-
Notifications
You must be signed in to change notification settings - Fork 4
/
Copy pathtrain_cifar.py
291 lines (250 loc) · 12.8 KB
/
train_cifar.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
""" Search cell """
import os
import torch
import torch.nn as nn
import torch.nn.functional as F
import numpy as np
from tensorboardX import SummaryWriter
from util_func.config import TrainCifarConfig
import util_func.utils as utils
# from models.search_cnn import SearchCNNController
import torchvision.transforms as transforms
import torchvision.datasets as datasets
import torch.optim
import torch.utils.data
import pytorch_warmup as warmup
from models_util import *
from models_cifar import *
from train_util import *
from util_func.dataset import get_dataset, DATASETS
import math
import copy
config = TrainCifarConfig()
device = torch.device("cuda")
# tensorboard
writer = SummaryWriter(log_dir=os.path.join(config.path, "tb"))
writer.add_text('config', config.as_markdown(), 0)
logger = utils.get_logger(os.path.join(config.path, "{}.log".format(config.dataset)))
config.print_params(logger.info)
def main():
logger.info("Logger is set - training start")
# set default gpu device id
torch.cuda.set_device(config.gpus[0])
device = torch.device("cuda")
# set seed
np.random.seed(config.seed)
torch.manual_seed(config.seed)
torch.cuda.manual_seed_all(config.seed)
torch.backends.cudnn.benchmark = True
model = model_ReLU_RP(config)
criterion = nn.CrossEntropyLoss().to(device)
if config.distil:
config_teacher = copy.deepcopy(config)
config_teacher.act_type = 'nn.ReLU'
config_teacher.arch = config_teacher.teacher_arch
teacher_model = model_ReLU_RP(config_teacher)
model.criterion = criterion
# if 'vgg' in config.arch:
# model = vgg.__dict__[config.arch](config, criterion, config.act_type, config.pool_type)
# elif 'resnet' in config.arch:
# model = resnet.__dict__[config.arch](config, criterion, config.act_type, config.pool_type, num_classes = config.num_classes)
# elif config.arch == "mobilenet_v2":
# model = mbnet.__dict__[config.arch](config, criterion, config.act_type, num_classes = config.num_classes)
# config.pretrained_path = '/data3/hop20001/mpc_proj/PASNET_cifar10_search/pretrained/cifar10_baseline/vgg_checkpoint_best.tar'
# config.pretrained_path = '/data3/hop20001/mpc_proj/PASNET_cifar10_search/pretrained/cifar10_baseline/resnet18.pt'
### finetune from checkpoint
if config.pretrained_path:
print("==> Load pretrained")
model.load_pretrained(pretrained_path = config.pretrained_path)
if config.distil:
teacher_model.load_pretrained(pretrained_path = config.teacher_path)
if(config.checkpoint_path):
config.start_epoch, best_top1 = model.load_check_point(check_point_path = config.checkpoint_path)
model = model.to(device)
if config.distil:
teacher_model = teacher_model.to(device)
teacher_model.eval()
criterion_kd = SoftTarget(4.0).to(device)
train_dataset = get_dataset(config, 'train')
val_dataset = get_dataset(config, 'test')
pin_memory = (config.dataset == "imagenet")
train_loader = torch.utils.data.DataLoader(train_dataset, shuffle=True, batch_size=config.batch_size,
num_workers=config.workers, pin_memory=pin_memory)
val_loader = torch.utils.data.DataLoader(val_dataset, shuffle=False, batch_size=config.batch_size,
num_workers=config.workers, pin_memory=pin_memory)
if config.evaluate:
# # ----------------------------
checkpoint = torch.load(config.evaluate, map_location="cpu")
model.load_state_dict(checkpoint['state_dict'])
# print(model)
from torchinfo import summary
device = "cuda"
summary(model, input_size=config.x_size, device=device, depth=3, verbose=2,
col_names=["input_size",
