Denoising Autoencoder

from keras.models import Model
from keras.layers import Dense, Input
from keras.datasets import mnist
import numpy as np

# Hyper parameters
batch_size = 128
nb_epoch = 100

# Parameters for MNIST dataset
img_rows, img_cols = 28, 28

# Parameters for denoising autoencoder
nb_visible = img_rows * img_cols
nb_hidden = 500
corruption_level = 0.3

Using TensorFlow backend.

# Load MNIST Dataset
(x_train, _), (x_test, _) = mnist.load_data()
x_train = x_train.astype('float32') / 255.
x_test = x_test.astype('float32') / 255.
x_train = x_train.reshape((len(x_train), np.prod(x_train.shape[1:])))
x_test = x_test.reshape((len(x_test), np.prod(x_test.shape[1:])))
print(x_train.shape)
print(x_test.shape)

(60000, 784)
(10000, 784)

# Add random noise
x_train_noisy = x_train + corruption_level * np.random.normal(loc=0.0, scale=1.0, size=x_train.shape)
x_test_noisy = x_test + corruption_level * np.random.normal(loc=0.0, scale=1.0, size=x_test.shape)
x_train_noisy = np.clip(x_train_noisy, 0., 1.)
x_test_noisy = np.clip(x_test_noisy, 0., 1.)
print(x_train_noisy.shape)
print(x_test_noisy.shape)

(60000, 784)
(10000, 784)

# Build autoencoder model
input_img = Input(shape=(nb_visible,))
encoded = Dense(nb_hidden, activation='relu')(input_img)
decoded = Dense(nb_visible, activation='sigmoid')(encoded)

autoencoder = Model(input=input_img, output=decoded)
autoencoder.compile(optimizer='adadelta', loss='binary_crossentropy')
autoencoder.summary()

____________________________________________________________________________________________________
Layer (type)                     Output Shape          Param #     Connected to                     
====================================================================================================
input_1 (InputLayer)             (None, 784)           0                                            
____________________________________________________________________________________________________
dense_1 (Dense)                  (None, 500)           392500      input_1[0][0]                    
____________________________________________________________________________________________________
dense_2 (Dense)                  (None, 784)           392784      dense_1[0][0]                    
====================================================================================================
Total params: 785284
____________________________________________________________________________________________________

# Train
autoencoder.fit(x_train_noisy, x_train,
                nb_epoch=nb_epoch, batch_size=batch_size, shuffle=True, verbose=1,
                validation_data=(x_test_noisy, x_test))

