forked from kuangliu/pytorch-cifar
-
Notifications
You must be signed in to change notification settings - Fork 2
/
train.py
346 lines (309 loc) · 14.3 KB
/
train.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
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
'''Train CIFAR10 with PyTorch.'''
from __future__ import print_function
import argparse
import cPickle as pickle
import pprint as pp
import numpy as np
import time
import torch
import torch.nn as nn
import torch.optim as optim
import torch.nn.functional as F
import torch.backends.cudnn as cudnn
import torchvision.models
import os
import random
import gc
from models import *
from utils import progress_bar
from lib.SelectiveBackpropper import SelectiveBackpropper
import lib.backproppers
import lib.calculators
import lib.datasets
import lib.forwardproppers
import lib.loggers
import lib.losses
import lib.selectors
import lib.trainer
BIAS_LOG_INTERVAL = 10
start_time_seconds = time.time()
def count_tensors():
num_tensors = 0
for obj in gc.get_objects():
try:
if torch.is_tensor(obj) or (hasattr(obj, 'data') and torch.is_tensor(obj.data)):
num_tensors += 1
except:
pass
return num_tensors
def set_random_seeds(seed):
if seed:
print("Setting static random seeds to {}".format(seed))
random.seed(seed)
np.random.seed(seed)
torch.manual_seed(seed)
torch.backends.cudnn.deterministic = True
return
def set_experiment_default_args(parser):
strategy_options = ['nofilter', 'sb', 'kath', 'logbias']
calculator_options = ['relative', 'random', 'hybrid']
fp_selector_options = ['alwayson', 'stale']
# Basic training options
parser.add_argument('--lr', default=0.1, type=float, help='learning rate')
parser.add_argument('--lr-sched', default=None, help='Path to learning rate schedule')
parser.add_argument('--momentum', default=0.9, type=float, help='learning rate')
parser.add_argument('--decay', default=5e-4, type=float, help='decay')
parser.add_argument('--resume-checkpoint-file', default=None, metavar='N',
help='checkpoint to resume from')
parser.add_argument('--augment', '-a', dest='augment', action='store_true',
help='turn on data augmentation for CIFAR10')
parser.add_argument('--batch-size', type=int, default=128, metavar='N',
help='input batch size for training (default: 1)')
parser.add_argument('--forward-batch-size', type=int, default=128, metavar='N',
help='batch size for informative forward pass')
parser.add_argument('--test-batch-size', type=int, default=100, metavar='N',
help='input batch size for testing (default: 100)')
parser.add_argument('--log-interval', type=int, default=1, metavar='N',
help='how many batches to wait before logging training status')
parser.add_argument('--net', default="resnet", metavar='N',
help='which network architecture to train')
parser.add_argument('--dataset', default="cifar10", metavar='N',
help='which network architecture to train')
parser.add_argument('--datadir', default="./", metavar='N',
help='path to directory for ImageData loader')
parser.add_argument('--seed', type=int, default=None,
help='seed for randomization; None to not set seed')
parser.add_argument('--optimizer', default="sgd", metavar='N',
help='Optimizer among {sgd, adam}')
parser.add_argument('--loss-fn', default="cross", metavar='N',
help='Loss function among {cross, hinge, cross_squared, cross_custom}')
parser.add_argument('--max-num-backprops', type=int, default=float('inf'), metavar='N',
help='how many images to backprop total')
# SB options
parser.add_argument('--strategy', default='nofilter', choices=strategy_options)
parser.add_argument('--calculator', default='relative', choices=calculator_options)
parser.add_argument('--fp_selector', default='alwayson', choices=fp_selector_options)
parser.add_argument('--sb-start-epoch', type=float, default=0,
help='epoch to start selective backprop')
parser.add_argument('--prob-pow', type=float, default=1, metavar='N',
help='Power to scale probability by')
parser.add_argument('--staleness', type=int, default=2,
