5, 5), data_format='channels_first', name='inception_3b_5x5_conv2')(X_5x5)
65
    X_5x5 = BatchNormalization(axis=1, epsilon=0.00001, name='inception_3b_5x5_bn2')(X_5x5)
66
    X_5x5 = Activation('relu')(X_5x5)
67

68
    X_pool = AveragePooling2D(pool_size=(3, 3), strides=(3, 3), data_format='channels_first')(X)
69
    X_pool = Conv2D(64, (1, 1), data_format='channels_first', name='inception_3b_pool_conv')(X_pool)
70
    X_pool = BatchNormalization(axis=1, epsilon=0.00001, name='inception_3b_pool_bn')(X_pool)
71
    X_pool = Activation('relu')(X_pool)
72
    X_pool = ZeroPadding2D(padding=(4, 4), data_format='channels_first')(X_pool)
73

74
    X_1x1 = Conv2D(64, (1, 1), data_format='channels_first', name='inception_3b_1x1_conv')(X)
75
    X_1x1 = BatchNormalization(axis=1, epsilon=0.00001, name='inception_3b_1x1_bn')(X_1x1)
76
    X_1x1 = Activation('relu')(X_1x1)
77

78
    inception = concatenate([X_3x3, X_5x5, X_pool, X_1x1], axis=1)
79

80
    return inception
81

82
def inception_block_1c(X):
83
    X_3x3 = fr_utils.conv2d_bn(X,
84
                           layer='inception_3c_3x3',
85
                           cv1_out=128,
86
                           cv1_filter=(1, 1),
87
                           cv2_out=256,
88
                           cv2_filter=(3, 3),
89
                           cv2_strides=(2, 2),
90
                           padding=(1, 1))
91

92
    X_5x5 = fr_utils.conv2d_bn(X,
93
                           layer='inception_3c_5x5',
94
                           cv1_out=32,
95
                           cv1_filter=(1, 1),
96
                           cv2_out=64,
97
                           cv2_filter=(5, 5),
98
                           cv2_strides=(2, 2),
99
                           padding=(2, 2))
100

101
    X_pool = MaxPooling2D(pool_size=3, strides=2, data_format='channels_first')(X)
102
    X_pool = ZeroPadding2D(padding=((0, 1), (0, 1)), data_format='channels_first')(X_pool)
103

104
    inception = concatenate([X_3x3, X_5x5, X_pool], axis=1)
105

106
    return inception
107

108
def inception_block_2a(X):
109
    X_3x3 = fr_utils.conv2d_bn(X,
110
                           layer='inception_4a_3x3',
111
                           cv1_out=96,
112
                           cv1_filter=(1, 1),
113
                           cv2_out=192,
114
                           cv2_filter=(3, 3),
115
                           cv2_strides=(1, 1),
116
                           padding=(1, 1))
117
    X_5x5 = fr_utils.conv2d_bn(X,
118
                           layer='inception_4a_5x5',
119
                           cv1_out=32,
120
                           cv1_filter=(1, 1),
121
                           cv2_out=64,
122
                           cv2_filter=(5, 5),
123
                           cv2_strides=(1, 1),
124
                           padding=(2, 2))
125

126
    X_pool = AveragePooling2D(pool_size=(3, 3), strides=(3, 3), data_format='channels_first')(X)
127
    X_pool = fr_utils.conv2d_bn(X_pool,
128
                           layer='inception_4a_pool',
129
                           cv1_out=128,
130
                           cv1_filter=(1, 1),
131
                           padding=(2, 2))
132
    X_1x1 = fr_utils.conv2d_bn(X,
133
                           layer='inception_4a_1x1',
134
                           cv1_out=256,
135
                           cv1_filter=(1, 1))
136
    inception = concatenate([X_3x3, X_5x5, X_pool, X_1x1], axis=1)
137

138
    return inception
139

140
def inception_block_2b(X):
141
    #inception4e
142
    X_3x3 = fr_utils.conv2d_bn(X,
143
                           layer='inception_4e_3x3',
144
                           cv1_out=160,
145
                           cv1_filter=(1, 1),
146
                           cv2_out=256,
147
                           cv2_filter=(3, 3),
148
                           cv2_strides=(2, 2),
149
                           padding=(1, 1))
150
    X_5x5 = fr_utils.conv2d_bn(X,
151
                           layer='inception_4e_5x5',
152
                           cv1_out=64,
153
                           cv1_filter=(1, 1),
154
                           cv2_out=128,
155
                           cv2_filter=(5, 5),
156
                           cv2_strides=(2, 2),
157
                           padding=(2, 2))
158
    
159
    X_pool = MaxPooling2D(pool_size=3, strides=2, data_format='channels_first')(X)
160
    X_pool = ZeroPadding2D(padding=((0, 1), (0, 1)), data_format='channels_first')(X_pool)
161