Train on 60000 samples, validate on 10000 samples
Epoch 1/100
60000/60000 [==============================] - 3s - loss: 0.2751 - val_loss: 0.2242
Epoch 2/100
60000/60000 [==============================] - 3s - loss: 0.2033 - val_loss: 0.1844
Epoch 3/100
60000/60000 [==============================] - 3s - loss: 0.1745 - val_loss: 0.1633
Epoch 4/100
60000/60000 [==============================] - 3s - loss: 0.1572 - val_loss: 0.1490
Epoch 5/100
60000/60000 [==============================] - 2s - loss: 0.1450 - val_loss: 0.1387
Epoch 6/100
60000/60000 [==============================] - 3s - loss: 0.1360 - val_loss: 0.1309
Epoch 7/100
60000/60000 [==============================] - 3s - loss: 0.1292 - val_loss: 0.1249
Epoch 8/100
60000/60000 [==============================] - 3s - loss: 0.1238 - val_loss: 0.1201
Epoch 9/100
60000/60000 [==============================] - 3s - loss: 0.1195 - val_loss: 0.1163
Epoch 10/100
60000/60000 [==============================] - 3s - loss: 0.1160 - val_loss: 0.1131
Epoch 11/100
60000/60000 [==============================] - 2s - loss: 0.1131 - val_loss: 0.1105
Epoch 12/100
60000/60000 [==============================] - 3s - loss: 0.1106 - val_loss: 0.1082
Epoch 13/100
60000/60000 [==============================] - 3s - loss: 0.1085 - val_loss: 0.1063
Epoch 14/100
60000/60000 [==============================] - 3s - loss: 0.1067 - val_loss: 0.1046
Epoch 15/100
60000/60000 [==============================] - 3s - loss: 0.1051 - val_loss: 0.1032
Epoch 16/100
60000/60000 [==============================] - 3s - loss: 0.1037 - val_loss: 0.1019
Epoch 17/100
60000/60000 [==============================] - 3s - loss: 0.1025 - val_loss: 0.1008
Epoch 18/100
60000/60000 [==============================] - 3s - loss: 0.1014 - val_loss: 0.0997
Epoch 19/100
60000/60000 [==============================] - 3s - loss: 0.1003 - val_loss: 0.0988
Epoch 20/100
60000/60000 [==============================] - 3s - loss: 0.0994 - val_loss: 0.0980
Epoch 21/100
60000/60000 [==============================] - 3s - loss: 0.0986 - val_loss: 0.0972
Epoch 22/100
60000/60000 [==============================] - 2s - loss: 0.0979 - val_loss: 0.0965
Epoch 23/100
60000/60000 [==============================] - 2s - loss: 0.0972 - val_loss: 0.0959
Epoch 24/100
60000/60000 [==============================] - 3s - loss: 0.0965 - val_loss: 0.0953
Epoch 25/100
60000/60000 [==============================] - 3s - loss: 0.0959 - val_loss: 0.0948
Epoch 26/100
60000/60000 [==============================] - 3s - loss: 0.0954 - val_loss: 0.0943
Epoch 27/100
60000/60000 [==============================] - 3s - loss: 0.0949 - val_loss: 0.0938
Epoch 28/100
60000/60000 [==============================] - 3s - loss: 0.0944 - val_loss: 0.0934
Epoch 29/100
60000/60000 [==============================] - 3s - loss: 0.0939 - val_loss: 0.0929
Epoch 30/100
60000/60000 [==============================] - 2s - loss: 0.0935 - val_loss: 0.0926
Epoch 31/100
60000/60000 [==============================] - 3s - loss: 0.0931 - val_loss: 0.0922
Epoch 32/100
60000/60000 [==============================] - 2s - loss: 0.0928 - val_loss: 0.0919
Epoch 33/100
60000/60000 [==============================] - 3s - loss: 0.0924 - val_loss: 0.0916
Epoch 34/100
60000/60000 [==============================] - 2s - loss: 0.0921 - val_loss: 0.0913
Epoch 35/100
60000/60000 [==============================] - 3s - loss: 0.0918 - val_loss: 0.0910
Epoch 36/100
60000/60000 [==============================] - 3s - loss: 0.0915 - val_loss: 0.0907
Epoch 37/100
60000/60000 [==============================] - 3s - loss: 0.0912 - val_loss: 0.0904
Epoch 38/100
60000/60000 [==============================] - 3s - loss: 0.0909 - val_loss: 0.0902
Epoch 39/100
60000/60000 [==============================] - 3s - loss: 0.0906 - val_loss: 0.0900
Epoch 40/100
60000/60000 [==============================] - 2s - loss: 0.0904 - val_loss: 0.0897
Epoch 41/100
60000/60000 [==============================] - 3s - loss: 0.0902 - val_loss: 0.0895
Epoch 42/100
60000/60000 [==============================] - 3s - loss: 0.0899 - val_loss: 0.0893
Epoch 43/100
60000/60000 [==============================] - 3s - loss: 0.0897 - val_loss: 0.0891
Epoch 44/100
60000/60000 [==============================] - 3s - loss: 0.0895 - val_loss: 0.0890
Epoch 45/100
60000/60000 [==============================] - 2s - loss: 0.0893 - val_loss: 0.0888
Epoch 46/100
60000/60000 [==============================] - 3s - loss: 0.0891 - val_loss: 0.0886
Epoch 47/100
60000/60000 [==============================] - 3s - loss: 0.0890 - val_loss: 0.0884
Epoch 48/100
60000/60000 [==============================] - 3s - loss: 0.0888 - val_loss: 0.0883
Epoch 49/100
60000/60000 [==============================] - 3s - loss: 0.0886 - val_loss: 0.0881
Epoch 50/100
60000/60000 [==============================] - 3s - loss: 0.0884 - val_loss: 0.0880
Epoch 51/100
60000/60000 [==============================] - 3s - loss: 0.0883 - val_loss: 0.0878
Epoch 52/100
60000/60000 [==============================] - 3s - loss: 0.0881 - val_loss: 0.0877
Epoch 53/100
60000/60000 [==============================] - 2s - loss: 0.0880 - val_loss: 0.0876
Epoch 54/100
60000/60000 [==============================] - 3s - loss: 0.0878 - val_loss: 0.0874
Epoch 55/100
60000/60000 [==============================] - 3s - loss: 0.0877 - val_loss: 0.0873
Epoch 56/100
