-
Notifications
You must be signed in to change notification settings - Fork 13
/
train_MixModel_VAE-2StochLayers.py
192 lines (145 loc) · 8.13 KB
/
train_MixModel_VAE-2StochLayers.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
import os
from os.path import join as pjoin
import h5py
import cPickle as cp
import numpy as np
import tensorflow as tf
from tensorflow.examples.tutorials.mnist import input_data
from models.gaussMMVAE_collapsed import DLGMM
from utils.sampling_utils import *
try:
import PIL.Image as Image
except ImportError:
import Image
# command line arguments
flags = tf.flags
flags.DEFINE_integer("batchSize", 100, "batch size.")
flags.DEFINE_integer("nEpochs", 1000, "number of epochs to train.")
flags.DEFINE_float("adamLr", 1e-4, "AdaM learning rate.")
flags.DEFINE_integer("hidden_size", 500, "number of hidden units in en/decoder.")
flags.DEFINE_integer("latent_size", 20, "dimensionality of latent variables.")
flags.DEFINE_integer("K", 3, "number of components in mixture model.")
flags.DEFINE_string("experimentDir", "MNIST/", "directory to save training artifacts.")
inArgs = flags.FLAGS
def get_file_name(expDir, vaeParams, trainParams):
# concat hyperparameters into file name
output_file_base_name = '_'+''.join('{}_{}_'.format(key, val) for key, val in sorted(vaeParams.items()) if key not in ['prior', 'input_d'])
output_file_base_name += ''.join('{}_{}_'.format(key, "-".join([str(x) for x in vaeParams['prior'][key]])) for key in sorted(['mu', 'sigma']))
output_file_base_name += 'dirichlet_alpha_'+str(vaeParams['prior']['dirichlet_alpha'])
output_file_base_name += '_adamLR_'+str(trainParams['adamLr'])
# check if results file already exists, if so, append a number
results_file_name = pjoin(expDir, "train_logs/dlgmm_trainResults"+output_file_base_name+".txt")
file_exists_counter = 0
while os.path.isfile(results_file_name):
file_exists_counter += 1
results_file_name = pjoin(expDir, "train_logs/dlgmm_trainResults"+output_file_base_name+"_"+str(file_exists_counter)+".txt")
if file_exists_counter > 0:
output_file_base_name += "_"+str(file_exists_counter)
return output_file_base_name
### Training function
def trainDLGMM(data, vae_hyperParams, hyperParams, param_save_path, logFile=None):
N_train, d = data['train'].shape
N_valid, d = data['valid'].shape
nTrainBatches = N_train/hyperParams['batchSize']
nValidBatches = N_valid/hyperParams['batchSize']
vae_hyperParams['batchSize'] = hyperParams['batchSize']
# init Mix Density VAE
model = DLGMM(vae_hyperParams)
# get training op
optimizer = tf.train.AdamOptimizer(hyperParams['adamLr']).minimize(-model.elbo_obj)
# get op to save the model
persister = tf.train.Saver()
with tf.Session(config=hyperParams['tf_config']) as s:
s.run(tf.initialize_all_variables())
# for early stopping
best_elbo = -10000000.
best_epoch = 0
for epoch_idx in xrange(hyperParams['nEpochs']):
# training
train_elbo = 0.
for batch_idx in xrange(nTrainBatches):
x = data['train'][batch_idx*hyperParams['batchSize']:(batch_idx+1)*hyperParams['batchSize'],:]
_, elbo_val = s.run([optimizer, model.elbo_obj], {model.X: x})
train_elbo += elbo_val/nTrainBatches
# validation
valid_elbo = 0.
