-
Notifications
You must be signed in to change notification settings - Fork 2
/
run.py
219 lines (192 loc) · 7.26 KB
/
run.py
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
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
'''The VAE code skeleton taken from the VAE MNIST repository.
# Reference
[1] Kingma, Diederik P., and Max Welling.
"Auto-encoding variational bayes."
https://arxiv.org/abs/1312.6114
'''
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
from keras.layers import Lambda, Input, Dense, Add, Activation
from keras.models import Model
from keras.datasets import mnist
from keras.losses import mse, binary_crossentropy
from keras.utils import plot_model
from keras.callbacks import History, ModelCheckpoint
from keras.optimizers import Adam, SGD
from keras import backend as K
from keras import objectives
from keras import initializers
import numpy as np
import matplotlib.pyplot as plt
import argparse
import os
import pandas as pd
def sampling(args):
"""Reparameterization trick by sampling fr an isotropic unit Gaussian.
# Arguments:
args (tensor): mean and log of variance of Q(z|X)
# Returns:
z (tensor): sampled latent vector
"""
z_mean, z_log_var = args
batch = K.shape(z_mean)[0]
dim = K.int_shape(z_mean)[1]
# by default, random_normal has mean=0 and std=1.0
epsilon = K.random_normal(shape=(batch, dim))
return z_mean + K.exp(0.5 * z_log_var) * epsilon
#load toy-model events
filename = 'ttb.csv'
ttb_df = pd.read_csv(filename, sep=' ', header=None)
data = ttb_df.values
data = data[:,0:26]
max = np.empty(26)
for i in range(0,data.shape[1]):
max[i] = np.max(np.abs(data[:,i]))
if np.abs(max[i]) > 0:
data[:,i] = data[:,i]/max[i]
else:
pass
trainsize = 100000
print(np.shape(data))
x_train = data[:trainsize]
x_test = data[100000:200000]
image_size = x_train.shape[1]
original_dim = image_size
x_train = np.reshape(x_train, [-1, original_dim])
x_test = np.reshape(x_test, [-1, original_dim])
x_train = x_train.astype('float32')
x_test = x_test.astype('float32')
# network parameters
input_shape = (original_dim, )
intermediate_dim = 128
encoder_dim = 128
batch_size = 1024
latent_dim = 20
epochs = 240
eluvar = np.sqrt(1.55/intermediate_dim)
# VAE model = encoder + decoder
# build encoder model
inputs = Input(shape=input_shape, name='encoder_input')
x1 = Dense(encoder_dim, activation='elu', kernel_initializer=initializers.random_normal(mean=0.0, stddev=eluvar))(inputs)
x2 = Dense(encoder_dim, kernel_initializer=initializers.random_normal(mean=0.0, stddev=eluvar))(x1)
x2 = Activation('elu')(x2)
x3 = Dense(encoder_dim, kernel_initializer=initializers.random_normal(mean=0.0, stddev=eluvar))(x2)
sc1 = Add()([x1,x3])
x3 = Activation('elu')(sc1)
x4 = Dense(encoder_dim, kernel_initializer=initializers.random_normal(mean=0.0, stddev=eluvar))(x3)
sc2 = Add()([x2,x4])
x4 = Activation('elu')(sc2)
x5 = Dense(encoder_dim, kernel_initializer=initializers.random_normal(mean=0.0, stddev=eluvar))(x4)
sc3 = Add()([x3,x5])
x5 = Activation('elu')(sc3)
x6 = Dense(encoder_dim, kernel_initializer=initializers.random_normal(mean=0.0, stddev=eluvar))(x5)
sc4 = Add()([x4,x6])
x6 = Activation('elu')(sc4)
z_mean = Dense(latent_dim, name='z_mean')(x6)
z_log_var = Dense(latent_dim, name='z_log_var')(x6)
# use reparameterization trick to push the sampling out as input
# note that "output_shape" isn't necessary with the TensorFlow backend
z = Lambda(sampling, output_shape=(latent_dim,), name='z')([z_mean, z_log_var])
# instantiate encoder model
encoder = Model(inputs, [z_mean, z_log_var, z], name='encoder')
encoder.summary()
