CoCalc -- fr_utils.py

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/fr_utils.py
Views: ¹³³⁷⁶
1
#### PART OF THIS CODE IS USING CODE FROM VICTOR SY WANG: https://github.com/iwantooxxoox/Keras-OpenFace/blob/master/utils.py ####
2

3
import tensorflow as tf
4
import numpy as np
5
import os
6
import cv2
7
from numpy import genfromtxt
8
from keras.layers import Conv2D, ZeroPadding2D, Activation, Input, concatenate
9
from keras.models import Model
10
from keras.layers.normalization import BatchNormalization
11
from keras.layers.pooling import MaxPooling2D, AveragePooling2D
12
import h5py
13
import matplotlib.pyplot as plt
14

15

16
_FLOATX = 'float32'
17

18
def variable(value, dtype=_FLOATX, name=None):
19
    v = tf.Variable(np.asarray(value, dtype=dtype), name=name)
20
    _get_session().run(v.initializer)
21
    return v
22

23
def shape(x):
24
    return x.get_shape()
25

26
def square(x):
27
    return tf.square(x)
28

29
def zeros(shape, dtype=_FLOATX, name=None):
30
    return variable(np.zeros(shape), dtype, name)
31

32
def concatenate(tensors, axis=-1):
33
    if axis < 0:
34
        axis = axis % len(tensors[0].get_shape())
35
    return tf.concat(axis, tensors)
36

37
def LRN2D(x):
38
    return tf.nn.lrn(x, alpha=1e-4, beta=0.75)
39

40
def conv2d_bn(x,
41
              layer=None,
42
              cv1_out=None,
43
              cv1_filter=(1, 1),
44
              cv1_strides=(1, 1),
45
              cv2_out=None,
46
              cv2_filter=(3, 3),
47
              cv2_strides=(1, 1),
48
              padding=None):
49
    num = '' if cv2_out == None else '1'
50
    tensor = Conv2D(cv1_out, cv1_filter, strides=cv1_strides, data_format='channels_first', name=layer+'_conv'+num)(x)
51
    tensor = BatchNormalization(axis=1, epsilon=0.00001, name=layer+'_bn'+num)(tensor)
52
    tensor = Activation('relu')(tensor)
53
    if padding == None:
54
        return tensor
55
    tensor = ZeroPadding2D(padding=padding, data_format='channels_first')(tensor)
56
    if cv2_out == None:
57
        return tensor
58
    tensor = Conv2D(cv2_out, cv2_filter, strides=cv2_strides, data_format='channels_first', name=layer+'_conv'+'2')(tensor)
59
    tensor = BatchNormalization(axis=1, epsilon=0.00001, name=layer+'_bn'+'2')(tensor)
60
    tensor = Activation('relu')(tensor)
61
    return tensor
62

