├── README.md
└── variational_autoencoder.py


/README.md:
--------------------------------------------------------------------------------
 1 | # autoencoder_explained
 2 | This is the code for "Autoencoder Explained" by Siraj Raval on Youtube 
 3 | 
 4 | # Coding Challenge - Due Date- ~Thursday, Feb 1 2018 
 5 | 
 6 | Create an autoencoder using Keras. Try to think of a cool use case. Bonus points for good documentation. Submit your github link in the comment section of this video. Good luck!
 7 | 
 8 | ## Overview
 9 | 
10 | This is the code for [this](https://youtu.be/H1AllrJ-_30) video on Youtube by Siraj Raval on autoencoders. 
11 | 
12 | ## Dependencies
13 | 
14 | * keras
15 | 
16 | Install keras [here](https://keras.io/). 
17 | 
18 | ## Usage 
19 | 
20 | Run the demo with python class_name.py in terminal 
21 | 
22 | ## Credits
23 | 
24 | Credits go to the Keras team
25 | 


--------------------------------------------------------------------------------
/variational_autoencoder.py:
--------------------------------------------------------------------------------
  1 | '''This script demonstrates how to build a variational autoencoder with Keras.
  2 | 
  3 |  #Reference
  4 | 
  5 |  - Auto-Encoding Variational Bayes
  6 |    https://arxiv.org/abs/1312.6114
  7 | '''
  8 | from __future__ import print_function
  9 | 
 10 | import numpy as np
 11 | import matplotlib.pyplot as plt
 12 | from scipy.stats import norm
 13 | 
 14 | from keras.layers import Input, Dense, Lambda
 15 | from keras.models import Model
 16 | from keras import backend as K
 17 | from keras import metrics
 18 | from keras.datasets import mnist
 19 | 
 20 | batch_size = 100
 21 | original_dim = 784
 22 | latent_dim = 2
 23 | intermediate_dim = 256
 24 | epochs = 50
 25 | epsilon_std = 1.0
 26 | 
 27 | 
 28 | x = Input(shape=(original_dim,))
 29 | h = Dense(intermediate_dim, activation='relu')(x)
 30 | z_mean = Dense(latent_dim)(h)
 31 | z_log_var = Dense(latent_dim)(h)
 32 | 
 33 | 
 34 | def sampling(args):
 35 |     z_mean, z_log_var = args
 36 |     epsilon = K.random_normal(shape=(K.shape(z_mean)[0], latent_dim), mean=0.,
 37 |                               stddev=epsilon_std)
 38 |     return z_mean + K.exp(z_log_var / 2) * epsilon
 39 | 
 40 | # note that "output_shape" isn't necessary with the TensorFlow backend
 41 | z = Lambda(sampling, output_shape=(latent_dim,))([z_mean, z_log_var])
 42 | 
 43 | # we instantiate these layers separately so as to reuse them later
 44 | decoder_h = Dense(intermediate_dim, activation='relu')
 45 | decoder_mean = Dense(original_dim, activation='sigmoid')
 46 | h_decoded = decoder_h(z)
 47 | x_decoded_mean = decoder_mean(h_decoded)
 48 | 
 49 | # instantiate VAE model
 50 | vae = Model(x, x_decoded_mean)
 51 | 
 52 | # Compute VAE loss
 53 | xent_loss = original_dim * metrics.binary_crossentropy(x, x_decoded_mean)
 54 | kl_loss = - 0.5 * K.sum(1 + z_log_var - K.square(z_mean) - K.exp(z_log_var), axis=-1)
 55 | vae_loss = K.mean(xent_loss + kl_loss)
 56 | 
 57 | vae.add_loss(vae_loss)
 58 | vae.compile(optimizer='rmsprop')
 59 | vae.summary()
 60 | 
 61 | 
 62 | # train the VAE on MNIST digits
 63 | (x_train, y_train), (x_test, y_test) = mnist.load_data()
 64 | 
 65 | x_train = x_train.astype('float32') / 255.
 66 | x_test = x_test.astype('float32') / 255.
 67 | x_train = x_train.reshape((len(x_train), np.prod(x_train.shape[1:])))
 68 | x_test = x_test.reshape((len(x_test), np.prod(x_test.shape[1:])))
 69 | 
 70 | vae.fit(x_train,
 71 |         shuffle=True,
 72 |         epochs=epochs,
 73 |         batch_size=batch_size,
 74 |         validation_data=(x_test, None))
 75 | 
 76 | # build a model to project inputs on the latent space
 77 | encoder = Model(x, z_mean)
 78 | 
 79 | # display a 2D plot of the digit classes in the latent space
 80 | x_test_encoded = encoder.predict(x_test, batch_size=batch_size)
 81 | plt.figure(figsize=(6, 6))
 82 | plt.scatter(x_test_encoded[:, 0], x_test_encoded[:, 1], c=y_test)
 83 | plt.colorbar()
 84 | plt.show()
 85 | 
 86 | # build a digit generator that can sample from the learned distribution
 87 | decoder_input = Input(shape=(latent_dim,))
 88 | _h_decoded = decoder_h(decoder_input)
 89 | _x_decoded_mean = decoder_mean(_h_decoded)
 90 | generator = Model(decoder_input, _x_decoded_mean)
 91 | 
 92 | # display a 2D manifold of the digits
 93 | n = 15  # figure with 15x15 digits
 94 | digit_size = 28
 95 | figure = np.zeros((digit_size * n, digit_size * n))
 96 | # linearly spaced coordinates on the unit square were transformed through the inverse CDF (ppf) of the Gaussian
 97 | # to produce values of the latent variables z, since the prior of the latent space is Gaussian
 98 | grid_x = norm.ppf(np.linspace(0.05, 0.95, n))
 99 | grid_y = norm.ppf(np.linspace(0.05, 0.95, n))
100 | 
101 | for i, yi in enumerate(grid_x):
102 |     for j, xi in enumerate(grid_y):
103 |         z_sample = np.array([[xi, yi]])
104 |         x_decoded = generator.predict(z_sample)
105 |         digit = x_decoded[0].reshape(digit_size, digit_size)
106 |         figure[i * digit_size: (i + 1) * digit_size,
107 |                j * digit_size: (j + 1) * digit_size] = digit
108 | 
109 | plt.figure(figsize=(10, 10))
110 | plt.imshow(figure, cmap='Greys_r')
111 | plt.show()
112 | 


--------------------------------------------------------------------------------