-
Notifications
You must be signed in to change notification settings - Fork 3
/
Copy pathtrain.py
202 lines (161 loc) · 7.9 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
import utils.utils
import torch.nn as nn
import torch
import numpy as np
import torch.nn.functional as F
import time
from utils.utils import format_time
class Mixup_Criterion(nn.Module):
def __init__(self, beta, cls_criterion):
super().__init__()
self.beta = beta
self.cls_criterion = cls_criterion
def get_mixup_data(self, image, target) :
beta = np.random.beta(self.beta, self.beta)
index = torch.randperm(image.size()[0]).to(image.device)
shuffled_image, shuffled_target = image[index], target[index]
mixed_image = beta * image + (1 - beta) * shuffled_image
return mixed_image, shuffled_target, beta
def forward(self, image, target, net):
mixed_image, shuffled_target, beta = self.get_mixup_data(image, target)
pred_mixed = net(mixed_image)
loss_mixup = beta * self.cls_criterion(pred_mixed, target) + (1 - beta) * self.cls_criterion(pred_mixed, shuffled_target)
return loss_mixup
class Correctness_Log(object):
def __init__(self, n_data):
self.correctness = np.zeros((n_data))
self.max_correctness = 1
# correctness update
def update(self, data_idx, correctness):
self.correctness[data_idx] += correctness.cpu().numpy()
def max_correctness_update(self, epoch):
if epoch > 1:
self.max_correctness += 1
# correctness normalize (0 ~ 1) range
def _normalize(self, data):
data_min = self.correctness.min()
data_max = float(self.max_correctness)
return (data - data_min) / (data_max - data_min)
# get target & margin
def get_target_margin(self, idx1, idx2):
idx1 = idx1.cpu().numpy()
idx2 = idx2.cpu().numpy()
correctness_norm = self._normalize(self.correctness)
target1, target2 = correctness_norm[idx1], correctness_norm[idx2]
# 1 for idx1 > idx2, 0 for idx1 = idx2, -1 for idx1 < idx2
target = np.array(target1 > target2, dtype='float') + np.array(target1 < target2, dtype='float') * (-1)
target = torch.from_numpy(target).float().cuda()
# calc margin
margin = abs(target1 - target2)
margin = torch.from_numpy(margin).float().cuda()
return target, margin
class CRL_Criterion(nn.Module):
'''
Confidence-Aware Learning for Deep Neural Networks
ICML 2020
http://proceedings.mlr.press/v119/moon20a/moon20a.pdf
code borrows from: https://github.com/daintlab/confidence-aware-learning/blob/master/crl_utils.py
'''
def __init__(self):
super().__init__()
self.rank_criterion = torch.nn.MarginRankingLoss(margin=0)
def forward(self, output, image_idx, correct_log):
conf, _ = F.softmax(output, dim=1).max(dim=1)
conf_roll, image_idx_roll = torch.roll(conf, -1), torch.roll(image_idx, -1)
# ranking target:
# 1 for image_idx > image_idx_roll
# 0 for image_idx = image_idx_roll
# -1 for image_idx < image_idx_roll
rank_target, rank_margin = correct_log.get_target_margin(image_idx, image_idx_roll)
conf_roll = conf_roll + rank_margin / (rank_target + 1e-7)
ranking_loss = self.rank_criterion(conf, conf_roll, rank_target)
return ranking_loss
def compute_loss(args, net, image, target, image_idx, correct_log, cls_criterion, mixup_criterion, rank_criterion):
'''
feature = net.module.forward_features(image)
output = net.module.forward_head(feature)
'''
output = net(image)
loss_ce = cls_criterion(output, target)
mixup_criterion.get_mixup_data(image, target)
loss_mixup = mixup_criterion(image, target, net)
loss_crl = rank_criterion(output, image_idx, correct_log)
'''
# energy loss
Ec_in = -torch.logsumexp(output, dim=1)
loss_energy = torch.pow(F.relu(Ec_in-args.m_in), 2).mean()
'''
# loss = loss_ce + args.mixup_weight * loss_mixup + args.crl_weight * loss_crl + args.energy_weight * loss_energy
loss = loss_ce + args.mixup_weight * loss_mixup + args.crl_weight * loss_crl
return loss, loss_ce, loss_mixup, loss_crl, output
def train(train_loader, net, optimizer, epoch, correct_log, logger, writer, args):
net.train()
## define criterion
cls_criterion = torch.nn.CrossEntropyLoss(label_smoothing=0.1)
mixup_criterion = Mixup_Criterion(beta=args.mixup_beta, cls_criterion=cls_criterion)
rank_criterion = CRL_Criterion()
train_log = {
'Top1 Acc.': utils.utils.AverageMeter(),
'CLS Loss': utils.utils.AverageMeter(),
'Mixup Loss': utils.utils.AverageMeter(),
'CRL Loss': utils.utils.AverageMeter(),
'Tot. Loss': utils.utils.AverageMeter(),
'LR': utils.utils.AverageMeter(),
# 'Energy Loss':utils.utils.AverageMeter(),
}
msg = '####### --- Training Epoch {:d} --- #######'.format(epoch)
if logger is not None:
logger.info(msg)
total_batches = args.epochs * len(train_loader)
batch_start_time = time.time()
for i, (image, target, image_idx) in enumerate(train_loader):
if i == len(train_loader):
break
image, target = image.cuda(), target.long().cuda()
loss, loss_ce, loss_mixup, loss_crl, output = compute_loss(args,
net,
image,
target,
image_idx,
correct_log,
cls_criterion,
mixup_criterion,
rank_criterion)
optimizer.zero_grad()
loss.backward()
if args.optim_name in ['sam', 'fmfp']:
optimizer.first_step(zero_grad=True)
compute_loss(args, net, image, target, image_idx, correct_log, cls_criterion, mixup_criterion,
rank_criterion)[0].backward()
optimizer.second_step(zero_grad=True)
else:
optimizer.step()
prec, correct = utils.utils.accuracy(output, target)
correct_log.update(image_idx, correct)
for param_group in optimizer.param_groups:
lr = param_group["lr"]
break
train_log['Tot. Loss'].update(loss.item(), image.size(0))
train_log['CLS Loss'].update(loss_ce.item(), image.size(0))
train_log['Mixup Loss'].update(loss_mixup.item(), image.size(0))
train_log['CRL Loss'].update(loss_crl.item(), image.size(0))
# train_log['Energy Loss'].update(loss_energy.item(), image.size(0))
train_log['Top1 Acc.'].update(prec.item(), image.size(0))
train_log['LR'].update(lr, image.size(0))
if i % 100 == 99:
elapsed_time = time.time() - batch_start_time
avg_batch_time = elapsed_time / (i + 1)
remaining_batches = total_batches - (i + 1) - (epoch - 1) * len(train_loader)
eta = avg_batch_time * remaining_batches
hrs, mins, secs = format_time(eta)
log = ['LR : {:.5f}'.format(train_log['LR'].avg)] + [key + ': {:.2f}'.format(train_log[key].avg) for key in train_log if key != 'LR']
msg = 'Epoch {:d} \t Batch [{:d}/{:d}]\t'.format(epoch, i + 1, len(train_loader)) + '\t'.join(log) + '\tETA: {:02d}:{:02d}:{:02d}'.format(hrs, mins, secs)
if logger is not None:
logger.info(msg)
for key in train_log:
train_log[key] = utils.utils.AverageMeter()
batch_start_time = time.time()
correct_log.max_correctness_update(epoch)
if writer is not None:
for key in train_log:
writer.add_scalar('./Train/' + key, train_log[key].avg, epoch)