/
train_comb.py
231 lines (192 loc) · 8.32 KB
/
train_comb.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
import os
import time
import shutil
from collections import defaultdict
import torch
import torch.optim as optim
import torch.nn.functional as F
import torch.backends.cudnn as cudnn
import torch.nn as nn
from utils import AverageMeter
import utils
# import cifar_models as cifar_models
# from torch.utils.data.sampler import SubsetRandomSampler
# import json
# from torchvision.utils import make_grid, save_image
# import math
# from warmup_scheduler import GradualWarmupScheduler
from data import get_dataloaders
from models import get_model, num_class
from augment_comb import Augment
def train_and_validate(config):
# data loaders
trainloader, testloader = get_dataloaders(config)
# model
bn_types = ['base']
if config.perturb_vae: bn_types.append('texture')
if config.aug_stn: bn_types.append('stn')
if config.deform_vae: bn_types.append('deform')
# if config.bn_num == 1:
# target_net = get_model(config, num_class(config.dataset))
# else:
target_net = get_model(config, num_class(config.dataset), bn_types=bn_types)
model = Augment(target_net=target_net, config=config)
start_epoch = 0
best_test_acc = 0.0
test_acc = 0.0
if config.resume:
best_test_acc, test_acc, start_epoch = \
utils.load_checkpoint(config, model.target_net, model.target_net_optim)
print('trainloader length: {}'.format(len(trainloader)))
print('testloader length: {}'.format(len(testloader)))
exp_dir = utils.get_log_dir_path(config.exp_dir, config.exp_id)
if not os.path.exists(exp_dir):
os.makedirs(exp_dir)
print('exp_dir: {}'.format(exp_dir))
log_file = os.path.join(exp_dir, 'log.txt')
names = ['epoch', 'lr', 'Train Acc', 'Test Acc', 'Best Test Acc']
with open(log_file, 'a') as f:
f.write('batch size: {}\n'.format(config.batch_size))
f.write('lr: {}\n'.format(config.lr))
f.write('momentum: {}\n'.format(config.momentum))
f.write('weight_decay: {}\n'.format(config.weight_decay))
for per_name in names:
f.write(per_name + '\t')
f.write('\n')
# print('=> Training the base model')
# print('start_epoch {}'.format(start_epoch))
# print(type(start_epoch))
# exit()
print('target net grad clip: {}'.format(config.grad_clip))
for epoch in range(start_epoch, config.epochs):
# lr = adjust_learning_rate(optimizer, epoch, model.module, config)
lr = model.target_net_optim.param_groups[0]['lr']
print('lr: {}'.format(lr))
# inner_lr = get_lr_cosine_decay(config, epoch)
# print('inner_lr: {}'.format(inner_lr))
# train for one epoch
train_acc = train_epoch_multi_bns(trainloader, model, epoch, config)
# evaluate on test set
# print('testing epoch ...')
test_acc = validate_epoch(testloader, model, config)
# remember best acc, evaluate on test set and save checkpoint
is_best = test_acc > best_test_acc
if is_best:
best_test_acc = test_acc
utils.save_checkpoint(model,{
'epoch': epoch + 1,
'state_dict': model.target_net.state_dict(),
'optimizer': model.target_net_optim.state_dict(),
'perturb_vae_state_dict': model.perturb_vae.state_dict() if model.perturb_vae else None,
'perturb_vae_optimizer': model.perturb_vae_optim.state_dict() if model.perturb_vae else None,
'aug_stn_state_dict': model.aug_stn.state_dict() if model.aug_stn else None,
'aug_stn_optimizer': model.aug_stn_optim.state_dict() if model.aug_stn else None,
'deform_vae_state_dict': model.deform_vae.state_dict() if model.deform_vae else None,
'deform_vae_optimizer': model.deform_vae_optim.state_dict() if model.deform_vae else None,
'test_acc': test_acc,
'best_test_acc': best_test_acc,
}, is_best, exp_dir)
values = [train_acc, test_acc, best_test_acc]
with open(log_file, 'a') as f:
f.write('{:d}\t'.format(epoch))
f.write('{:g}\t'.format(lr))
for per_value in values:
f.write('{:2.2f}\t'.format(per_value))
f.write('\n')
print('exp_dir: {}'.format(exp_dir))
# not using implicit gradients from validation data
def train_epoch_multi_bns(trainloader, model, epoch, config):
print('using function train_epoch_multi_bns...')