# build decoder model
latent_inputs = Input(shape=(latent_dim,), name='z_sampling')
x1 = Dense(intermediate_dim, activation='elu', kernel_initializer=initializers.random_normal(mean=0.0, stddev=eluvar))(latent_inputs)
x2 = Dense(encoder_dim, kernel_initializer=initializers.random_normal(mean=0.0, stddev=eluvar))(x1)
x2 = Activation('elu')(x2)
x3 = Dense(encoder_dim, kernel_initializer=initializers.random_normal(mean=0.0, stddev=eluvar))(x2)
sc1 = Add()([x1,x3])
x3 = Activation('elu')(sc1)
x4 = Dense(encoder_dim, kernel_initializer=initializers.random_normal(mean=0.0, stddev=eluvar))(x3)
sc2 = Add()([x2,x4])
x4 = Activation('elu')(sc2)
x5 = Dense(encoder_dim, kernel_initializer=initializers.random_normal(mean=0.0, stddev=eluvar))(x4)
sc3 = Add()([x3,x5])
x5 = Activation('elu')(sc3)
x6 = Dense(encoder_dim, kernel_initializer=initializers.random_normal(mean=0.0, stddev=eluvar))(x5)
sc4 = Add()([x4,x6])
x6 = Activation('elu')(sc4)
outputs = Dense(original_dim, activation='tanh')(x6)
# instantiate decoder model
decoder = Model(latent_inputs, outputs, name='decoder')
decoder.summary()
# instantiate VAE model
outputs = decoder(encoder(inputs)[2])
vae = Model(inputs, outputs, name='ttbar_vae')
if __name__ == '__main__':
parser = argparse.ArgumentParser()
help_ = "Load h5 model trained weights"
parser.add_argument("-w", "--weights", help=help_)
help_ = "Use mse loss instead of binary cross entropy (default)"
parser.add_argument("-m",
"--mse",
help=help_, action='store_true')
args = parser.parse_args()
models = (encoder, decoder)
data = (x_test, x_test)
def vae_loss(x, x_decoded_mean):
mse_loss = objectives.mse(x, x_decoded_mean)
kl_loss = 1 + z_log_var - K.square(z_mean) - K.exp(z_log_var)
kl_loss = K.sum(kl_loss, axis=-1)
kl_loss *= -0.5
beta=10**(-6)
loss = K.mean((1-beta)*mse_loss + beta*kl_loss)
return loss
learnrate = 0.001
iterations = 7
lr_limit = 0.001/(2**iterations)
history = History()
k=0
checkpointer = ModelCheckpoint(filepath='ttbar_20d_e-6.hdf5', verbose=1, save_best_only=True)
opt = Adam(lr=learnrate, beta_1=0.9, beta_2=0.999, epsilon=1e-08, decay=0.0)
vae.compile(optimizer=opt, loss=vae_loss)
vae.summary()
k=0
if args.weights:
vae.load_weights(args.weights)
else:
while learnrate > lr_limit:
if k < 4:
opt = Adam(lr=learnrate, beta_1=0.9, beta_2=0.999, epsilon=1e-08, decay=0.0)
else:
opt = SGD(lr=learnrate, decay=1e-6, momentum=0.9, nesterov=True)
epochs=120
vae.compile(loss=vae_loss, optimizer=opt, metrics=['mse'])
vae.fit(x_train, x_train,
epochs=epochs,
batch_size=batch_size,
validation_data=(x_test, x_test),
callbacks = [checkpointer, history])
vae.load_weights('ttbar_20d_e-6.hdf5')
learnrate /= 2
k=k+1
# train the autoencoder
vae.save_weights('ttbar_20d_e-6.h5')
latent_mean = encoder.predict(x_train)[0]
latent_logvar = encoder.predict(x_train)[1]
latent_var = np.exp(latent_logvar)
latent_std = np.sqrt(latent_var)
np.savetxt('latent_mean_20d_e-6.csv', latent_mean)
np.savetxt('latent_std_20d_e-6.csv', latent_std)
filename = 'latent_mean_20d_e-6.csv'
means_df = pd.read_csv(filename, sep=' ', header=None)
mean = means_df.values
filename = 'latent_std_20d_e-6.csv'
stds_df = pd.read_csv(filename, sep=' ', header=None)
std = stds_df.values
lat_dim = 20
b = 'e-6'
z_samples = np.empty([1200000,lat_dim])
l=0
#sampling from the new prior with gamma=0.05
l=0
for i in range(0,1200000):
for j in range(0,lat_dim):
z_samples[l,j] = np.random.normal(mean[i%100000,j], 0.05+std[i%100000,j])
l=l+1
new_events = decoder.predict(z_samples)
for i in range(0,new_events.shape[1]):
new_events[:,i]=new_events[:,i]*max[i]
np.savetxt('B-VAE_events.csv', new_events)