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: amanchadha/coursera-deep-learning-specialization
Path: blob/master/C4 - Convolutional Neural Networks/Week 4/Face Recognition/fr_utils.py
Views: ⁴⁸⁰⁴
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 tensorflow.keras.layers import Conv2D, ZeroPadding2D, Activation, Input, concatenate
9
from tensorflow.keras.models import Model
10
from tensorflow.keras.layers import BatchNormalization
11
from tensorflow.keras.layers import MaxPooling2D, AveragePooling2D
12
import h5py
13
import matplotlib.pyplot as plt
14
import PIL
15
from tensorflow.keras import backend as K
16

17
K.set_image_data_format('channels_first')
18

19

20
_FLOATX = 'float32'
21

22
def variable(value, dtype=_FLOATX, name=None):
23
    v = tf.Variable(np.asarray(value, dtype=dtype), name=name)
24
    _get_session().run(v.initializer)
25
    return v
26

27
def shape(x):
28
    return x.get_shape()
29

30
def square(x):
31
    return tf.square(x)
32

33
def zeros(shape, dtype=_FLOATX, name=None):
34
    return variable(np.zeros(shape), dtype, name)
35

36
def concatenate(tensors, axis=-1):
37
    if axis < 0:
38
        axis = axis % len(tensors[0].get_shape())
39
    return tf.concat(axis, tensors)
40

41
def LRN2D(x):
42
    return tf.nn.lrn(x, alpha=1e-4, beta=0.75)
43

44
def conv2d_bn(x,
45
              layer=None,
46
              cv1_out=None,
47
              cv1_filter=(1, 1),
48
              cv1_strides=(1, 1),
49
              cv2_out=None,
50
              cv2_filter=(3, 3),
51
              cv2_strides=(1, 1),
52
              padding=None):
53
    num = '' if cv2_out == None else '1'
54
    tensor = Conv2D(cv1_out, cv1_filter, strides=cv1_strides, data_format='channels_first', name=layer+'_conv'+num)(x)
55
    tensor = BatchNormalization(axis=1, epsilon=0.00001, name=layer+'_bn'+num)(tensor)
56
    tensor = Activation('relu')(tensor)
57
    if padding == None:
58
        return tensor
59
    tensor = ZeroPadding2D(padding=padding, data_format='channels_first')(tensor)
60
    if cv2_out == None:
61
        return tensor
62
    tensor = Conv2D(cv2_out, cv2_filter, strides=cv2_strides, data_format='channels_first', name=layer+'_conv'+'2')(tensor)
63
    tensor = BatchNormalization(axis=1, epsilon=0.00001, name=layer+'_bn'+'2')(tensor)
64
    tensor = Activation('relu')(tensor)
65
    return tensor
66