for batch_idx in xrange(nValidBatches):
x = data['valid'][batch_idx*hyperParams['batchSize']:(batch_idx+1)*hyperParams['batchSize'],:]
valid_elbo += s.run(model.elbo_obj, {model.X: x})/nValidBatches
# check for ELBO improvement
star_printer = ""
if valid_elbo > best_elbo:
best_elbo = valid_elbo
best_epoch = epoch_idx
star_printer = "***"
# save the parameters
persister.save(s, param_save_path)
# log training progress
logging_str = "Epoch %d. Train ELBO: %.3f, Validation ELBO: %.3f %s" %(epoch_idx+1, train_elbo, valid_elbo, star_printer)
print logging_str
if logFile:
logFile.write(logging_str + "\n")
logFile.flush()
# check for convergence
if epoch_idx - best_epoch > hyperParams['lookahead_epochs'] or np.isnan(train_elbo): break
return model
### Marginal Likelihood Calculation
def calc_margLikelihood(data, model, param_file_path, vae_hyperParams, nSamples=50):
N,d = data.shape
# get op to load the model
persister = tf.train.Saver()
with tf.Session() as s:
persister.restore(s, param_file_path)
sample_collector = []
for s_idx in xrange(nSamples):
samples = s.run(model.get_log_margLL(N), {model.X: data})
if not np.isnan(samples.mean()) and not np.isinf(samples.mean()):
sample_collector.append(samples)
if len(sample_collector) < 1:
print "\tMARG LIKELIHOOD CALC: No valid samples were collected!"
return np.nan
all_samples = np.hstack(sample_collector)
m = np.amax(all_samples, axis=1)
mLL = m + np.log(np.mean(np.exp( all_samples - m[np.newaxis].T ), axis=1))
return mLL.mean()
### Sample Images
def sample_from_model(model, param_file_path, vae_hyperParams, image_file_path, nImages=100):
# get op to load the model
persister = tf.train.Saver()
with tf.Session() as s:
persister.restore(s, param_file_path)
sample_list = s.run(model.get_samples(nImages))
for i, samples in enumerate(sample_list):
image = Image.fromarray(tile_raster_images(X=samples, img_shape=(28, 28), tile_shape=(int(np.sqrt(nImages)), int(np.sqrt(nImages))), tile_spacing=(1, 1)))
image.save(image_file_path+"_component"+str(i)+".png")
if __name__ == "__main__":
# load MNIST
f = h5py.File('./MNIST/data/binarized_mnist.h5')
mnist = {'train':np.copy(f['train']), 'valid':np.copy(f['valid']), 'test':np.copy(f['test'])}
np.random.shuffle(mnist['train'])
# set architecture params
vae_hyperParams = {'input_d':mnist['train'].shape[1], 'hidden_d':inArgs.hidden_size, 'latent_d':inArgs.latent_size, 'K':inArgs.K, \
'prior':{'dirichlet_alpha':1., 'mu':[0.]*inArgs.K, 'sigma':[1.]*inArgs.K}}
#'prior':{'dirichlet_alpha':1., 'mu':[-.5, -.25, 0., .25, .5], 'sigma':[1.]*inArgs.K}}
assert len(vae_hyperParams['prior']['mu']) == len(vae_hyperParams['prior']['sigma']) == vae_hyperParams['K']
# set training hyperparameters
train_hyperParams = {'adamLr':inArgs.adamLr, 'nEpochs':inArgs.nEpochs, 'batchSize':inArgs.batchSize, 'lookahead_epochs':25, \
'tf_config': tf.ConfigProto(gpu_options = tf.GPUOptions(per_process_gpu_memory_fraction=0.5), log_device_placement=False)}
# setup files to write results and save parameters
outfile_base_name = get_file_name(inArgs.experimentDir, vae_hyperParams, train_hyperParams)
logging_file = open(inArgs.experimentDir+"train_logs/dlgmm_trainResults"+outfile_base_name+".txt", 'w')
param_file_name = inArgs.experimentDir+"params/dlgmm_params"+outfile_base_name+".ckpt"
# train
print "Training model..."
model = trainDLGMM(mnist, vae_hyperParams, train_hyperParams, param_file_name, logging_file)
# evaluate marginal likelihood
print "Calculating the marginal likelihood..."
margll_valid = calc_margLikelihood(mnist['valid'], model, param_file_name, vae_hyperParams)
margll_test = calc_margLikelihood(mnist['test'], model, param_file_name, vae_hyperParams)
logging_str = "\n\nValidation Marginal Likelihood: %.3f, Test Marginal Likelihood: %.3f" %(margll_valid, margll_test)
print logging_str
logging_file.write(logging_str+"\n")
logging_file.close()
# draw some samples
#print "Drawing samples..."
#sample_from_model(model, param_file_name, vae_hyperParams, inArgs.experimentDir+'samples/dlgmm_samples'+outfile_base_name)