63
WEIGHTS = [
64
  'conv1', 'bn1', 'conv2', 'bn2', 'conv3', 'bn3',
65
  'inception_3a_1x1_conv', 'inception_3a_1x1_bn',
66
  'inception_3a_pool_conv', 'inception_3a_pool_bn',
67
  'inception_3a_5x5_conv1', 'inception_3a_5x5_conv2', 'inception_3a_5x5_bn1', 'inception_3a_5x5_bn2',
68
  'inception_3a_3x3_conv1', 'inception_3a_3x3_conv2', 'inception_3a_3x3_bn1', 'inception_3a_3x3_bn2',
69
  'inception_3b_3x3_conv1', 'inception_3b_3x3_conv2', 'inception_3b_3x3_bn1', 'inception_3b_3x3_bn2',
70
  'inception_3b_5x5_conv1', 'inception_3b_5x5_conv2', 'inception_3b_5x5_bn1', 'inception_3b_5x5_bn2',
71
  'inception_3b_pool_conv', 'inception_3b_pool_bn',
72
  'inception_3b_1x1_conv', 'inception_3b_1x1_bn',
73
  'inception_3c_3x3_conv1', 'inception_3c_3x3_conv2', 'inception_3c_3x3_bn1', 'inception_3c_3x3_bn2',
74
  'inception_3c_5x5_conv1', 'inception_3c_5x5_conv2', 'inception_3c_5x5_bn1', 'inception_3c_5x5_bn2',
75
  'inception_4a_3x3_conv1', 'inception_4a_3x3_conv2', 'inception_4a_3x3_bn1', 'inception_4a_3x3_bn2',
76
  'inception_4a_5x5_conv1', 'inception_4a_5x5_conv2', 'inception_4a_5x5_bn1', 'inception_4a_5x5_bn2',
77
  'inception_4a_pool_conv', 'inception_4a_pool_bn',
78
  'inception_4a_1x1_conv', 'inception_4a_1x1_bn',
79
  'inception_4e_3x3_conv1', 'inception_4e_3x3_conv2', 'inception_4e_3x3_bn1', 'inception_4e_3x3_bn2',
80
  'inception_4e_5x5_conv1', 'inception_4e_5x5_conv2', 'inception_4e_5x5_bn1', 'inception_4e_5x5_bn2',
81
  'inception_5a_3x3_conv1', 'inception_5a_3x3_conv2', 'inception_5a_3x3_bn1', 'inception_5a_3x3_bn2',
82
  'inception_5a_pool_conv', 'inception_5a_pool_bn',
83
  'inception_5a_1x1_conv', 'inception_5a_1x1_bn',
84
  'inception_5b_3x3_conv1', 'inception_5b_3x3_conv2', 'inception_5b_3x3_bn1', 'inception_5b_3x3_bn2',
85
  'inception_5b_pool_conv', 'inception_5b_pool_bn',
86
  'inception_5b_1x1_conv', 'inception_5b_1x1_bn',
87
  'dense_layer'
88
]
89

90
conv_shape = {
91
  'conv1': [64, 3, 7, 7],
92
  'conv2': [64, 64, 1, 1],
93
  'conv3': [192, 64, 3, 3],
94
  'inception_3a_1x1_conv': [64, 192, 1, 1],
95
  'inception_3a_pool_conv': [32, 192, 1, 1],
96
  'inception_3a_5x5_conv1': [16, 192, 1, 1],
97
  'inception_3a_5x5_conv2': [32, 16, 5, 5],
98
  'inception_3a_3x3_conv1': [96, 192, 1, 1],
99
  'inception_3a_3x3_conv2': [128, 96, 3, 3],
100
  'inception_3b_3x3_conv1': [96, 256, 1, 1],
101
  'inception_3b_3x3_conv2': [128, 96, 3, 3],
102
  'inception_3b_5x5_conv1': [32, 256, 1, 1],
103
  'inception_3b_5x5_conv2': [64, 32, 5, 5],
104
  'inception_3b_pool_conv': [64, 256, 1, 1],
105
  'inception_3b_1x1_conv': [64, 256, 1, 1],
106
  'inception_3c_3x3_conv1': [128, 320, 1, 1],
107
  'inception_3c_3x3_conv2': [256, 128, 3, 3],
108
  'inception_3c_5x5_conv1': [32, 320, 1, 1],
109
  'inception_3c_5x5_conv2': [64, 32, 5, 5],
110
  'inception_4a_3x3_conv1': [96, 640, 1, 1],
111
  'inception_4a_3x3_conv2': [192, 96, 3, 3],
112
  'inception_4a_5x5_conv1': [32, 640, 1, 1,],
113
  'inception_4a_5x5_conv2': [64, 32, 5, 5],
114
  'inception_4a_pool_conv': [128, 640, 1, 1],
115
  'inception_4a_1x1_conv': [256, 640, 1, 1],
116
  'inception_4e_3x3_conv1': [160, 640, 1, 1],
117
  'inception_4e_3x3_conv2': [256, 160, 3, 3],
118
  'inception_4e_5x5_conv1': [64, 640, 1, 1],
119
  'inception_4e_5x5_conv2': [128, 64, 5, 5],
120
  'inception_5a_3x3_conv1': [96, 1024, 1, 1],
121
  'inception_5a_3x3_conv2': [384, 96, 3, 3],
122
  'inception_5a_pool_conv': [96, 1024, 1, 1],
123
  'inception_5a_1x1_conv': [256, 1024, 1, 1],
124
  'inception_5b_3x3_conv1': [96, 736, 1, 1],
125
  'inception_5b_3x3_conv2': [384, 96, 3, 3],
126
  'inception_5b_pool_conv': [96, 736, 1, 1],
127
  'inception_5b_1x1_conv': [256, 736, 1, 1],
128
}
129