help='Number of epochs to use stale losses for fp_selector')
parser.add_argument('--kath-oversampling-rate', type=int, default=3, metavar='N',
help='how much to oversample by when running kath')
parser.add_argument('--std-multiplier', type=float, default=1, metavar='N',
help='stdev multiplier for forward pass prob calculator')
parser.add_argument('--sample-size', type=int, default=64, metavar='N',
help='input batch size for training (default: 1)')
parser.add_argument('--sampling-strategy', default="square", metavar='N',
help='Selective backprop sampling strategy among {nosquare, square}')
parser.add_argument('--sampling-min', type=float, default=0,
help='Minimum sampling rate for sampling strategy')
parser.add_argument('--sampling-max', type=float, default=1,
help='Maximum sampling rate for sampling strategy')
parser.add_argument('--selectivity-scalar', type=float, default=1,
help='scale the select probability')
parser.add_argument('--forwardlr', dest='forwardlr', action='store_true',
help='LR schedule based on forward passes')
# Logging and checkpointing interval
parser.add_argument('--no-logging', dest='no_logging', action='store_true',
help='turn off unnecessary logging')
parser.add_argument('--pickle-dir', default="/tmp/",
help='directory for pickles')
parser.add_argument('--pickle-prefix', default="stats",
help='file prefix for pickles')
parser.add_argument('--imageids-log-interval', type=int, default=10,
help='How often to write image ids to file (in epochs)')
parser.add_argument('--losses-log-interval', type=int, default=10,
help='How often to write losses to file (in epochs)')
parser.add_argument('--confidences-log-interval', type=int, default=10,
help='How often to write target confidences to file (in epochs)')
parser.add_argument('--checkpoint-interval', type=int, default=None, metavar='N',
help='how often to save snapshot')
parser.add_argument('--log-bias', dest='log_bias', action='store_true',
help='Log bias by epoch')
# Random features
parser.add_argument('--randomize-labels', type=float, default=None,
help='fraction of labels to randomize')
parser.add_argument('--write-images', default=False, type=bool,
help='whether or not write png images by id')
return parser
def print_config(args):
print("config sb-start-epoch {}".format(args.sb_start_epoch))
print("config lr {}".format(args.lr))
print("config lr-sched {}".format(args.lr_sched))
print("config momentum {}".format(args.momentum))
print("config decay {}".format(args.decay))
print("config batch-size {}".format(args.batch_size))
print("config net {}".format(args.net))
print("config dataset {}".format(args.dataset))
print("config seed {}".format(args.seed))
print("config optimizer {}".format(args.optimizer))
print("config loss-fn {}".format(args.loss_fn))
print("config strategy {}".format(args.strategy))
print("config calculator {}".format(args.calculator))
print("config sampling-min {}".format(args.sampling_min))
print("config sampling-max {}".format(args.sampling_max))
print("config prob_pow {}".format(args.prob_pow))
print("config forwardlr {}".format(args.forwardlr))
def test_sb(cnn, loader, epoch, sb):
cnn.eval() # Change model to 'eval' mode (BN uses moving mean/var).
correct = 0.
total = 0.
test_loss = 0.
for images, labels, ids in loader:
images = images.cuda()
labels = labels.cuda()
with torch.no_grad():
pred = cnn(images)
loss = nn.CrossEntropyLoss()(pred, labels)
test_loss += loss.item()
pred = torch.max(pred.data, 1)[1]
total += labels.size(0)
correct += (pred == labels).sum().item()
test_loss /= total
val_acc = correct / total
print('test_debug,{},{},{},{:.6f},{:.6f},{},{}'.format(
epoch,
sb.logger.global_num_backpropped,
sb.logger.global_num_skipped,
test_loss,
100.*val_acc,
sb.logger.global_num_skipped_fp,
time.time() - start_time_seconds))
cnn.train()
return 100. * val_acc
def main(args):
device = 'cuda' if torch.cuda.is_available() else 'cpu'
assert device == "cuda"
set_random_seeds(args.seed)
# Model case
print('==> Building model..')