60000/60000 [==============================] - 2s - loss: 0.0876 - val_loss: 0.0872
Epoch 57/100
60000/60000 [==============================] - 3s - loss: 0.0874 - val_loss: 0.0871
Epoch 58/100
60000/60000 [==============================] - 3s - loss: 0.0873 - val_loss: 0.0870
Epoch 59/100
60000/60000 [==============================] - 3s - loss: 0.0872 - val_loss: 0.0869
Epoch 60/100
60000/60000 [==============================] - 3s - loss: 0.0871 - val_loss: 0.0868
Epoch 61/100
60000/60000 [==============================] - 3s - loss: 0.0870 - val_loss: 0.0867
Epoch 62/100
60000/60000 [==============================] - 3s - loss: 0.0868 - val_loss: 0.0866
Epoch 63/100
60000/60000 [==============================] - 3s - loss: 0.0867 - val_loss: 0.0865
Epoch 64/100
60000/60000 [==============================] - 2s - loss: 0.0866 - val_loss: 0.0864
Epoch 65/100
60000/60000 [==============================] - 3s - loss: 0.0865 - val_loss: 0.0863
Epoch 66/100
60000/60000 [==============================] - 3s - loss: 0.0864 - val_loss: 0.0862
Epoch 67/100
60000/60000 [==============================] - 3s - loss: 0.0863 - val_loss: 0.0861
Epoch 68/100
60000/60000 [==============================] - 3s - loss: 0.0862 - val_loss: 0.0860
Epoch 69/100
60000/60000 [==============================] - 3s - loss: 0.0861 - val_loss: 0.0859
Epoch 70/100
60000/60000 [==============================] - 3s - loss: 0.0860 - val_loss: 0.0859
Epoch 71/100
60000/60000 [==============================] - 3s - loss: 0.0860 - val_loss: 0.0858
Epoch 72/100
60000/60000 [==============================] - 3s - loss: 0.0859 - val_loss: 0.0857
Epoch 73/100
60000/60000 [==============================] - 3s - loss: 0.0858 - val_loss: 0.0857
Epoch 74/100
60000/60000 [==============================] - 3s - loss: 0.0857 - val_loss: 0.0856
Epoch 75/100
60000/60000 [==============================] - 3s - loss: 0.0856 - val_loss: 0.0855
Epoch 76/100
60000/60000 [==============================] - 2s - loss: 0.0855 - val_loss: 0.0854
Epoch 77/100
60000/60000 [==============================] - 2s - loss: 0.0855 - val_loss: 0.0854
Epoch 78/100
60000/60000 [==============================] - 3s - loss: 0.0854 - val_loss: 0.0853
Epoch 79/100
60000/60000 [==============================] - 3s - loss: 0.0853 - val_loss: 0.0852
Epoch 80/100
60000/60000 [==============================] - 2s - loss: 0.0852 - val_loss: 0.0852
Epoch 81/100
60000/60000 [==============================] - 3s - loss: 0.0852 - val_loss: 0.0851
Epoch 82/100
60000/60000 [==============================] - 3s - loss: 0.0851 - val_loss: 0.0850
Epoch 83/100
60000/60000 [==============================] - 3s - loss: 0.0850 - val_loss: 0.0850
Epoch 84/100
60000/60000 [==============================] - 3s - loss: 0.0850 - val_loss: 0.0849
Epoch 85/100
60000/60000 [==============================] - 3s - loss: 0.0849 - val_loss: 0.0849
Epoch 86/100
60000/60000 [==============================] - 3s - loss: 0.0848 - val_loss: 0.0848
Epoch 87/100
60000/60000 [==============================] - 3s - loss: 0.0848 - val_loss: 0.0848
Epoch 88/100
60000/60000 [==============================] - 3s - loss: 0.0847 - val_loss: 0.0847
Epoch 89/100
60000/60000 [==============================] - 3s - loss: 0.0847 - val_loss: 0.0847
Epoch 90/100
60000/60000 [==============================] - 3s - loss: 0.0846 - val_loss: 0.0846
Epoch 91/100
60000/60000 [==============================] - 3s - loss: 0.0846 - val_loss: 0.0846
Epoch 92/100
60000/60000 [==============================] - 3s - loss: 0.0845 - val_loss: 0.0845
Epoch 93/100
60000/60000 [==============================] - 2s - loss: 0.0844 - val_loss: 0.0845
Epoch 94/100
60000/60000 [==============================] - 3s - loss: 0.0844 - val_loss: 0.0844
Epoch 95/100
60000/60000 [==============================] - 3s - loss: 0.0843 - val_loss: 0.0844
Epoch 96/100
60000/60000 [==============================] - 2s - loss: 0.0843 - val_loss: 0.0843
Epoch 97/100
60000/60000 [==============================] - 3s - loss: 0.0842 - val_loss: 0.0843
Epoch 98/100
60000/60000 [==============================] - 3s - loss: 0.0842 - val_loss: 0.0842
Epoch 99/100
60000/60000 [==============================] - 3s - loss: 0.0841 - val_loss: 0.0842
Epoch 100/100
60000/60000 [==============================] - 2s - loss: 0.0841 - val_loss: 0.0842





<keras.callbacks.History at 0x7f905a312ba8>

# Evaluate
evaluation = autoencoder.evaluate(x_test_noisy, x_test, batch_size=batch_size, verbose=1)
print('\nSummary: Loss over the test dataset: %.2f' % (evaluation))

 9856/10000 [============================>.] - ETA: 0s
Summary: Loss over the test dataset: 0.08

import matplotlib.pyplot as plt
%matplotlib inline

# Decode test images 
decoded_imgs = autoencoder.predict(x_test_noisy)

n = 10  # how many digits we will display
plt.figure(figsize=(20, 4))
for i in range(n):
    # display original
    ax = plt.subplot(2, n, i + 1)
    plt.imshow(x_test_noisy[i].reshape(28, 28))
    plt.gray()
    ax.get_xaxis().set_visible(False)
    ax.get_yaxis().set_visible(False)

    # display reconstruction
    ax = plt.subplot(2, n, i + 1 + n)
    plt.imshow(decoded_imgs[i].reshape(28, 28))
    plt.gray()
    ax.get_xaxis().set_visible(False)
    ax.get_yaxis().set_visible(False)
plt.show()

Denoising Autoencoder

Denoising Autoencoder

results matching ""

No results matching ""