"output_size",
"kernel_size"],
)
if config.act_type != 'nn.ReLU':
model.print_alphas(logger)
validate(val_loader, model, 0, len(val_loader), device, config, logger, writer)
return
# # ----------------------------
# weights optimizer
# w_optim = torch.optim.SGD(model.weights(), config.w_mask_lr, momentum=config.w_momentum,
# weight_decay=config.w_weight_decay)
# #w_optim = torch.optim.Adam(model.weights(), config.w_mask_lr)
# # alphas optimizer
# if config.act_type != 'nn.ReLU':
# alpha_optim = torch.optim.Adam(model.alpha_aux(), config.alpha_lr, betas=(0.5, 0.999),
# weight_decay=config.alpha_weight_decay)
w_optim = torch.optim.Adam(model.weights_and_alpha(), config.w_mask_lr)
# w_optim = torch.optim.Adam(model.weights_and_alpha(), config.w_mask_lr, weight_decay=config.w_weight_decay)
# w_optim = torch.optim.SGD(model.weights_and_alpha(), config.w_mask_lr)
# w_optim = torch.optim.SGD(model.weights_and_alpha(), config.w_mask_lr, momentum=config.w_momentum,
# weight_decay=config.w_weight_decay)
# param_groups = [
# {'optimizer':w_optim,'T_max':config.mask_epochs, 'eta_min':config.w_lr_min},
# {'optimizer':alpha_optim,'T_max':config.mask_epochs}
# ]
lr_scheduler_w = torch.optim.lr_scheduler.CosineAnnealingLR(optimizer = w_optim, T_max = config.mask_epochs,
eta_min = config.w_lr_min)
# if config.act_type != 'nn.ReLU':
# lr_scheduler_alpha = torch.optim.lr_scheduler.CosineAnnealingLR(optimizer = alpha_optim, T_max = config.mask_epochs)
# warmup_scheduler = warmup.UntunedLinearWarmup(alpha_optim)
#### Freeze batch normalization ####
# model.train_fz_bn(freeze_bn=True)
model.train_fz_bn(freeze_bn=False)
# model.change_mask_dropout_ratio(config.mask_dropout)
# lambda_l1 = 1e-6
# lambda_l2 = 5e-4
lambda0 = config.lamda
# training loop
best_top1 = 0.
for epoch in range(config.start_epoch, config.mask_epochs):
if config.act_type != 'nn.ReLU':
model.update_sparse_list()
model.print_sparse_list(logger)
if config.precision == 'full':
if(config.distil):
top1_train, global_density, total_mask = train_mask_distil(train_loader, model, w_optim, lambda0, teacher_model, criterion_kd, epoch,
device, config, logger, writer)
else:
top1_train, global_density, total_mask = train_mask(train_loader, model, w_optim, lambda0, epoch,
device, config, logger, writer)
else:
if(config.distil):
top1_train, global_density, total_mask = train_mask_distil_fp16(train_loader, model, w_optim, lambda0, teacher_model, criterion_kd, epoch,
device, config, logger, writer)
else:
top1_train, global_density, total_mask = train_mask_fp16(train_loader, model, w_optim, lambda0, epoch,
device, config, logger, writer)
# training without mask update
# adjust_learning_rate(w_optim, epoch, config)
lr_scheduler_w.step()
# lr_scheduler_alpha.step()
# validation
cur_step = (epoch+1) * len(train_loader)
if config.precision == 'full':
top1 = validate(val_loader, model, epoch, cur_step, device, config, logger, writer)
else:
top1 = validate_fp16(val_loader, model, epoch, cur_step, device, config, logger, writer)
# save
# save
relu_count = global_density * total_mask/1000
# if (global_density - config.ReLU_count*1000/total_mask) < 0.01:
if (relu_count - config.ReLU_count) < 2:
if best_top1 < top1:
best_top1 = top1
# best_genotype = genotype
is_best = True
else:
is_best = False
# utils.save_checkpoint(model, config.path, is_best)
# if is_best:
# save_path = os.path.join(config.path, 'best_mask_train.pth.tar')
# else:
# save_path = os.path.join(config.path, 'checkpoint_mask_train.pth.tar')
# model.save_checkpoint(epoch, best_top1, is_best, filename=save_path)
if is_best:
save_path = os.path.join(config.path, 'best_mask_train.pth.tar')
