QuickExperiments/tf_vae_mnist.py at master · MoMe36/QuickExperiments · GitHub

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
import numpy as np
import tensorflow as tf
import pickle
import matplotlib.pyplot as plt

from PIL import Image


class Loader():

	def __init__(self, path = '', max_items = 60000):

		self.path = path
		self.max_count = max_items
		self.counter = 0

	def sample(self,batch_size = 32, use_cuda = False):

		data = np.zeros((batch_size, 784))
		ind = np.random.randint(0,self.max_count, (batch_size))
		for i in range(batch_size):
			name = self.path+'{}'.format(ind[i])
			image = pickle.load(open(name,'rb'))
			data[i,:] = image.reshape(784)
			self.counter = (self.counter+1)%self.max_count

		data /= 255.

		return data

def make_grid(x, rows = 4, decal = 2):

	# x is a (batch_size, nb_channels, height, width) nd-array

	col = int(x.shape[0]/rows)
	image = np.zeros(((x.shape[2] + decal)*rows, (decal + x.shape[3])*col, 3))
	ligne = 0
	column = 0
	for i in range(x.shape[0]):
		current = x[i,:,:,:]
		current = np.transpose(current, [1,2,0])


		image[decal + ligne*(x.shape[2]):(ligne+1)*(x.shape[2]) + decal, decal + column*(x.shape[3]):(column+1)*(x.shape[3]) + decal,:] = current

		column = (column + 1)%col
		if(column == 0):
			ligne += 1

	# input(image.shape)
	return image

dense = tf.layers.dense
# relu = tf.nn.relu(alpha = 0.2)

code_size = 20
x = tf.placeholder(tf.float32, shape = [None, 784], name = 'Input_x')

with tf.variable_scope('Encoder'):
	e1 = dense(x, 512, activation = tf.nn.relu, name = 'encoding_1')
	e1 = dense(e1, 384, activation = tf.nn.relu, name = 'encoding_2')
	e1= dense(e1, 256, activation = tf.nn.relu, name = 'encoding_3')

with tf.variable_scope('Reparametrization'):
	z_means = dense(e1, code_size, name = 'means')
	z_stds = dense(e1, code_size, name = 'logvar')

with tf.variable_scope('Code'):
	z = z_means + tf.random_normal(tf.shape(z_stds))*tf.exp(z_stds)


with tf.variable_scope('Decoder'):
	d1 = dense(z, 256, activation = tf.nn.relu, name = 'decoding_1')
	d1 = dense(d1, 384, activation = tf.nn.relu, name = 'decoding_2')
	d2 = dense(d1, 512, activation = tf.nn.relu, name = 'decoding_3')
	out = dense(d2, 784, name = 'reconstruction', activation = tf.nn.sigmoid)


with tf.variable_scope('Losses'):

	kl_loss = -0.5*tf.reduce_sum(1. + z_stds - tf.pow(z_means,2) - tf.exp(z_stds), axis = 1)
	# recon_loss = tf.losses.mean_squared_error(x, out, reduction_indices = 1)
	recon_loss = tf.reduce_sum(tf.square(out - x), axis = 1)

with tf.variable_scope('Training'):

	full_loss = tf.reduce_mean(kl_loss + recon_loss)
	update_vae = tf.train.AdamOptimizer(5e-4).minimize(full_loss)


epochs = 2500
loader = Loader('/home/mehdi/Codes/MNIST/',60000)

f, ax = plt.subplots(1,2)

with tf.Session() as sess:

	sess.run(tf.global_variables_initializer())

	mean_loss = 0.
	for epoch in range(1,epochs+1):

		data_x = loader.sample(64)

		loss, _ = sess.run([full_loss, update_vae], feed_dict = {x:data_x})
		mean_loss += np.mean(loss)

		if epoch% 100 == 0:
			# input(loss)
			print('Epoch: {} | Loss: {:.6f}'.format(epoch, mean_loss/100.))
			mean_loss = 0.

		if epoch % 100 == 0:

			data_x = loader.sample(32)
			result = sess.run([out], feed_dict = {z:np.random.normal(0.,1., (32,code_size))})[0]
			recons = sess.run([out], feed_dict = {x:data_x})[0]

			for a in ax:
				a.clear()

			ax[0].imshow(make_grid(recons.reshape(32,1,28,28)))
			ax[1].imshow(make_grid(result.reshape(32,1,28,28)))

			ax[0].set_title('Recons')
			ax[1].set_title('Prods')

			plt.pause(0.1)

	plt.show()