130
def load_weights_from_FaceNet(FRmodel):
131
    # Load weights from csv files (which was exported from Openface torch model)
132
    weights = WEIGHTS
133
    weights_dict = load_weights()
134

135
    # Set layer weights of the model
136
    for name in weights:
137
        if FRmodel.get_layer(name) != None:
138
            FRmodel.get_layer(name).set_weights(weights_dict[name])
139
        elif model.get_layer(name) != None:
140
            model.get_layer(name).set_weights(weights_dict[name])
141

142
def load_weights():
143
    # Set weights path
144
    dirPath = './weights'
145
    fileNames = filter(lambda f: not f.startswith('.'), os.listdir(dirPath))
146
    paths = {}
147
    weights_dict = {}
148

149
    for n in fileNames:
150
        paths[n.replace('.csv', '')] = dirPath + '/' + n
151

152
    for name in WEIGHTS:
153
        if 'conv' in name:
154
            conv_w = genfromtxt(paths[name + '_w'], delimiter=',', dtype=None)
155
            conv_w = np.reshape(conv_w, conv_shape[name])
156
            conv_w = np.transpose(conv_w, (2, 3, 1, 0))
157
            conv_b = genfromtxt(paths[name + '_b'], delimiter=',', dtype=None)
158
            weights_dict[name] = [conv_w, conv_b]     
159
        elif 'bn' in name:
160
            bn_w = genfromtxt(paths[name + '_w'], delimiter=',', dtype=None)
161
            bn_b = genfromtxt(paths[name + '_b'], delimiter=',', dtype=None)
162
            bn_m = genfromtxt(paths[name + '_m'], delimiter=',', dtype=None)
163
            bn_v = genfromtxt(paths[name + '_v'], delimiter=',', dtype=None)
164
            weights_dict[name] = [bn_w, bn_b, bn_m, bn_v]
165
        elif 'dense' in name:
166
            dense_w = genfromtxt(dirPath+'/dense_w.csv', delimiter=',', dtype=None)
167
            dense_w = np.reshape(dense_w, (128, 736))
168
            dense_w = np.transpose(dense_w, (1, 0))
169
            dense_b = genfromtxt(dirPath+'/dense_b.csv', delimiter=',', dtype=None)
170
            weights_dict[name] = [dense_w, dense_b]
171

172
    return weights_dict
173

174

175
def load_dataset():
176
    train_dataset = h5py.File('datasets/train_happy.h5', "r")
177
    train_set_x_orig = np.array(train_dataset["train_set_x"][:]) # your train set features
178
    train_set_y_orig = np.array(train_dataset["train_set_y"][:]) # your train set labels
179

180
    test_dataset = h5py.File('datasets/test_happy.h5', "r")
181
    test_set_x_orig = np.array(test_dataset["test_set_x"][:]) # your test set features
182
    test_set_y_orig = np.array(test_dataset["test_set_y"][:]) # your test set labels
183

184
    classes = np.array(test_dataset["list_classes"][:]) # the list of classes
185
    
186
    train_set_y_orig = train_set_y_orig.reshape((1, train_set_y_orig.shape[0]))
187
    test_set_y_orig = test_set_y_orig.reshape((1, test_set_y_orig.shape[0]))
188
    
189
    return train_set_x_orig, train_set_y_orig, test_set_x_orig, test_set_y_orig, classes
190

191
def img_to_encoding(image_path, model):
192
    img1 = cv2.imread(image_path, 1)
193
    img = img1[...,::-1]
194
    img = np.around(np.transpose(img, (2,0,1))/255.0, decimals=12)
195
    x_train = np.array([img])
196
    embedding = model.predict_on_batch(x_train)
197
    return embedding
198
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.