batch_time = AverageMeter()
data_time = AverageMeter()
losses = AverageMeter()
losses_adv = AverageMeter()
losses_div = AverageMeter()
# losses3 = AverageMeter()
# losses4 = AverageMeter()
top1 = AverageMeter()
top1_aug_texture = AverageMeter()
top1_aug_stn = AverageMeter()
top1_aug_deform = AverageMeter()
model.target_net.train()
model.train_mode()
loader_len = len(trainloader)
end = time.time()
for i, (input_list, target) in enumerate(trainloader):
# measure data loading time
# print('iter: {}'.format(i))
data_time.update(time.time() - end)
assert isinstance(input_list, list)
assert len(input_list) == 2
input, input_preaug = input_list[0], input_list[1]
# print('input size: {}'.format(input.size()))
# print('input_autoaug size: {}'.format(input_autoaug.size()))
# print('target size: {}'.format(target.size()))
input, input_preaug, target = input.cuda(), input_preaug.cuda(), target.cuda()
# update aug_net and target_net
model.target_net_optim.zero_grad()
if model.aug_stn:
for j in range(config.inner_num):
model.stn_step(input, target)
if model.deform_vae:
for j in range(config.inner_num):
model.deform_step(input, target)
if model.perturb_vae:
for j in range(config.inner_num):
model.texture_step(input, target)
# for j in range(config.inner_num):
# model.comb_step(input, target)
model.target_net_optim.zero_grad()
output_preaug = model.target_net(input_preaug, 'base')
loss_preaug = model.criterion(output_preaug, target)
loss_preaug.backward()
if config.grad_clip and config.grad_clip > 0:
nn.utils.clip_grad_norm_(model.target_net.parameters(), config.grad_clip)
model.target_net_optim.step()
# update lr
model.lr_scheduler.step(epoch + float(i + 1) / loader_len)
acc = utils.accuracy(output_preaug, target)[0]
losses.update(loss_preaug.item(), input.size(0))
top1.update(acc.item(), input.size(0))
# measure elapsed time
batch_time.update(time.time() - end)
end = time.time()
# print('momentum_buffer: {}'.format(momentum_buffer[0][0, 0, 0:10]))
if i % config.print_freq == 0:
print('Epoch: [{0}][{1}/{2}]\t'
'Acc {top1.val:.3f}% ({top1.avg:.3f}%)\t'
'Loss {losses.val:.4f} ({losses.avg:.4f})\t'.format(
epoch, i, len(trainloader), top1=top1, losses=losses))
# exit()
print(' * Acc {top1.avg:.3f}% '.format(top1=top1))
# exit()
return top1.avg
def validate_epoch(val_loader, model, config):
batch_time = AverageMeter()
losses = AverageMeter()
top1 = AverageMeter()
# switch to evaluate mode
model.target_net.eval()
end = time.time()
with torch.no_grad():
for i, (input, target) in enumerate(val_loader):
input, target = input.cuda(), target.cuda()
# compute output
output = model.target_net(input)
loss = model.criterion(output, target)
# measure accuracy and record loss
acc = utils.accuracy(output, target)[0]
losses.update(loss.item(), input.size(0))
top1.update(acc.item(), input.size(0))
# measure elapsed time
batch_time.update(time.time() - end)
end = time.time()
if i % config.print_freq == 0:
print('Test: [{0}/{1}]\t'
'Time {batch_time.val:.3f} ({batch_time.avg:.3f})\t'
'Loss {loss.val:.4f} ({loss.avg:.4f})\t'
'Acc {top1.val:.3f}% ({top1.avg:.3f}%)'.format(
i, len(val_loader), batch_time=batch_time, loss=losses,
top1=top1))
print(' * Acc {top1.avg:.3f}% '.format(top1=top1))
return top1.avg