-
Notifications
You must be signed in to change notification settings - Fork 0
/
NEURAL_STYLE_TRANSFER_TF.py
221 lines (193 loc) · 8.28 KB
/
NEURAL_STYLE_TRANSFER_TF.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
# -*- coding: utf-8 -*-
"""
Created on Sat Feb 24 13:11:44 2018
@author: ashima.garg
"""
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import tensorflow as tf
import time
import os
import sys
import cv2
import numpy as np
import tensorflow as tf
import matplotlib.pyplot as plt
import scipy.io
from keras.applications.vgg19 import VGG19
import matplotlib.pyplot as plt
from matplotlib.pyplot import imshow
from keras.utils.vis_utils import plot_model
from keras.applications.vgg19 import preprocess_input
from keras.preprocessing.image import img_to_array
from keras import backend as K
from PIL import Image
IMAGE_WIDTH = 400
IMAGE_HEIGHT = 300
COLOR_CHANNELS = 3
noise_ratio = 0.6
alpha = 10
beta = 40
learning_rate = 2.0
num_iterations = 400
STYLE_LAYERS = [('conv1_1', 0.2), ('conv2_1', 0.2), ('conv3_1', 0.2), ('conv4_1', 0.2), ('conv5_1', 0.2)]
vggmodel = scipy.io.loadmat('pretrainedModel\imagenet-vgg-verydeep-19.mat')
print("model loaded")
vgg_layers = vggmodel['layers'][0]
def get_weights_bias(layer):
W = vgg_layers[layer][0][0][2][0][0]
b = vgg_layers[layer][0][0][2][0][1]
W = tf.constant(W)
b = tf.constant(np.reshape(b, (b.size)))
return W, b
def conv2d_layer(pre_layer, layer_name, layer):
weight, bias = get_weights_bias(layer)
conv = tf.nn.conv2d(pre_layer, weight, [1, 1, 1, 1], padding="SAME")
return tf.nn.bias_add(conv, bias)
def avg_pool_layer(pre_layer, layer_name):
return tf.nn.avg_pool(pre_layer, ksize = [1, 2, 2, 1], strides = [1, 2, 2, 1], padding = "SAME")
def relu_layer(conv2d_layer, layer_name, layer):
return tf.nn.relu(conv2d_layer)
def build_model():
print('\nBUILDING VGG-19 NETWORK')
graph = {}
print('INPUT')
graph['input'] = tf.Variable(np.zeros((1, IMAGE_HEIGHT, IMAGE_WIDTH, COLOR_CHANNELS)), dtype = np.float32)
print('LAYER GROUP 1')
graph['conv1_1'] = conv2d_layer(graph['input'], 'conv1_1', 0)
graph['relu1_1'] = relu_layer(graph['conv1_1'], 'relu1_1', 0)
graph['conv1_2'] = conv2d_layer(graph['relu1_1'], 'conv1_2', 2)
graph['relu1_2'] = relu_layer(graph['conv1_2'], 'relu1_2', 2)
graph['pool1'] = avg_pool_layer(graph['relu1_2'], 'pool1')
print('LAYER GROUP 2')
graph['conv2_1'] = conv2d_layer(graph['pool1'], 'conv2_1', 5)
graph['relu2_1'] = relu_layer(graph['conv2_1'], 'relu2_1', 5)
graph['conv2_2'] = conv2d_layer(graph['relu2_1'], 'conv2_2', 7)
graph['relu2_2'] = relu_layer(graph['conv2_2'], 'relu2_2', 7)
graph['pool2'] = avg_pool_layer(graph['relu2_2'], 'pool2')
print('LAYER GROUP 3')
graph['conv3_1'] = conv2d_layer(graph['pool2'], 'conv3_1', 10)
graph['relu3_1'] = relu_layer(graph['conv3_1'], 'relu3_1', 10)
graph['conv3_2'] = conv2d_layer(graph['relu3_1'], 'conv3_2', 12)
graph['relu3_2'] = relu_layer(graph['conv3_2'], 'relu3_2', 12)
graph['conv3_3'] = conv2d_layer(graph['relu3_2'], 'conv3_3', 14)
graph['relu3_3'] = relu_layer(graph['conv3_3'], 'relu3_3', 14)
graph['conv3_4'] = conv2d_layer(graph['relu3_3'], 'conv3_4', 16)
graph['relu3_4'] = relu_layer(graph['conv3_4'], 'relu3_4', 16)
graph['pool3'] = avg_pool_layer(graph['relu3_4'], 'pool3')
print('LAYER GROUP 4')
graph['conv4_1'] = conv2d_layer(graph['pool3'], 'conv4_1', 19)
graph['relu4_1'] = relu_layer(graph['conv4_1'], 'relu4_1', 19)
graph['conv4_2'] = conv2d_layer(graph['relu4_1'], 'conv4_2', 21)
graph['relu4_2'] = relu_layer(graph['conv4_2'], 'relu4_2', 21)
graph['conv4_3'] = conv2d_layer(graph['relu4_2'], 'conv4_3', 23)
graph['relu4_3'] = relu_layer(graph['conv4_3'], 'relu4_3', 23)
graph['conv4_4'] = conv2d_layer(graph['relu4_3'], 'conv4_4', 25)
graph['relu4_4'] = relu_layer(graph['conv4_4'], 'relu4_4', 25)
graph['pool4'] = avg_pool_layer(graph['relu4_4'], 'pool4')
print('LAYER GROUP 5')
graph['conv5_1'] = conv2d_layer(graph['pool4'], 'conv5_1', 28)