if args.net == "resnet":
if args.dataset == "imagenet":
net = torchvision.models.__dict__["resnet18"]()
else:
net = ResNet18()
elif args.net == "vgg":
net = VGG('VGG19')
elif args.net == "preact_resnet":
net = PreActResNet18()
elif args.net == "googlenet":
net = GoogLeNet()
elif args.net == "densenet":
net = DenseNet121()
elif args.net == "resnext":
net = ResNeXt29_2x64d()
elif args.net == "mobilenet":
net = MobileNet()
elif args.net == "mobilenetv2":
net = MobileNetV2()
elif args.net == "dpn":
net = DPN92()
elif args.net == "shufflenet":
net = ShuffleNetG2()
elif args.net == "senet":
net = SENet18()
else:
net = ResNet18()
net = net.to(device)
# Device case
if device == 'cuda':
net = torch.nn.DataParallel(net)
cudnn.benchmark = True
# Dataset case
if args.dataset == "cifar10":
dataset = lib.datasets.CIFAR10(net,
args.test_batch_size,
args.augment,
#args.batch_size * 4,
None,
randomize_labels=args.randomize_labels)
elif args.dataset == "mnist":
dataset = lib.datasets.MNIST(
#10000,
None,
args.test_batch_size)
elif args.dataset == "svhn":
dataset = lib.datasets.SVHN(net,
args.test_batch_size,
#100000,
None,
args.augment)
elif args.dataset == "imagenet":
traindir = os.path.join(args.datadir, "train")
valdir = os.path.join(args.datadir, "val")
dataset = lib.datasets.ImageNet(net,
args.test_batch_size,
traindir,
valdir,
100000)
else:
print("Only cifar10, mnist, svhn and imagenet are implemented")
exit()
print(dataset.num_training_images)
print_config(args)
# Optimizer case
if args.optimizer == "sgd":
optimizer = optim.SGD(dataset.model.parameters(),
lr=args.lr,
momentum=args.momentum,
weight_decay=args.decay)
elif args.optimizer == "adam":
optimizer = optim.Adam(dataset.model.parameters(),
lr=args.lr,
weight_decay=args.decay)
# Loss function case
if args.loss_fn == "cross":
loss_fn = nn.CrossEntropyLoss
elif args.loss_fn == "cross_squared":
loss_fn = lib.losses.CrossEntropySquaredLoss
elif args.loss_fn == "cross_custom":
loss_fn = lib.losses.CrossEntropyLoss
elif args.loss_fn == "cross_regulated":
loss_fn = lib.losses.CrossEntropyRegulatedLoss
elif args.loss_fn == "cross_regulated_boosted":
loss_fn = lib.losses.CrossEntropyRegulatedBoostedLoss
elif args.loss_fn == "hinge":
loss_fn = nn.MultiMarginLoss
else:
print("Error: Loss function cannot be {}".format(args.loss_fn))
exit()
num_images_to_prime = int(args.sb_start_epoch * dataset.num_training_images)
sb = SelectiveBackpropper(net,
optimizer,
args.prob_pow,
args.batch_size,
args.lr_sched,
len(dataset.classes),
dataset.num_training_images,
args.forwardlr,
args.strategy,
args.kath_oversampling_rate,
args.calculator,
args.fp_selector,
args.staleness)
eval_every_n = args.batch_size * 10
last_global_num_backpropped = 0
epoch = 0
while True:
for dataset_split in dataset.get_dataset_splits(first_split_size=num_images_to_prime):
if not args.no_logging:
if sb.logger.global_num_backpropped - last_global_num_backpropped > eval_every_n:
test_sb(net, dataset.testloader, epoch, sb)
last_global_num_backpropped = sb.logger.global_num_backpropped
dataset_sampler = torch.utils.data.SubsetRandomSampler(dataset_split)
trainloader = torch.utils.data.DataLoader(dataset.trainset,
batch_size=args.batch_size,
sampler=dataset_sampler,
num_workers=2)
sb.trainer.train(trainloader)
sb.next_partition()
test_sb(net, dataset.testloader, epoch, sb)
sb.next_epoch()
epoch += 1
if __name__ == '__main__':
parser = argparse.ArgumentParser(description='PyTorch CIFAR10 Training')
parser = set_experiment_default_args(parser)
args = parser.parse_args()
main(args)