67
WEIGHTS = [
68
  'conv1', 'bn1', 'conv2', 'bn2', 'conv3', 'bn3',
69
  'inception_3a_1x1_conv', 'inception_3a_1x1_bn',
70
  'inception_3a_pool_conv', 'inception_3a_pool_bn',
71
  'inception_3a_5x5_conv1', 'inception_3a_5x5_conv2', 'inception_3a_5x5_bn1', 'inception_3a_5x5_bn2',
72
  'inception_3a_3x3_conv1', 'inception_3a_3x3_conv2', 'inception_3a_3x3_bn1', 'inception_3a_3x3_bn2',
73
  'inception_3b_3x3_conv1', 'inception_3b_3x3_conv2', 'inception_3b_3x3_bn1', 'inception_3b_3x3_bn2',
74
  'inception_3b_5x5_conv1', 'inception_3b_5x5_conv2', 'inception_3b_5x5_bn1', 'inception_3b_5x5_bn2',
75
  'inception_3b_pool_conv', 'inception_3b_pool_bn',
76
  'inception_3b_1x1_conv', 'inception_3b_1x1_bn',
77
  'inception_3c_3x3_conv1', 'inception_3c_3x3_conv2', 'inception_3c_3x3_bn1', 'inception_3c_3x3_bn2',
78
  'inception_3c_5x5_conv1', 'inception_3c_5x5_conv2', 'inception_3c_5x5_bn1', 'inception_3c_5x5_bn2',
79
  'inception_4a_3x3_conv1', 'inception_4a_3x3_conv2', 'inception_4a_3x3_bn1', 'inception_4a_3x3_bn2',
80
  'inception_4a_5x5_conv1', 'inception_4a_5x5_conv2', 'inception_4a_5x5_bn1', 'inception_4a_5x5_bn2',
81
  'inception_4a_pool_conv', 'inception_4a_pool_bn',
82
  'inception_4a_1x1_conv', 'inception_4a_1x1_bn',
83
  'inception_4e_3x3_conv1', 'inception_4e_3x3_conv2', 'inception_4e_3x3_bn1', 'inception_4e_3x3_bn2',
84
  'inception_4e_5x5_conv1', 'inception_4e_5x5_conv2', 'inception_4e_5x5_bn1', 'inception_4e_5x5_bn2',
85
  'inception_5a_3x3_conv1', 'inception_5a_3x3_conv2', 'inception_5a_3x3_bn1', 'inception_5a_3x3_bn2',
86
  'inception_5a_pool_conv', 'inception_5a_pool_bn',
87
  'inception_5a_1x1_conv', 'inception_5a_1x1_bn',
88
  'inception_5b_3x3_conv1', 'inception_5b_3x3_conv2', 'inception_5b_3x3_bn1', 'inception_5b_3x3_bn2',
89
  'inception_5b_pool_conv', 'inception_5b_pool_bn',
90
  'inception_5b_1x1_conv', 'inception_5b_1x1_bn',
91
  'dense_layer'
92
]
93

94
conv_shape = {
95
  'conv1': [64, 3, 7, 7],
96
  'conv2': [64, 64, 1, 1],
97
  'conv3': [192, 64, 3, 3],
98
  'inception_3a_1x1_conv': [64, 192, 1, 1],
99
  'inception_3a_pool_conv': [32, 192, 1, 1],
100
  'inception_3a_5x5_conv1': [16, 192, 1, 1],
101
  'inception_3a_5x5_conv2': [32, 16, 5, 5],
102
  'inception_3a_3x3_conv1': [96, 192, 1, 1],
103
  'inception_3a_3x3_conv2': [128, 96, 3, 3],
104
  'inception_3b_3x3_conv1': [96, 256, 1, 1],
105
  'inception_3b_3x3_conv2': [128, 96, 3, 3],
106
  'inception_3b_5x5_conv1': [32, 256, 1, 1],
107
  'inception_3b_5x5_conv2': [64, 32, 5, 5],
108
  'inception_3b_pool_conv': [64, 256, 1, 1],
109
  'inception_3b_1x1_conv': [64, 256, 1, 1],
110
  'inception_3c_3x3_conv1': [128, 320, 1, 1],
111
  'inception_3c_3x3_conv2': [256, 128, 3, 3],
112
  'inception_3c_5x5_conv1': [32, 320, 1, 1],
113
  'inception_3c_5x5_conv2': [64, 32, 5, 5],
114
  'inception_4a_3x3_conv1': [96, 640, 1, 1],
115
  'inception_4a_3x3_conv2': [192, 96, 3, 3],
116
  'inception_4a_5x5_conv1': [32, 640, 1, 1,],
117
  'inception_4a_5x5_conv2': [64, 32, 5, 5],
118
  'inception_4a_pool_conv': [128, 640, 1, 1],
119
  'inception_4a_1x1_conv': [256, 640, 1, 1],
120
  'inception_4e_3x3_conv1': [160, 640, 1, 1],
121
  'inception_4e_3x3_conv2': [256, 160, 3, 3],
122
  'inception_4e_5x5_conv1': [64, 640, 1, 1],
123
  'inception_4e_5x5_conv2': [128, 64, 5, 5],
124
  'inception_5a_3x3_conv1': [96, 1024, 1, 1],
125
  'inception_5a_3x3_conv2': [384, 96, 3, 3],
126
  'inception_5a_pool_conv': [96, 1024, 1, 1],
127
  'inception_5a_1x1_conv': [256, 1024, 1, 1],
128
  'inception_5b_3x3_conv1': [96, 736, 1, 1],
129
  'inception_5b_3x3_conv2': [384, 96, 3, 3],
130
  'inception_5b_pool_conv': [96, 736, 1, 1],
131
  'inception_5b_1x1_conv': [256, 736, 1, 1],
132
}
133

