File size: 6,177 Bytes
5620c0c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
e09159b
5620c0c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import gradio as gr
import tensorflow as tf
from PIL import Image
from tensorflow.keras.applications import vgg19

from tensorflow.python.keras import models 
import numpy as np
from tensorflow import keras
from werkzeug.utils import secure_filename
import os
from flask import send_file

os.environ["CUDA_VISIBLE_DEVICES"] = "-1"

img_nrows = 256
img_ncols = 300

def img_preprocess(image_path):
    # Util function to open, resize and format pictures into appropriate tensors
    # img = keras.preprocessing.image.load_img(
    #     image_path, target_size=(img_nrows, img_ncols)
    # )
    img = keras.preprocessing.image.img_to_array(image_path)
    img = np.expand_dims(img, axis=0)
    img = vgg19.preprocess_input(img)
    return tf.convert_to_tensor(img)

def deprocess_img(processed_img):
  x = processed_img.copy()
  if len(x.shape) == 4:
    x = np.squeeze(x, 0)
  assert len(x.shape) == 3 #Input dimension must be [1, height, width, channel] or [height, width, channel]
  
  
  # perform the inverse of the preprocessing step
  x[:, :, 0] += 103.939
  x[:, :, 1] += 116.779
  x[:, :, 2] += 123.68
  x = x[:, :, ::-1] # converting BGR to RGB channel

  x = np.clip(x, 0, 255).astype('uint8')
  return x

content_layers = ['block5_conv2']
style_layers = ['block1_conv1',
                'block2_conv1',
                'block3_conv1', 
                'block4_conv1', 
                'block5_conv1']
number_content=len(content_layers)
number_style =len(style_layers)
def get_model():
    
    vgg=tf.keras.applications.vgg19.VGG19(include_top=False,weights='imagenet')
    vgg.trainable=False
    content_output=[vgg.get_layer(layer).output for layer in content_layers]
    style_output=[vgg.get_layer(layer).output for layer in style_layers]
    model_output= style_output+content_output
    return models.Model(vgg.input,model_output)

def get_content_loss(noise,target):
    loss = tf.reduce_mean(tf.square(noise-target))
    return loss

def gram_matrix(tensor):
    channels=int(tensor.shape[-1])
    vector=tf.reshape(tensor,[-1,channels])
    n=tf.shape(vector)[0]
    gram_matrix=tf.matmul(vector,vector,transpose_a=True)
    return gram_matrix/tf.cast(n,tf.float32)

def get_style_loss(noise,target):
    gram_noise=gram_matrix(noise)
    #gram_target=gram_matrix(target)
    loss=tf.reduce_mean(tf.square(target-gram_noise))
    return loss

def get_features(model,content_path,style_path):
    content_img=img_preprocess(content_path)
    style_image=img_preprocess(style_path)
    
    content_output=model(content_img)
    style_output=model(style_image)
    
    content_feature = [layer[0] for layer in content_output[number_style:]]
    style_feature = [layer[0] for layer in style_output[:number_style]]
    return content_feature,style_feature

def compute_loss(model, loss_weights,image, gram_style_features, content_features):
    style_weight,content_weight = loss_weights #style weight and content weight are user given parameters
                                               #that define what percentage of content and/or style will be preserved in the generated image
    
    output=model(image)
    content_loss=0
    style_loss=0
    
    noise_style_features = output[:number_style]
    noise_content_feature = output[number_style:]
    
    weight_per_layer = 1.0/float(number_style)
    for a,b in zip(gram_style_features,noise_style_features):
        style_loss+=weight_per_layer*get_style_loss(b[0],a)
        
    
    weight_per_layer =1.0/ float(number_content)
    for a,b in zip(noise_content_feature,content_features):
        content_loss+=weight_per_layer*get_content_loss(a[0],b)
        
    style_loss *= style_weight
    content_loss *= content_weight
    
    total_loss = content_loss + style_loss
    
    
    return total_loss,style_loss,content_loss

def compute_grads(dictionary):
    with tf.GradientTape() as tape:
        all_loss=compute_loss(**dictionary)
        
    total_loss=all_loss[0]
    return tape.gradient(total_loss,dictionary['image']),all_loss

def run_style_transfer(content_path,style_path,epochs=20,content_weight=1e3, style_weight=1e-2):
    
    model=get_model()
    
    for layer in model.layers:
        layer.trainable = False
        
    content_feature,style_feature = get_features(model,content_path,style_path)
    style_gram_matrix=[gram_matrix(feature) for feature in style_feature]
    
    noise = img_preprocess(content_path)
    noise=tf.Variable(noise,dtype=tf.float32)
    
    optimizer = tf.keras.optimizers.Adam(learning_rate=5, beta_1=0.99, epsilon=1e-1)
    
    best_loss,best_img=float('inf'),None
    
    loss_weights = (style_weight, content_weight)
    dictionary={'model':model,
              'loss_weights':loss_weights,
              'image':noise,
              'gram_style_features':style_gram_matrix,
              'content_features':content_feature}
    
    norm_means = np.array([103.939, 116.779, 123.68])
    min_vals = -norm_means
    max_vals = 255 - norm_means   
  
    imgs = []
    for i in range(1,epochs+1):
        grad,all_loss=compute_grads(dictionary)
        total_loss,style_loss,content_loss=all_loss
        optimizer.apply_gradients([(grad,noise)])
        clipped=tf.clip_by_value(noise,min_vals,max_vals)
        noise.assign(clipped)
        
        if total_loss<best_loss:
            best_loss = total_loss
            best_img = deprocess_img(noise.numpy())
            
         #for visualization   
            
        if i%1==0:
            plot_img = noise.numpy()
            plot_img = deprocess_img(plot_img)
            imgs.append(plot_img)

    
    return best_img,best_loss,imgs

content_path = "3.jpg"
style_path = "4.jpg"
def predict(image1_input, image2_input):
   
    return run_style_transfer(image1_input,image2_input,epochs=60)[0]
    

image1_input = gr.inputs.Image(label="Image 1")
image2_input = gr.inputs.Image(label="Image 2")
output_image = gr.outputs.Image(label="Merged Image", type="filepath")

title = "Image Merger"
description = "Merge two input images"

gr.Interface(fn=predict, inputs=[image1_input, image2_input], outputs=output_image, title=title, description=description).launch()