In [63]:

    

%matplotlib inline

from keras.datasets import mnist
from keras.layers import Input, Dense, Lambda
from keras.models import Model
from keras.objectives import binary_crossentropy
from keras.callbacks import LearningRateScheduler


    

In [3]:

    

import numpy as np
import matplotlib.pyplot as plt
import keras.backend as K
import tensorflow as tf


    

In [75]:

    

# バッチサイズ
m = 50

# 潜在変数の次元数
n_z = 2

n_epoch = 10


    

In [76]:

    

# Q(z|X) -- encoder
inputs = Input(shape=(784, ))
h_q = Dense(512, activation='relu')(inputs)

# 中間層のh_qからmuとsigmaを生成
# zは N(mu, sigma) からサンプリングされる
mu = Dense(n_z, activation='linear')(h_q)
log_sigma = Dense(n_z, activation='linear')(h_q)


    

In [77]:

    

def sample_z(args):
    mu, log_sigma = args
    eps = K.random_normal(shape=(m, n_z), mean=0.0, stddev=1.0)
    return mu + K.exp(log_sigma / 2) * eps


    

In [78]:

    

sample_z((0, 1))


    

    
Out[78]:





<tf.Tensor 'add_69:0' shape=(50, 2) dtype=float32>

In [79]:

    

# encoderの出力 [mu, log_sigma] を使ってサンプリング
# m個のサンプルzが得られる
# sample z ~ Q(z|X)
z = Lambda(sample_z)([mu, log_sigma])


    

In [80]:

    

# P(X|z) -- decoder
decoder_hidden = Dense(512, activation='relu')
decoder_out = Dense(784, activation='sigmoid')

h_p = decoder_hidden(z)
outputs = decoder_out(h_p)


    

In [81]:

    

outputs


    

    
Out[81]:





<tf.Tensor 'dense_12/Sigmoid:0' shape=(50, 784) dtype=float32>

In [82]:

    

# Overall VAE model, for reconstruction and training
vae = Model(inputs, outputs)

# Encoder model, to encode input into latent variable
# We use the mean as the output as it is the center point, representative of Gaussian
encoder = Model(inputs, mu)


    

In [83]:

    

# Generator model, generate new data given latent variable z
d_in = Input(shape=(n_z, ))
d_h = decoder_hidden(d_in)
d_out = decoder_out(d_h)
decoder = Model(d_in, d_out)


    

In [84]:

    

def vae_loss(y_true, y_pred):
    # E[log P(X|z)] => max  <=>  zから生成されたXと元のXの誤差が小さい
    # axis=1はBatch方向に和をとる
    recon = K.sum(K.binary_crossentropy(y_pred, y_true), axis=1)
    # D_KL(Q(z|x) || P(z)) => min  <=> 小さくしたいのでそのままlossとする
    kl = 0.5 * K.sum(K.exp(log_sigma) + K.square(mu) - 1.0 - log_sigma, axis=1)
    return recon + kl


    

In [85]:

    

vae.compile(optimizer='adam', loss=vae_loss)


    

In [86]:

    

vae


    

    
Out[86]:





<keras.engine.training.Model at 0x122380160>

In [87]:

    

(X_train, y_train), (X_test, y_test) = mnist.load_data()


    

In [88]:

    

X_train = X_train.astype('float32') / 255.0
X_test = X_test.astype('float32') / 255.0


    

In [39]:

    

X_train.shape


    

    
Out[39]:





(60000, 28, 28)

In [90]:

    

X_train = X_train.reshape((-1, 28 * 28))
X_test = X_test.reshape((-1, 28 * 28))


    

In [91]:

    

X_train.shape


    

    
Out[91]:





(60000, 784)

In [97]:

    

vae.fit(X_train, X_train, batch_size=m, epochs=5)


    

    




Epoch 1/5
60000/60000 [==============================] - 17s - loss: 156.5659    
Epoch 2/5
60000/60000 [==============================] - 16s - loss: 155.5695    
Epoch 3/5
60000/60000 [==============================] - 15s - loss: 154.8109    
Epoch 4/5
60000/60000 [==============================] - 15s - loss: 154.1780    
Epoch 5/5
60000/60000 [==============================] - 15s - loss: 153.5644    






    
Out[97]:





<keras.callbacks.History at 0x12b5fd240>

In [98]:

    

z_test = encoder.predict(X_test)


    

In [99]:

    

z_test.shape


    

    
Out[99]:





(10000, 2)

In [100]:

    

y_test.shape


    

    
Out[100]:





(10000,)

In [101]:

    

plt.figure()
plt.scatter(z_test[:, 0], z_test[:, 1], c=y_test)
plt.colorbar()


    

    
Out[101]:





<matplotlib.colorbar.Colorbar at 0x128dbb518>






    

In [103]:

    

z = np.random.normal(size=100)


    

In [105]:

    

z = z.reshape((50, 2))


    

In [106]:

    

z.shape


    

    
Out[106]:





(50, 2)

In [107]:

    

decoder.summary()


    

    




_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
input_5 (InputLayer)         (None, 2)                 0         
_________________________________________________________________
dense_11 (Dense)             multiple                  1536      
_________________________________________________________________
dense_12 (Dense)             multiple                  402192    
=================================================================
Total params: 403,728
Trainable params: 403,728
Non-trainable params: 0
_________________________________________________________________

In [110]:

    

x = decoder.predict(z)


    

In [111]:

    

x.shape


    

    
Out[111]:





(50, 784)

In [112]:

    

x = x.reshape(-1, 28, 28)


    

In [113]:

    

x.shape


    

    
Out[113]:





(50, 28, 28)

In [122]:

    

plt.imshow(x[0], cmap='gray')


    

    
Out[122]:





<matplotlib.image.AxesImage at 0x12a4897f0>






    

In [ ]: