-
Notifications
You must be signed in to change notification settings - Fork 0
/
main.py
442 lines (336 loc) · 18.5 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
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
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
import warnings
# warnings.filterwarnings(action='ignore', category=DeprecationWarning)
warnings.filterwarnings(action='ignore')
import torch
import torch.nn as nn
import torch.optim as optim
import torchvision
import torchvision.transforms as transforms
import os
import sys
import argparse
import copy
import json
import datetime
import logging
logger = logging.getLogger()
import numpy as np
np.seterr(all="ignore")
from PIL import ImageFile
ImageFile.LOAD_TRUNCATED_IMAGES = True
import random
torch.manual_seed(0)
random.seed(0)
np.random.seed(0)
start_time = datetime.datetime.utcnow().strftime("%Y-%m-%dT%H:%M:%S.%f%Z")
parser = argparse.ArgumentParser(description='PyTorch training/testing')
parser.add_argument('--exp_desc', default='', type=str, help='experriment description')
parser.add_argument('--setting_dir', type=str, help='path of setting JSON file')
parser.add_argument('--dataset', default='cifar10', type=str, help='decide which dataset to run on.')
parser.add_argument('--dataset_dir', default='../data', type=str, help='folder path of dataset.')
parser.add_argument('--model_dir', default='', type=str, help='path of pretrain model')
parser.add_argument('--model_state_dict_dir', default='', type=str, help='path of pretrain model')
parser.add_argument('--snapshot_folder_dir', default='', type=str, help='path of model training snapshots')
parser.add_argument('--output_folder_dir', default='', type=str, help='path of output model')
parser.add_argument('--multi_prune_flag', default=False, type=bool, help='prune all conv layes or just 3x3, False means just prune 3x3')
parser.add_argument('--resnet_type', default='basicblock', type=str, help='ResNet type, basicblock or bottleneck.')
parser.add_argument('--zero_mask', default=False, type=bool, help='zero mask pruned network')
parser.add_argument('--baseline', default=False, type=bool, help='test baseline before prune/group')
# parser.add_argument('--iterative_flag', default=False, type=bool, help='prune iteratively or as one-shot')
parser.add_argument('--task', default='finetune', type=str, help='Select from one of the following tasks: finetune, train, test')
parser.add_argument('--gpu', default='', help='gpu available')
args = parser.parse_args()
with open(args.setting_dir) as setting_f:
global setting
setting = json.load(setting_f)
if args.output_folder_dir is not '':
setting['management']['output_folder_dir'] = args.output_folder_dir
if args.output_folder_dir[-1] != '/':
setting['management']['output_folder_dir'] += '/'
if not os.path.isdir(setting['management']['output_folder_dir']):
os.makedirs(setting['management']['output_folder_dir'])
log_formatter = logging.Formatter("%(asctime)s | %(levelname)s : %(message)s")
logger = logging.getLogger()
file_handler = logging.FileHandler(setting['management']['output_folder_dir'] + 'experiment.log', mode = 'w')
file_handler.setFormatter(log_formatter)
logger.addHandler(file_handler)
console_handler = logging.StreamHandler(sys.stdout)
console_handler.setFormatter(log_formatter)
logger.addHandler(console_handler)
logger.setLevel(logging.INFO)
logger.info(f'Experiment {args.exp_desc} starts at {start_time}.')
logger.info(f'Parsed setting file from {args.setting_dir}...')
if args.gpu is not '':
os.environ["CUDA_VISIBLE_DEVICES"] = args.gpu
device = 'cuda' if torch.cuda.is_available() else 'cpu'
# device = 'cpu'
setting['results'] = dict()
if args.exp_desc is not '':
setting['management']['exp_desc'] = args.exp_desc
else:
setting['management']['exp_desc'] = None
if args.model_dir is not '':
setting['management']['model_dir'] = args.model_dir
if args.snapshot_folder_dir is not '':
setting['management']['snapshot_folder_dir'] = args.snapshot_folder_dir
if args.snapshot_folder_dir[-1] != '/':
setting['management']['snapshot_folder_dir'] += '/'
setting['management']['multi_prune_flag'] = args.multi_prune_flag
setting['management']['task'] = args.task
setting['management']['dataset'] = args.dataset
setting['management']['dataset_dir'] = args.dataset_dir
logger.info(f"==> Preparing dataset: {setting['management']['dataset']}")
if args.task != 'prune':
if setting['management']['dataset'] == 'cifar10':
transform_train = transforms.Compose([
transforms.RandomCrop(32, padding=4),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010)),])
transform_test = transforms.Compose([
transforms.ToTensor(),
transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010)),])
trainset = torchvision.datasets.CIFAR10(root=setting['management']['dataset_dir'] , train=True, download=True, transform=transform_train)
trainloader = torch.utils.data.DataLoader(trainset, batch_size=setting['train_params']['train_batch_size'], shuffle=True, num_workers=2)
testset = torchvision.datasets.CIFAR10(root=setting['management']['dataset_dir'] , train=False, download=True, transform=transform_test)
testloader = torch.utils.data.DataLoader(testset, batch_size=setting['train_params']['test_batch_size'], shuffle=False, num_workers=2)
elif setting['management']['dataset'] == 'imagenet':
transform_train = transforms.Compose([
transforms.Scale(256),
transforms.RandomSizedCrop(224),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
transforms.Normalize((0.4914, 0.4824, 0.4467), (0.2471, 0.2435, 0.2616)),
])
transform_test = transforms.Compose([
transforms.Scale(256),
transforms.CenterCrop(224),
transforms.ToTensor(),
transforms.Normalize((0.4914, 0.4824, 0.4467), (0.2471, 0.2435, 0.2616)),
])
trainset = torchvision.datasets.ImageFolder(root=setting['management']['dataset_dir']+'train', transform=transform_train)
trainloader = torch.utils.data.DataLoader(trainset, batch_size=setting['train_params']['train_batch_size'], shuffle=True, num_workers=4, pin_memory=True)
testset = torchvision.datasets.ImageFolder(root=setting['management']['dataset_dir']+'val', transform=transform_test)
testloader = torch.utils.data.DataLoader(testset, batch_size=setting['train_params']['test_batch_size'], shuffle=False, num_workers=4, pin_memory=True)
# pass
elif setting['management']['dataset'] == 'tiny_imagenet':
transform_train = transforms.Compose([
# transforms.RandomSizedCrop(32),
transforms.Scale(64),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
transforms.Normalize((0.4802, 0.4481, 0.3975), (0.2770, 0.2691, 0.2821)),
])
transform_test = transforms.Compose([
# transforms.Scale(32),
transforms.Scale(64),
transforms.ToTensor(),
transforms.Normalize((0.4802, 0.4481, 0.3975), (0.2770, 0.2691, 0.2821)),
])
trainset = torchvision.datasets.ImageFolder(root=setting['management']['dataset_dir']+'train', transform=transform_train)
trainloader = torch.utils.data.DataLoader(trainset, batch_size=setting['train_params']['train_batch_size'], shuffle=True, num_workers=1)
testset = torchvision.datasets.ImageFolder(root=setting['management']['dataset_dir']+'val', transform=transform_test)
testloader = torch.utils.data.DataLoader(testset, batch_size=setting['train_params']['test_batch_size'], shuffle=False, num_workers=1)
# pass
else:
logger.error(f"Invalid dataset_dir: {setting['management']['dataset']}")
logger.info(f"==> Building baseline model: {setting['management']['model_dir']} (dataset: {setting['management']['dataset']})")
if setting['management']['dataset'] == 'cifar10':
from group_cifar import *
from prune_cifar import *
elif setting['management']['dataset'] == 'imagenet' and setting['management']['multi_prune_flag']:
from group_imagenet_multi import *
from prune_imagenet_multi import *
elif setting['management']['dataset'] == 'imagenet' and not setting['management']['multi_prune_flag']:
from group_imagenet import *
from prune_imagenet import *
elif setting['management']['dataset'] == 'tiny_imagenet' and not setting['management']['multi_prune_flag'] and args.resnet_type == 'bottleneck':
from models.resnet_bottleneck import *
from group_imagenet import *
from prune_imagenet import *
elif setting['management']['dataset'] == 'tiny_imagenet' and not setting['management']['multi_prune_flag'] and args.resnet_type == 'basicblock':
print('is basicblock')
from group_cifar import *
from prune_cifar import *
else:
logger.error(f"Invalid input on task: {setting['management']['task']} or multi_prune_flag {setting['management']['multi_prune_flag']}")
sys.exit()
net = torch.load(args.model_dir)
# net = torch.load(args.model_dir, map_location=torch.device('cpu'))
if args.model_state_dict_dir != '':
temp_checkpoint = torch.load(args.model_state_dict_dir)
net.load_state_dict(temp_checkpoint['model_state_dict'])
setting['management']['model_state_dict_dir'] = args.model_state_dict_dir
logger.info(f"Model state dict at {setting['management']['model_state_dict_dir']} loaded into {args.model_dir}.")
net = net.to(device)
criterion = nn.CrossEntropyLoss()
logger.info(f"Model {setting['management']['model_dir']} loaded ({type(net)})")
def test(epoch=None):
net.eval()
test_loss = 0
correct = 0
total = 0
with torch.no_grad():
for batch_idx, (inputs, targets) in enumerate(testloader):
inputs, targets = inputs.to(device), targets.to(device)
outputs = net(inputs)
loss = criterion(outputs, targets)
test_loss += loss.item()
_, predicted = outputs.max(1)
total += targets.size(0)
correct += predicted.eq(targets).sum().item()
acc = 100.*correct/total
return acc, test_loss
baseline_test_acc = -1
setting['results']['baseline'] = baseline_test_acc
if args.baseline and not isinstance(net, GroupResNet) and not isinstance(net, ModifiedResNet):
baseline_test_acc, _ = test()
baseline_net_output_path = setting['management']['output_folder_dir'] + 'baseline'
setting['results']['baseline'] = baseline_test_acc
torch.save(net, baseline_net_output_path)
logger.info(f'Model {setting["management"]["model_dir"]} {type(net)} has a baseline of {baseline_test_acc}.')
if not isinstance(net, GroupResNet) and args.task == 'finetune':
if not isinstance(net, ModifiedResNet):
original_model = copy.deepcopy(net)
net = ModifiedResNet(net, setting, original_model, dataset = args.dataset, pruning_rate = sum(setting['prune_params']['pruning_rate']), pruned_flag = False).cuda()
pruned_net_output_path = setting['management']['output_folder_dir'] + 'pruned'
torch.save(net, pruned_net_output_path)
logger.info(f'Model {setting["management"]["model_dir"]} {type(net)} now pruned.')
net = GroupResNet(net, dataset = args.dataset).cuda()
grouped_net_output_path = setting['management']['output_folder_dir'] + 'grouped'
torch.save(net, grouped_net_output_path)
logger.info(f'Model {setting["management"]["model_dir"]} {type(net)} now grouped.')
net.to(device)
else:
logger.info(f'Model {setting["management"]["model_dir"]} {type(net)} is already grouped or task is test or train (task: {args.task} {args.zero_mask}).')
lr = setting['train_params']['lr']
lr_step_size = setting['train_params']['lr_step_size']
weight_decay = setting['train_params']['weight_decay']
momentum = setting['train_params']['momentum']
gamma = setting['train_params']['gamma']
optimizer = optim.SGD(net.parameters(), lr=lr, momentum=momentum, weight_decay=weight_decay)
scheduler = optim.lr_scheduler.StepLR(optimizer, step_size=lr_step_size, gamma=gamma)
logger.info(f"==> Starting task: {setting['management']['task']} for {sum(setting['train_params']['epoch_num'])} epochs; multi_prune_flag: {setting['management']['multi_prune_flag']})")
def train(epoch):
net.train()
train_loss = 0
correct = 0
total = 0
for batch_idx, (inputs, targets) in enumerate(trainloader):
inputs, targets = inputs.to(device), targets.to(device)
optimizer.zero_grad()
outputs = net(inputs)
loss = criterion(outputs, targets)
loss.backward()
optimizer.step()
train_loss += loss.item()
_, predicted = outputs.max(1)
total += targets.size(0)
correct += predicted.eq(targets).sum().item()
if args.zero_mask:
zeroize_pruned_filters(net)
acc = 100.*correct/total
return acc, loss
train_acc_list = []
train_loss_list = []
test_acc_list = []
test_loss_list = []
def process_epoch_results(current_epoch_num, epoch_snapshot, epoch_eval_info, save_subdir = None, overwrite_save_flag = False):
train_acc, train_loss, test_acc, test_loss, best_acc, best_epoch, current_lr = epoch_eval_info
epoch_dict = {
'model_state_dict': epoch_snapshot,
'epoch': current_epoch_num,
'train_acc': train_acc,
'train_loss': train_loss,
'test_acc': test_acc,
'test_loss': test_loss,
}
train_acc_list.append(train_acc)
train_loss_list.append(train_loss)
test_acc_list.append(test_acc)
test_loss_list.append(test_loss)
if test_acc >= best_acc:
best_acc = test_acc
best_epoch = current_epoch_num
logger.info(f'Epoch #{current_epoch_num}')
logger.info(f'\t train_acc: {train_acc}; train_loss: {train_loss :.3f}; test_loss: {test_loss :.3f}; lr: {current_lr}')
logger.info(f'\t test_acc: {test_acc} (best: {best_acc} from #{best_epoch})')
if save_subdir is not None:
snapshot_save_path = setting['management']['output_folder_dir'] + str(save_subdir) + '/'
if not os.path.isdir(snapshot_save_path):
os.mkdir(snapshot_save_path)
if not overwrite_save_flag:
torch.save(epoch_dict, snapshot_save_path + f'epoch_{current_epoch_num}.pt')
else:
epoch_dict['optimizer'] = optimizer
epoch_dict['scheduler'] = scheduler
epoch_dict['model'] = net
if (current_epoch_num + 1) % 10 == 0:
torch.save(epoch_dict, snapshot_save_path + f'epoch_{current_epoch_num}.pt')
torch.save(epoch_dict, snapshot_save_path + 'most_recent_checkpoint.pt')
return epoch_dict
def finetune_model(num_of_epoch, baseline_net_output_path, finetuned_net_output_path):
baseline_acc = -1
if args.baseline:
test_acc, _ = test()
baseline_acc = test_acc
logger.info(f'Model {setting["management"]["model_dir"]} has a baseline of {test_acc}.')
torch.save(net, baseline_net_output_path)
best_acc = baseline_acc
best_epoch = -1
for current_epoch_num in range(num_of_epoch):
train_acc, train_loss = train(current_epoch_num)
test_acc, test_loss = test(current_epoch_num)
epoch_snapshot = net.state_dict()
current_lr = scheduler.get_last_lr()
epoch_eval_info = (train_acc, train_loss, test_acc, test_loss, best_acc, best_epoch, current_lr)
current_epoch_dict = process_epoch_results(current_epoch_num, epoch_snapshot, epoch_eval_info, save_subdir = setting['management']['save_subdir'], overwrite_save_flag = setting['management']['overwrite_save_flag'])
scheduler.step()
if test_acc >= best_acc:
best_acc = test_acc
best_epoch = current_epoch_num
torch.save(net, finetuned_net_output_path)
logger.info(f'Model {setting["management"]["model_dir"]} finetuned {baseline_acc} --> {best_acc}.')
return baseline_acc, best_acc
if setting['management']['task'] == 'test':
# torch.backends.cudnn.benchmark=True
# torch.cuda.synchronize()
test_start_time = datetime.datetime.utcnow()
test_acc, _ = test()
test_end_time = datetime.datetime.utcnow()
# torch.cuda.synchronize()
test_duration = test_end_time - test_start_time
test_duration_ms = test_duration.total_seconds() * 1000
logger.info(f'Model {setting["management"]["model_dir"]} has a baseline of {test_acc};\nTest time: {test_start_time.strftime("%Y-%m-%dT%H:%M:%S.%f%Z")} --> {test_end_time.strftime("%Y-%m-%dT%H:%M:%S.%f%Z")} (duration: {test_duration_ms} ms).')
sys.exit(0)
elif setting['management']['task'] == 'train':
num_of_epoch = sum(setting['train_params']['epoch_num'])
baseline_net_output_path = setting['management']['output_folder_dir'] + 'initialized'
finetuned_net_output_path = setting['management']['output_folder_dir'] + 'trained'
baseline_test_acc, finetuned_test_acc = finetune_model(num_of_epoch, baseline_net_output_path, finetuned_net_output_path)
sys.exit(0)
elif setting['management']['task'] == 'finetune':
num_of_epoch = sum(setting['train_params']['epoch_num'])
grouped_test_acc = -1
if args.baseline:
grouped_test_acc, _ = test()
logger.info(f'Model {setting["management"]["model_dir"]} (type: {type(net)}) pruned and grouped (grouped baseline: {grouped_test_acc}); now to finetune with {num_of_epoch} epochs.\n')
grouped_net_output_path = setting['management']['output_folder_dir'] + 'grouped'
finetuned_net_output_path = setting['management']['output_folder_dir'] + 'finetuned'
grouped_test_acc, finetuned_test_acc = finetune_model(num_of_epoch, grouped_net_output_path, finetuned_net_output_path)
setting["results"]["grouped_test_acc"] = grouped_test_acc
setting["results"]["best_acc"] = finetuned_test_acc
logger.info(f"Model {setting['management']['model_dir']} pruned, grouped, and finetuned: {setting['results']['baseline']} --> {grouped_test_acc} --> {finetuned_test_acc}.\n")
end_time = datetime.datetime.utcnow().strftime("%Y-%m-%dT%H:%M:%S.%f%Z")
setting['results']['start_time'] = start_time
setting['results']['end_time'] = end_time
setting['results']['train_acc_list'] = train_acc_list
setting['results']['train_loss_list'] = [i.item() for i in train_loss_list]
setting['results']['test_acc_list'] = test_acc_list
setting['results']['test_loss_list'] = [i for i in test_loss_list]
setting_output_path = setting['management']['output_folder_dir'] + 'setting.json'
logger.info('setting', setting)
with open(setting_output_path, 'w+') as out_setting_f:
logger.info(f'Saving setting file to {setting_output_path}...')
json.dump(setting, out_setting_f, indent = 4)
logger.info(f'Experiment {args.exp_desc} (task: {setting["management"]["task"]}) ends at {end_time}; result: {setting["results"]["baseline"]} --> {setting["results"]["best_acc"]}.')