model.save_checkpoint(epoch, best_top1, is_best, filename=save_path)
logger.info("Current mask training best Prec@1 = {:.4%}".format(best_top1))
save_path = os.path.join(config.path, 'checkpoint_mask_train.pth.tar')
model.save_checkpoint(epoch, best_top1, False, filename=save_path)
#### Start to finetune ####
para_group = [{'params': model.weights(), 'initial_lr': config.w_lr}]
w_optim = torch.optim.SGD(para_group, config.w_lr, momentum=config.w_momentum,
weight_decay=config.w_weight_decay)
# param_groups = [
# {'optimizer':w_optim,'T_max':config.mask_epochs, 'eta_min':config.w_lr_min},
# ]
if config.optim == 'cosine_rst':
# lr_scheduler = torch.optim.lr_scheduler.CosineAnnealingWarmRestarts(w_optim, 1, T_mult=2, eta_min=config.w_lr_min) #, last_epoch = config.epochs
T_mult = 2
lr_scheduler = torch.optim.lr_scheduler.CosineAnnealingWarmRestarts(w_optim, 1, T_mult=T_mult, eta_min=config.w_lr_min) #, last_epoch = config.epochs
elif config.optim == 'cosine':
lr_scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(
w_optim, config.epochs, eta_min=config.w_lr_min)
elif config.optim == 'cosine_finetune':
lr_scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(
w_optim, T_max = 400, eta_min=config.w_lr_min, last_epoch= 400 - config.epochs)
# lr_scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(optimizer = w_optim, T_max = config.epochs,
# eta_min = config.w_lr_min)
model.train_fz_bn(freeze_bn=False)
model.change_dropout_ratio(config.dropout)
model.change_mask_dropout_ratio(config.mask_dropout)
best_top1 = 0.
for epoch in range(config.start_epoch, config.epochs):
if config.act_type != 'nn.ReLU':
model.update_sparse_list()
model.print_sparse_list(logger)
if config.precision == 'full':
if(config.distil):
top1_train = train_distil(train_loader, model, w_optim, teacher_model, criterion_kd, epoch,
device, config, logger, writer)
else:
top1_train = train(train_loader, model, w_optim, epoch,
device, config, logger, writer)
else:
if(config.distil):
top1_train = train_distil_fp16(train_loader, model, w_optim, teacher_model, criterion_kd, epoch,
device, config, logger, writer)
else:
top1_train = train_fp16(train_loader, model, w_optim, epoch,
device, config, logger, writer)
# adjust_learning_rate(w_optim, epoch, config)
if config.optim == 'cos_modified':
cos_modified_learning_rate(w_optim, epoch, config)
else:
lr_scheduler.step()
# validation
cur_step = (epoch+1) * len(train_loader)
if config.precision == 'full':
top1 = validate(val_loader, model, epoch, cur_step, device, config, logger, writer)
else:
top1 = validate_fp16(val_loader, model, epoch, cur_step, device, config, logger, writer)
# save
if best_top1 < top1:
best_top1 = top1
# best_genotype = genotype
is_best = True
else:
is_best = False
# utils.save_checkpoint(model, config.path, is_best)
if is_best:
save_path = os.path.join(config.path, 'best.pth.tar')
else:
save_path = os.path.join(config.path, 'checkpoint.pth.tar')
model.save_checkpoint(epoch, best_top1, is_best, filename=save_path)
logger.info("Current best Prec@1 = {:.4%}".format(best_top1))
if (epoch % 20) == 0:
logger.info("Perform validation on training dataset. ")
if config.precision == 'full':
top1_train_wo_dropout = validate_train(train_loader, model, epoch, cur_step, device, config, logger, writer)
else:
top1_train_wo_dropout = validate_train_fp16(train_loader, model, epoch, cur_step, device, config, logger, writer)
logger.info("Final train Prec@1 = {:.4%}".format(top1_train_wo_dropout))
logger.info("Final best validation Prec@1 = {:.4%}".format(best_top1))
# logger.info("Best Genotype = {}".format(best_genotype))
if __name__ == "__main__":
main()