162
    inception = concatenate([X_3x3, X_5x5, X_pool], axis=1)
163

164
    return inception
165

166
def inception_block_3a(X):
167
    X_3x3 = fr_utils.conv2d_bn(X,
168
                           layer='inception_5a_3x3',
169
                           cv1_out=96,
170
                           cv1_filter=(1, 1),
171
                           cv2_out=384,
172
                           cv2_filter=(3, 3),
173
                           cv2_strides=(1, 1),
174
                           padding=(1, 1))
175
    X_pool = AveragePooling2D(pool_size=(3, 3), strides=(3, 3), data_format='channels_first')(X)
176
    X_pool = fr_utils.conv2d_bn(X_pool,
177
                           layer='inception_5a_pool',
178
                           cv1_out=96,
179
                           cv1_filter=(1, 1),
180
                           padding=(1, 1))
181
    X_1x1 = fr_utils.conv2d_bn(X,
182
                           layer='inception_5a_1x1',
183
                           cv1_out=256,
184
                           cv1_filter=(1, 1))
185

186
    inception = concatenate([X_3x3, X_pool, X_1x1], axis=1)
187

188
    return inception
189

190
def inception_block_3b(X):
191
    X_3x3 = fr_utils.conv2d_bn(X,
192
                           layer='inception_5b_3x3',
193
                           cv1_out=96,
194
                           cv1_filter=(1, 1),
195
                           cv2_out=384,
196
                           cv2_filter=(3, 3),
197
                           cv2_strides=(1, 1),
198
                           padding=(1, 1))
199
    X_pool = MaxPooling2D(pool_size=3, strides=2, data_format='channels_first')(X)
200
    X_pool = fr_utils.conv2d_bn(X_pool,
201
                           layer='inception_5b_pool',
202
                           cv1_out=96,
203
                           cv1_filter=(1, 1))
204
    X_pool = ZeroPadding2D(padding=(1, 1), data_format='channels_first')(X_pool)
205

206
    X_1x1 = fr_utils.conv2d_bn(X,
207
                           layer='inception_5b_1x1',
208
                           cv1_out=256,
209
                           cv1_filter=(1, 1))
210
    inception = concatenate([X_3x3, X_pool, X_1x1], axis=1)
211

212
    return inception
213

214
def faceRecoModel(input_shape):
215
    """
216
    Implementation of the Inception model used for FaceNet
217
    
218
    Arguments:
219
    input_shape -- shape of the images of the dataset
220

221
    Returns:
222
    model -- a Model() instance in Keras
223
    """
224
        
225
    # Define the input as a tensor with shape input_shape
226
    X_input = Input(input_shape)
227

228
    # Zero-Padding
229
    X = ZeroPadding2D((3, 3))(X_input)
230
    
231
    # First Block
232
    X = Conv2D(64, (7, 7), strides = (2, 2), name = 'conv1', data_format='channels_first')(X)
233
    X = BatchNormalization(axis = 1, name = 'bn1')(X)
234
    X = Activation('relu')(X)
235
    
236
    # Zero-Padding + MAXPOOL
237
    X = ZeroPadding2D((1, 1),data_format='channels_first' )(X)
238
    X = MaxPooling2D((3, 3), strides = 2, data_format='channels_first')(X)
239
    
240
    # Second Block
241
    X = Conv2D(64, (1, 1), strides = (1, 1), name = 'conv2', data_format='channels_first')(X)
242
    X = BatchNormalization(axis = 1, epsilon=0.00001, name = 'bn2')(X)
243
    X = Activation('relu')(X)
244
    
245
    # Zero-Padding + MAXPOOL
246
    X = ZeroPadding2D((1, 1), data_format='channels_first')(X)
247

248
    # Second Block
249
    X = Conv2D(192, (3, 3), strides = (1, 1), name = 'conv3', data_format='channels_first')(X)
250
    X = BatchNormalization(axis = 1, epsilon=0.00001, name = 'bn3')(X)
251
    X = Activation('relu')(X)
252
    
253
    # Zero-Padding + MAXPOOL
254
    X = ZeroPadding2D((1, 1), data_format='channels_first')(X)
255
    X = MaxPooling2D(pool_size = 3, strides = 2,data_format='channels_first')(X)
256
    
257
    # Inception 1: a/b/c
258
    X = inception_block_1a(X)
259
    X = inception_block_1b(X)
260
    X = inception_block_1c(X)
261
    
262
    # Inception 2: a/b
263
    X = inception_block_2a(X)
264
    X = inception_block_2b(X)
265
    
266
    # Inception 3: a/b
267
    X = inception_block_3a(X)
268
    X = inception_block_3b(X)
269
    
270
    # Top layer
271
    X = AveragePooling2D(pool_size=(3, 3), strides=(1, 1), data_format='channels_first')(X)
272
    X = Flatten()(X)
273
    X = Dense(128, name='dense_layer')(X)
274
    
275
    # L2 normalization
276
    X = Lambda(lambda  x: K.l2_normalize(x,axis=1))(X)
277

278
    # Create model instance
279
    model = Model(inputs = X_input, outputs = X, name='FaceRecoModel')
280
        
281
    return model
282
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.