134
def load_weights_from_FaceNet(FRmodel):
135
    # Load weights from csv files (which was exported from Openface torch model)
136
    weights = WEIGHTS
137
    weights_dict = load_weights()
138

139
    # Set layer weights of the model
140
    for name in weights:
141
        if FRmodel.get_layer(name) != None:
142
            FRmodel.get_layer(name).set_weights(weights_dict[name])
143
        elif model.get_layer(name) != None:
144
            model.get_layer(name).set_weights(weights_dict[name])
145

146
def load_weights():
147
    # Set weights path
148
    dirPath = './weights'
149
    fileNames = filter(lambda f: not f.startswith('.'), os.listdir(dirPath))
150
    paths = {}
151
    weights_dict = {}
152

153
    for n in fileNames:
154
        paths[n.replace('.csv', '')] = dirPath + '/' + n
155

156
    for name in WEIGHTS:
157
        if 'conv' in name:
158
            conv_w = genfromtxt(paths[name + '_w'], delimiter=',', dtype=None)
159
            conv_w = np.reshape(conv_w, conv_shape[name])
160
            conv_w = np.transpose(conv_w, (2, 3, 1, 0))
161
            conv_b = genfromtxt(paths[name + '_b'], delimiter=',', dtype=None)
162
            weights_dict[name] = [conv_w, conv_b]     
163
        elif 'bn' in name:
164
            bn_w = genfromtxt(paths[name + '_w'], delimiter=',', dtype=None)
165
            bn_b = genfromtxt(paths[name + '_b'], delimiter=',', dtype=None)
166
            bn_m = genfromtxt(paths[name + '_m'], delimiter=',', dtype=None)
167
            bn_v = genfromtxt(paths[name + '_v'], delimiter=',', dtype=None)
168
            weights_dict[name] = [bn_w, bn_b, bn_m, bn_v]
169
        elif 'dense' in name:
170
            dense_w = genfromtxt(dirPath+'/dense_w.csv', delimiter=',', dtype=None)
171
            dense_w = np.reshape(dense_w, (128, 736))
172
            dense_w = np.transpose(dense_w, (1, 0))
173
            dense_b = genfromtxt(dirPath+'/dense_b.csv', delimiter=',', dtype=None)
174
            weights_dict[name] = [dense_w, dense_b]
175

176
    return weights_dict
177

178

179
def load_dataset():
180
    train_dataset = h5py.File('datasets/train_happy.h5', "r")
181
    train_set_x_orig = np.array(train_dataset["train_set_x"][:]) # your train set features
182
    train_set_y_orig = np.array(train_dataset["train_set_y"][:]) # your train set labels
183

184
    test_dataset = h5py.File('datasets/test_happy.h5', "r")
185
    test_set_x_orig = np.array(test_dataset["test_set_x"][:]) # your test set features
186
    test_set_y_orig = np.array(test_dataset["test_set_y"][:]) # your test set labels
187

188
    classes = np.array(test_dataset["list_classes"][:]) # the list of classes
189
    
190
    train_set_y_orig = train_set_y_orig.reshape((1, train_set_y_orig.shape[0]))
191
    test_set_y_orig = test_set_y_orig.reshape((1, test_set_y_orig.shape[0]))
192
    
193
    return train_set_x_orig, train_set_y_orig, test_set_x_orig, test_set_y_orig, classes
194

195
def img_to_encoding(image_path, model):
196
    #img = PIL.Image.open(image_path)
197
    img = tf.keras.preprocessing.image.load_img(image_path)
198
    #img = img1[...,::-1]
199
    img = np.around(np.transpose(img, (2,0,1))/255.0, decimals=12)
200
    x_train = np.expand_dims(img, axis=0)
201
    print(x_train.shape)
202
    embedding = model.predict_on_batch(x_train)
203
    return embedding
204
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.
