-
Notifications
You must be signed in to change notification settings - Fork 8
/
main.py
262 lines (209 loc) · 11.4 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
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
from utils.utils import *
from utils.logging_utils import Logger
import utils.datasets
from nn_utils.NNTrainer import NNTrainer
from probes_lib.top import *
import torch.backends.cudnn as cudnn
import torch.nn as nn
import torch
import argparse
import models
from time import time
import pickle
import os
import socket
import sys
from ast import literal_eval
from nn_utils.init_utils import init_model
import optimizers_lib
###########################################################################
# Script's arguments
###########################################################################
archs_names = sorted(name for name in models.__dict__
if name.islower() and not name.startswith("__")
and callable(models.__dict__[name]))
optimizers_names = sorted(name for name in optimizers_lib.__dict__
if name.islower() and not name.startswith("__")
and callable(optimizers_lib.__dict__[name]))
datasets_names = sorted(name for name in utils.datasets.__dict__
if name.islower() and name.startswith("ds_")
and callable(utils.datasets.__dict__[name]))
parser = argparse.ArgumentParser(description='Train and record statistics of a Neural Network with BGD')
parser.add_argument('--dataset', default="ds_mnist", type=str, choices=datasets_names,
help='The name of the dataset to train. [Default: ds_mnist]')
parser.add_argument('--nn_arch', type=str, required=True, choices=archs_names,
help='Neural network architecture')
parser.add_argument('--logname', type=str, required=True,
help='Prefix of logfile name')
parser.add_argument('--results_dir', type=str, default="TMP",
help='Results dir name')
parser.add_argument('--seed', type=int,
help='Seed for randomization.')
parser.add_argument('--num_workers', default=8, type=int,
help='Num of workers for data loader')
parser.add_argument('--num_epochs', default=400, type=int,
help='Maximum number of training epochs. [Default: 400]')
parser.add_argument('--batch_size', default=128, type=int,
help='Batch size [Default: 128]')
parser.add_argument('--pruning_percents', default=[], type=int, nargs='*',
help='A list of percents to check pruning with [Default: []]')
# BGD
parser.add_argument('--train_mc_iters', default=10, type=int,
help='Number of MonteCarlo samples during training(default 10)')
parser.add_argument('--std_init', default=5e-2, type=float,
help='STD init value (default 5e-2)')
parser.add_argument('--mean_eta', default=1, type=float,
help='Eta for mean step (default 1)')
parser.add_argument('--permanent_prune_on_epoch', default=-1, type=int,
help='Permanent prune on epoch')
parser.add_argument('--permanent_prune_on_epoch_percent', default=90, type=float,
help='Permanent prune percent of weights')
parser.add_argument('--test_freq', default=1, type=int,
help='Run test set every test_freq epochs [default: 1]')
parser.add_argument('--contpermuted_beta', default=3, type=int,
help='Beta value for continuous permuted mnist. [default: 3]')
parser.add_argument('--optimizer', type=str, default="bgd", choices=optimizers_names,
help='Optimizer.')
parser.add_argument('--optimizer_params', default="{}", type=str, nargs='*',
help='Optimizer parameters')
parser.add_argument('--inference_mc', default=False, action='store_true',
help='Use MonteCarlo samples as inference method.')
parser.add_argument('--inference_map', default=False, action='store_true',
help='Use MAP as inference method.')
parser.add_argument('--inference_committee', default=False, action='store_true',
help='Use committee as inference method.')
parser.add_argument('--inference_aggsoftmax', default=False, action='store_true',
help='Use aggsoftmax as inference method.')
parser.add_argument('--inference_initstd', default=False, action='store_true',
help='Use initstd as inference method.')
parser.add_argument('--committee_size', default=0, type=int,
help='Size of committee (when using committee inference)')
parser.add_argument('--test_mc_iters', default=0, type=int,
help='Number of MC iters when testing (when using MC inference)')
parser.add_argument('--init_params',
default=["{\"bias_type\":", "\"xavier\",", "\"conv_type\":", "\"xavier\",",
"\"bn_init\":", "\"01\"}"], type=str, nargs='*', help='Initialization parameters')
parser.add_argument('--desc', default="", type=str, nargs='*',
help='Desc file content')
parser.add_argument('--bw_to_rgb', default=False, action='store_true',
help='Convert black&white (e.g. MNIST) images to RGB format')
parser.add_argument('--permuted_offset', default=False, action='store_true',
help='Use offset for permuted mnist experiment')
parser.add_argument('--labels_trick', default=False, action='store_true',
help='Use labels trick (train only the heads of current batch labels)')
parser.add_argument('--num_of_permutations', default=9, type=int,
help='Number of permutations (in addition to the original MNIST) '
'when using Permuted MNIST dataset [default: 9]')
parser.add_argument('--iterations_per_virtual_epc', default=468, type=int,
help='When using continuous dataset, number of iterations per epoch (in continuous mode, '
'epoch is not defined)')
parser.add_argument('--separate_labels_space', default=False, action='store_true',
help='Use separate label space for each task')
parser.add_argument('--permute_seed', type=int,
help='Seed for creating the permutations.')
