-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathmain.py
228 lines (211 loc) · 9.27 KB
/
main.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
220
221
222
import numpy as np
import tensorflow as tf
import argparse
from util import *
import models
import models_vae
def parse_args():
parser = argparse.ArgumentParser(
description='VI and HMC')
parser.add_argument('--dataset',
help='benchmark dataset to use (funnel / survey / mnist / minst_dyn / cifar10).',
default='mnist')
parser.add_argument('--method',
help='method, please choose from: {vi_klqp, vi_klpq, hmc, vae, vae_mcmc}.',
default='vi_klpq')
parser.add_argument('--v_fam',
help='variational family, please choose from: {gaussian, flow, iaf} (the last two are the same here).',
default='gaussian')
parser.add_argument('--space',
help='for HSC (VI with KLpq) and HMC only, please choose from: {eps, theta}.',
default='eps')
parser.add_argument('--vae_num_flow',
help='Number of flows to stack in VAE. For no flow (i.e. Gaussian), type 0.',
type=int,
default=0)
parser.add_argument('--epochs',
help='VI total training epochs.',
type=int,
default=100)
parser.add_argument('--lr',
help='learning rate (of the initial epoch).',
type=float,
default=0.1)
parser.add_argument('--decay_rate',
help='lr decay rate in inverse time decay.',
type=float,
default=0.001)
parser.add_argument('--batch_size',
help='batch size.',
type=int,
default=100)
parser.add_argument('--latent_dim',
help='VAE latent dimension.',
type=int,
default=2)
parser.add_argument('--architecture',
help='Architecture for encoder and decoder.',
default='cnn')
parser.add_argument('--num_samp',
help='number of samples in VI methods.',
type=int,
default=1)
parser.add_argument('--likelihood_sigma',
help='Sigma when VAE likelihood is diagonal Gaussian.',
type=float,
default=1.)
parser.add_argument('--iters',
help='HMC iterations.',
type=int,
default=1000)
parser.add_argument('--chains',
help='HMC chains.',
type=int,
default=1)
parser.add_argument('--hmc_e',
help='HMC step size.',
type=float,
default=0.25)
parser.add_argument('--hmc_L',
help='HMC number of leapfrog steps.',
type=int,
default=4)
parser.add_argument('--hmc_L_cap',
help='Cap on HMC number of leapfrog steps.',
type=int,
default=4)
parser.add_argument('--q_factor',
help='Factor to weight q - p during training.',
type=float,
default=1.)
parser.add_argument('--q_not_train',
help='Do not train q for __ epochs.',
type=int,
default=0)
parser.add_argument('--target_accept',
help='Target acceptance rate.',
type=float,
default=0.67)
parser.add_argument('--cis',
help='Number of samples in conditional importance sampling MSC. If 0, no CIS is used.',
type=int,
default=0)
parser.add_argument('--hmc_e_differs',
help='Training like Hoffman 2017, where each datapoint has its own step size, if True.',
default=False,
type=lambda x: (str(x).lower() == 'true'))
parser.add_argument('--shear',
default=True,
type=lambda x: (str(x).lower() == 'true'))
parser.add_argument('--reinitialize_from_q',
default=False,
type=lambda x: (str(x).lower() == 'true'))
parser.add_argument('--warm_up',
default=True,
type=lambda x: (str(x).lower() == 'true'))
parser.add_argument('--train_encoder',
default=True,
type=lambda x: (str(x).lower() == 'true'))
parser.add_argument('--natural_gradient',
default=False,
type=lambda x: (str(x).lower() == 'true'))
parser.add_argument('--stop_idx',
help='Index after which training data is not used.',
type=int,
default=1e8)
parser.add_argument('--default_path',
help='Redefine default_path to save results (if omitted, there will already be a default path).',
default='na')
parser.add_argument('--load_path',
help='Define load_path if you want to load a trained model.',
default='na')
parser.add_argument('--load_epoch',
help='Which epoch do you want the loaded model to start trainig at.',
type=int,
default=1)
parser.add_argument('--random_seed',
help='Random seed.',
type=int,
default=0)
args = parser.parse_args()
return args
args = parse_args()
tf.random.set_seed(args.random_seed)
if args.default_path.lower() == 'na':
path = None
else:
path = args.default_path
if args.load_path.lower() == 'na':
load_path = None
else:
load_path = args.load_path
if args.method.lower() == 'vi_klqp':
model = models.VI_KLqp(
dataset=args.dataset,
v_fam=args.v_fam.lower(),
num_samp=args.num_samp)
model.train(epochs=args.epochs, lr=args.lr, decay_rate=args.decay_rate, save=True, path=path)
if args.method.lower() == 'vi_klpq':
model = models.VI_KLpq(
dataset=args.dataset,
v_fam=args.v_fam.lower(),
space=args.space.lower(),
num_samp=args.num_samp,
chains=args.chains,
hmc_e=args.hmc_e,
hmc_L=args.hmc_L)
model.train(epochs=args.epochs, lr=args.lr, decay_rate=args.decay_rate, natural_gradient=args.natural_gradient,
save=True, path=path, load_path=load_path, load_epoch=args.load_epoch)
if args.method.lower() == 'hmc':
model = models.HMC(
space=args.space.lower(),
iters=args.iters,
chains=args.chains,
hmc_e=args.hmc_e,
hmc_L=args.hmc_L)
model.run(load_path='results/checkpoints/iaf_qp', save=True, path=path)
if args.method.lower() == 'vae' or args.method.lower() == 'vae_mcmc':
batch_size = args.batch_size
train_dataset, test_dataset, train_size, test_size = load_image_data(args.dataset, batch_size)
random_vector_for_generation = np.genfromtxt(
'data/vae_random_vector_' + str(args.latent_dim) + '.csv').astype('float32')
for test_batch in test_dataset.take(1):
test_sample = test_batch[0:16, :, :, :]
if args.method.lower() == 'vae':
model = models_vae.VAE(
args.latent_dim,
num_flow=args.vae_num_flow,
batch_size=batch_size,
K=args.num_samp,
dataset=args.dataset.lower(),
architecture=args.architecture.lower(),
likelihood_sigma=args.likelihood_sigma)
model.train(train_dataset, test_dataset, epochs=args.epochs, lr=args.lr,
load_path=load_path, load_epoch=args.load_epoch,
test_sample=test_sample, random_vector_for_generation=random_vector_for_generation, generation=True)
if args.method.lower() == 'vae_mcmc':
model = models_vae.VAE_HSC(
args.latent_dim,
num_flow=args.vae_num_flow,
space=args.space.lower(),
cis=args.cis,
dataset=args.dataset.lower(),
architecture=args.architecture.lower(),
likelihood_sigma=args.likelihood_sigma,
num_samp=args.num_samp,
chains=args.chains,
hmc_e=args.hmc_e,
hmc_L=args.hmc_L,
hmc_L_cap=args.hmc_L_cap,
q_factor=args.q_factor,
target_accept=args.target_accept,
batch_size=batch_size,
train_size=train_size,
reinitialize_from_q=args.reinitialize_from_q,
hmc_e_differs=args.hmc_e_differs,
shear=args.shear)
model.train(train_dataset, test_dataset,
epochs=args.epochs, lr=args.lr, stop_idx=args.stop_idx,
warm_up=args.warm_up, q_not_train=args.q_not_train, train_encoder=args.train_encoder,
test_sample=test_sample, random_vector_for_generation=random_vector_for_generation, generation=True,
load_path=load_path, load_epoch=args.load_epoch)