graph['relu5_1'] = relu_layer(graph['conv5_1'], 'relu5_1', 28)
graph['conv5_2'] = conv2d_layer(graph['relu5_1'], 'conv5_2', 30)
graph['relu5_2'] = relu_layer(graph['conv5_2'], 'relu5_2', 30)
graph['conv5_3'] = conv2d_layer(graph['relu5_2'], 'conv5_3', 32)
graph['relu5_3'] = relu_layer(graph['conv5_3'], 'relu5_3', 32)
graph['conv5_4'] = conv2d_layer(graph['relu5_3'], 'conv5_4', 34)
graph['relu5_4'] = relu_layer(graph['conv5_4'], 'relu5_4', 34)
graph['pool5'] = avg_pool_layer(graph['relu5_4'], 'pool5')
return graph
MEANS = np.array([123.68, 116.779, 103.939]).reshape((1,1,1,3))
def preprocess(image) :
image = cv2.resize(image, (IMAGE_WIDTH, IMAGE_HEIGHT), interpolation=cv2.INTER_AREA)
image = image.reshape((1, image.shape[0], image.shape[1], image.shape[2]))
image = image - MEANS
return image
def generate_noise_image(contentImage):
noiseImage = np.random.uniform(-20., 20., (1, IMAGE_HEIGHT, IMAGE_WIDTH, COLOR_CHANNELS)).astype(np.float32)
genimage = noiseImage * noise_ratio + contentImage * (1. - noise_ratio)
return genimage
def compute_content_cost(a_C, a_G) :
m, n_H, n_W, n_C = a_G.get_shape().as_list()
con = tf.reshape(a_C,[n_H*n_W, n_C])
gen = tf.reshape(a_G,[n_H*n_W, n_C])
con = tf.transpose(con)
gen = tf.transpose(gen)
cost = (1./(4 * n_H * n_W * n_C)) * tf.reduce_sum(tf.square(tf.subtract(con, gen)))
return cost
def gram_matrix(image, m, n_H, n_W, n_C):
img = tf.transpose(tf.reshape(image, [n_H*n_W, n_C]))
gram = tf.matmul(img, tf.transpose(img))
return gram
def compute_style_cost(styleImg, generatedImg):
m, n_H, n_W, n_C = generatedImg.get_shape().as_list()
styleG = gram_matrix(styleImg, m, n_H, n_W, n_C)
generateG = gram_matrix(generatedImg, m, n_H, n_W, n_C)
size = n_H*n_W
cost = tf.multiply(tf.divide(1, (4 * size**2 * n_C**2)) , tf.reduce_sum(tf.square(styleG - generateG)))
return cost
def compute_total_style_cost(model, sess):
style_cost = 0
for layer_name, coeff in STYLE_LAYERS:
a_G = model[layer_name]
a_S = sess.run(model[layer_name])
style_cost += coeff * compute_style_cost(a_S, a_G)
return style_cost
def read_image(path):
img = cv2.imread(path, 1)
img = img.astype(np.float32)
img = preprocess(img)
return img
def postprocess(image):
image = image + MEANS
image = image[0]
image = np.clip(image, 0, 255).astype('uint8')
return image
def display(name, image):
cv2.imshow(name, image)
cv2.waitKey(0)
return
def trainNetwork():
print('print style and content images')
path = os.path.dirname(os.path.realpath(__file__))
file = "\\Images\\"
dirname = path + file
stylePath = dirname + "1-style.jpg"
styleImg = read_image(stylePath)
contentPath = dirname + "1-content.jpg"
contentImg = read_image(contentPath)
generatedImg = generate_noise_image(contentImg)
print("generated Image " + str(generatedImg.shape))
display('STYLE', styleImg[0])
display('CONTENT', contentImg[0])
display('GENERATED', generatedImg[0])
tf.reset_default_graph()
with tf.Session() as sess:
model = build_model()
print("model created ")
sess.run(model['input'].assign(contentImg))
a_G = model['conv4_2']
a_C = sess.run(model['conv4_2'])
J_content = compute_content_cost(a_C, a_G)
sess.run(model['input'].assign(styleImg))
J_style = compute_total_style_cost(model, sess)
J_total = beta * J_style + alpha *
optimizer = tf.train.AdamOptimizer(learning_rate)
train_step = optimizer.minimize(J_total)
sess.run(tf.global_variables_initializer())
sess.run(model['input'].assign(generatedImg))
print("generated image assigned")
for i in range(num_iterations):
sess.run(train_step)
generatedImage = sess.run(model['input'])
if i%20 == 0:
J_Content, J_Style, J_Total = sess.run([J_content, J_style, J_total])
print("J_Content " +str(J_Content))
print("J_Style " +str(J_Style))
print("J_total " + str(J_Total))
generatedImage = postprocess(generatedImage)
cv2.imwrite('Outputs\generatedImg' + str(i) + '.jpg', generatedImage)
return
if __name__ == '__main__':
trainNetwork()
print("Completed")