args = parser.parse_args()
###########################################################################
# Verify arguments
###########################################################################
inference_methods = set()
if args.inference_committee:
assert (args.committee_size > 0)
inference_methods.add("committee")
if args.inference_mc:
assert (args.test_mc_iters > 0)
inference_methods.add("test_mc")
if args.inference_map:
inference_methods.add("map")
if args.inference_aggsoftmax:
inference_methods.add("agg_softmax")
if args.inference_initstd:
inference_methods.add("init_std")
assert(len(inference_methods) > 0)
if args.optimizer != "bgd":
assert args.train_mc_iters == 1, "Monte Carlo iterations are for BGD optimizer only"
assert len(inference_methods) == 1 and "map" in inference_methods, "When not using BGD, must use MAP for inference"
###########################################################################
# CUDA and seeds
###########################################################################
# Create permuations
all_permutation = []
if not args.permute_seed:
args.permute_seed = int(time()*10000) % (2**31)
set_seed(args.permute_seed, fully_deterministic=False)
for p_idx in range(args.num_of_permutations):
input_size = 32 * 32
if "padded" not in args.dataset:
input_size = 28 * 28
permutation = list(range(input_size))
random.shuffle(permutation)
all_permutation.append(permutation)
# Set seed
if not args.seed:
args.seed = int(time()*10000) % (2**31)
set_seed(args.seed, fully_deterministic=False)
if torch.cuda.is_available():
cudnn.benchmark = True
###########################################################################
# Logging
###########################################################################
# Create logger
save_path = os.path.join("./logs", str(args.results_dir) + "/")
if not os.path.exists(save_path):
os.makedirs(save_path)
logger = Logger(True, save_path + args.logname, True, True)
logger.info("Script args: " + str(args))
if args.desc != "":
logger.create_desc_file(" ".join(args.desc))
else:
logger.create_desc_file(str(args))
logger.info("Computer name: " + str(socket.gethostname()) + " with pytorch version: " + str(torch.__version__))
lastlogs_logger = Logger(add_timestamp=False, logfile_name="last_logs.txt", logfile_name_time_suffix=False,
print_to_screen=False)
lastlogs_logger.info(logger.get_log_basename() + " ")
lastlogs_logger = None
###########################################################################
# Model and training
###########################################################################
# Dataset
train_loaders, test_loaders = utils.datasets.__dict__[args.dataset](batch_size=args.batch_size,
num_workers=args.num_workers,
permutations=all_permutation,
separate_labels_space=args.separate_labels_space,
num_epochs=args.num_epochs,
iterations_per_virtual_epc=
args.iterations_per_virtual_epc,
contpermuted_beta=args.contpermuted_beta,
logger=logger)
# Probes manager
probes_manager = ProbesManager()
# Model
model = models.__dict__[args.nn_arch](probes_manager=probes_manager)
if torch.cuda.is_available():
model = torch.nn.DataParallel(model).cuda()
logger.info("Transformed model to CUDA")
criterion = nn.CrossEntropyLoss()
init_params = {"logger": logger}
if args.init_params != "":
init_params = dict(init_params, **literal_eval(" ".join(args.init_params)))
init_model(get_model(model), **init_params)
# optimizer model
optimizer_model = optimizers_lib.__dict__[args.optimizer]
optimizer_params = dict({"logger": logger,
"mean_eta": args.mean_eta,
"std_init": args.std_init,
"mc_iters": args.train_mc_iters}, **literal_eval(" ".join(args.optimizer_params)))
optimizer = optimizer_model(model, probes_manager=probes_manager, **optimizer_params)
trainer = NNTrainer(train_loader=train_loaders, test_loader=test_loaders,
criterion=criterion, net=model, logger=logger, probes_manager=probes_manager,
std_init=args.std_init, mean_eta=args.mean_eta, train_mc_iters=args.train_mc_iters,
test_mc_iters=args.test_mc_iters, committee_size=args.committee_size, batch_size=args.batch_size,
inference_methods=inference_methods,
pruning_percents=args.pruning_percents,
bw_to_rgb=args.bw_to_rgb,
labels_trick=args.labels_trick,
test_freq=args.test_freq,
optimizer=optimizer)
trainer.train_epochs(verbose_freq=100, max_epoch=args.num_epochs,
permanent_prune_on_epoch=args.permanent_prune_on_epoch,
permanent_prune_on_epoch_percent=args.permanent_prune_on_epoch_percent)
print("Done")