1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
# hiding warning tensorflow
import os
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '3' 

# import cac thu vien
import tensorflow as tf
from tensorflow import keras
import matplotlib.pyplot as plt
import cv2
import numpy as np

# Load du lieu mnist tu keras
(x_train, y_train), (x_test, y_test) = tf.keras.datasets.mnist.load_data()

# Hien thi anh trong dataset voi cmap = 'gray'
plt.imshow(x_train[1234], cmap='gray')

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
# data processing (chuan hoa du lieu [0, 1])
x_train = x_train / 255.0
x_test = x_test / 255.0

# autoencoder

# su dung ham toi uu la Adam voi lr=1e-3, decay = 1e-6
opt = keras.optimizers.Adam(learning_rate=0.001, decay = 1e-6)

# dau vao encoder (28, 28, 1)
# processing cua encoder: tu dau vao --> flatten --> dense
encoder_input = keras.Input((28, 28, 1), name = 'img')
x = keras.layers.Flatten()(encoder_input)
encoder_output = keras.layers.Dense(64, activation='relu') (x)

# encoder (day la mot mo hinh voi 2 tham so (encoder_input, encoder_output))
encoder = keras.Model(encoder_input, encoder_output, name = 'encoder')

# processing decoder: dau vao la encoder_output --> qua lop Dense --> reshape thanh (28, 28, 1)
decoder_input = keras.layers.Dense(784, activation='relu') (encoder_output)
decoder_output = keras.layers.Reshape((28, 28, 1)) (decoder_input)

# autoencoder no la mot model voi 2 so (encoder_input vs decoder_output)
autoencoder = keras.Model(encoder_input, decoder_output, name = 'autoencoder')
autoencoder.summary()

# compilie vs fit data
autoencoder.compile(opt, loss='mse')
autoencoder.fit(x_train, x_train, epochs=10, batch_size = 32, validation_split=0.1)
# luu y, cho nay minh dung ca (x, y) = (x_train, x_train)

1
2
3
4
5
6
7
8
9
10
example = encoder.predict([x_test[3].reshape(-1, 28, 28, 1)]) [0]
print(example)
print(example.shape)
print(64/784)

plt.imshow(example.reshape((8, 8)), cmap = 'gray')
plt.imshow(x_test[3], cmap='gray')

ae_out = autoencoder.predict([x_test[3].reshape(-1, 28, 28, 1)])[0]
plt.imshow(ae_out, cmap='gray')

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
import random
def add_noise(img, random_chance=5):
    noisy = []
    for row in img:
        new_row = []
        for pix in row:
            if random.choice(range(100)) <= random_chance:
                new_val = random.uniform(0, 1)
                new_row.append(new_val)
            else:
                new_row.append(pix)
        noisy.append(new_row)
    return np.array(noisy)

noisy = add_noise(x_test[3])
plt.imshow(noisy, cmap = 'gray')

noisy = add_noise(x_test[3])
plt.imshow(noisy, cmap = 'gray')
	

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
# mot mo hinh auto encoder khac.
import tensorflow as tf
from tensorflow import keras

opt = keras.optimizers.Adam(learning_rate = 0.001, decay = 1e-6)

#1. 
encoder_input = keras.Input((64, 64, 3), name = 'img_in')
#2. 
x = keras.layers.Conv2D(64, (3, 3), activation = 'relu', padding='same') (encoder_input)
#3. 
x = keras.layers.MaxPooling2D((2, 2), padding = 'same')(x)
#4. 
x = keras.layers.Conv2D(64, (3, 3), activation='relu', padding='same') (x)
#5. 
x= keras.layers.MaxPooling2D((2, 2), padding = 'same') (x)
#6. 
x = keras.layers.Conv2D(32, (3, 3), activation='relu', padding='same') (x)
#7. 
x = keras.layers.MaxPooling2D((2, 2), padding='same') (x)
#8. 
x = keras.layers.Flatten() (x)
#9. 
encoder_output = keras.layers.Dense(512, activation='relu') (x)


encoder = keras.Model(encoder_input, encoder_output, name = 'encoder')


#9. 
decoder_input = keras.layers.Dense(2048, activation='relu') (encoder_output)
#8. 
x = keras.layers.Reshape((8, 8, 32)) (decoder_input)
#7. 
x = keras.layers.Conv2D(32, (3, 3), activation='relu', padding='same') (x)
#6. 
x = keras.layers.UpSampling2D((2, 2)) (x)
#5. 
x = keras.layers.Conv2D(64, (3, 3), activation='relu', padding='same') (x)
#4. 
x = keras.layers.UpSampling2D((2, 2)) (x)
#3. 
x = keras.layers.Conv2D(64, (3, 3), activation='relu', padding='same')(x)
#2. 
x = keras.layers.UpSampling2D((2, 2)) (x)
#1. 
decoder_output = keras.layers.Conv2D(3, (3, 3), activation='relu', padding='same') (x)

autoencoder = keras.Model(encoder_input, decoder_output, name = 'ae')
autoencoder.summary()