forked from Gorilla-Lab-SCUT/OrthDNNs
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Trainer.py
277 lines (237 loc) · 11.9 KB
/
Trainer.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
from __future__ import division
import time
import numpy as np
import math
import random
import torch
import torch.nn as nn
import torch.nn.parallel
import torch.nn.functional as F
import torch.optim as optim
from torch.autograd import Variable
import torchvision
from Utility import Average_meter
from Utility import Training_aux
#from Utility import progress_bar
class Trainer(object):
"""a method that packaging dataloader and model and optim_methods"""
"""the model are trained here"""
"""the mixup operation and data_agu operation are perform here"""
def __init__(self, train_loader, val_loader, model, criterion,
optimizer, nEpoch, lr_base = 0.1, lr_end = 0.001, lr_decay_method = 'exp',
is_soft_regu=False, is_SRIP=False, soft_lambda = 1e-4,
svb_flag = False, iter_svb_flag=False, svb_factor = 0.5,
bbn_flag = False, bbn_factor = 0.2, bbn_type = 'rel',
fsave = './Save', print_freq = 10, is_evaluate = False, dataset = 'CIFAR10'):
self.train_loader = train_loader
self.val_loader = val_loader
self.model = model
self.criterion = criterion
self.optimizer = optimizer
self.nEpoch = nEpoch
self.lr_base = lr_base
self.lr_end = lr_end
self.lr_decay_method = lr_decay_method
self.is_soft_regu = is_soft_regu
self.is_SRIP = is_SRIP
self.soft_lambda = soft_lambda
self.svb_flag = svb_flag
self.iter_svb_flag = iter_svb_flag
self.svb_factor = svb_factor
self.bbn_flag = bbn_flag
self.bbn_factor = bbn_factor
self.bbn_type = bbn_type
self.training_aux = Training_aux(fsave)
self.is_evaluate = is_evaluate
self.print_freq = print_freq
self.best_prec1 = 0
def train(self, epoch):
"""Train for one epoch on the training set"""
batch_time = Average_meter()
data_time = Average_meter()
losses = Average_meter()
top1 = Average_meter()
top5 = Average_meter()
# switch to train mode
self.model.train()
begin = time.time()
for i, (image, target) in enumerate(self.train_loader):
batch_size= image.size(0)
# measure data loading time
data_time.update(time.time() - begin)
image = image.cuda()
input_var = Variable(image)
target = target.cuda()
target_var = Variable(target)
output = self.model(input_var)
if self.is_soft_regu or self.is_SRIP:
loss = self.criterion(output, target_var, self.model, self.soft_lambda)
else:
loss = self.criterion(output, target_var)
# measure accuracy and record loss
prec1, prec5 = self.training_aux.accuracy(output.data, target, topk=(1, 5))
losses.update(loss.data.item(), batch_size)
top1.update(prec1.item(), batch_size)
top5.update(prec5.item(), batch_size)
# compute gradient and do SGD step
self.optimizer.zero_grad()
loss.backward()
self.optimizer.step()
# measure elapsed time
batch_time.update(time.time() - begin)
if i % self.print_freq == 0:
#progress_bar(i, len(self.train_loader), 'Loss: {loss.avg:.4f} | Prec@1 {top1.avg:.3f} | Prec@5 {top5.avg:.3f}'.format(loss=losses, top1=top1, top5=top5))
print('Epoch: [{0}][{1}/{2}]\t'
'Time {batch_time.avg:.3f}\t'
'Data {data_time.avg:.3f}\t'
'Loss {loss.avg:.4f}\t'
'Prec@1 {top1.avg:.3f}\t'
'Prec@5 {top5.avg:.3f}'.format(
epoch, i, len(self.train_loader), batch_time=batch_time,
data_time=data_time, loss=losses, top1=top1, top5=top5))
begin = time.time()
if (self.iter_svb_flag) and epoch != (self.nEpoch -1) and i != (self.train_loader.__len__() -1):
self.fcConvWeightReguViaSVB()
self.training_aux.write_err_to_file(epoch = epoch, top1 = top1, top5 = top5, trn_loss = losses, mode = 'train')
return
def validate(self, epoch, img_size=320):
"""Perform validation on the validation set"""
batch_time = Average_meter()
losses = Average_meter()
top1 = Average_meter()
top5 = Average_meter()
self.model.eval()
begin = time.time()
with torch.no_grad():
for i, (raw_img, raw_label) in enumerate(self.val_loader):
raw_label = raw_label.cuda()
raw_img = raw_img.cuda()
input_var = Variable(raw_img)
target_var = Variable(raw_label)
# compute output
output = self.model(input_var)
# measure accuracy and record loss
criterion = nn.CrossEntropyLoss()
loss = criterion(output, target_var)
# measure accuracy and record loss
prec1, prec5 = self.training_aux.accuracy(output.data, raw_label, topk=(1, 5))
top1.update(prec1.item(), raw_img.size(0))
top5.update(prec5.item(), raw_img.size(0))
losses.update(loss.data.item(), raw_img.size(0))
# measure elapsed time
batch_time.update(time.time() - begin)
if i % self.print_freq == 0:
#progress_bar(i, len(self.train_loader), 'Loss: {loss.avg:.4f} | Prec@1 {top1.avg:.3f} | Prec@5 {top5.avg:.3f}'.format(loss=losses, top1=top1, top5=top5))
print('Test: [{0}/{1}]\t'
'Time {batch_time.avg:.3f}\t'
'Loss {loss.avg:.4f}\t'
'{top1.avg:.3f}\t'
'{top5.avg:.3f}'.format(
i, len(self.val_loader), batch_time=batch_time,
loss=losses, top1=top1, top5=top5))
begin = time.time()
print(' * Loss {loss.avg:.4f} Prec@1 {top1.avg:.3f} Prec@5 {top5.avg:.3f}'
.format(loss=losses, top1=top1, top5=top5))
self.is_best = top1.avg > self.best_prec1
self.best_prec1 = max(top1.avg, self.best_prec1)
if self.is_evaluate:
return top1.avg
else:
self.training_aux.write_err_to_file(epoch = epoch, top1 = top1, top5 = top5, mode = 'val')
return top1.avg
def adjust_learning_rate(self, epoch, warm_up_epoch = 0,scheduler=None):
"""Sets the learning rate to the initial LR decayed by 10 after 0.5 and 0.75 epochs"""
if self.lr_decay_method == 'exp':
lr = self.lr_base
if epoch < warm_up_epoch:
lr = 0.001 + (self.lr_base - 0.001) * epoch / warm_up_epoch
if epoch >= warm_up_epoch:
lr_series = torch.logspace(math.log(self.lr_base, 10), math.log(self.lr_end, 10), int(self.nEpoch/2))
lr = lr_series[int(math.floor((epoch-warm_up_epoch)/2))]
for param_group in self.optimizer.param_groups:
param_group['lr'] = lr
elif self.lr_decay_method == 'noDecay':
lr = self.lr_base
for param_group in self.optimizer.param_groups:
param_group['lr'] = lr
print('lr:{0}'.format(self.optimizer.param_groups[-1]['lr']))
return
def save_checkpoint(self, epoch, save_flag = 'learning', filename = False):
if save_flag == 'standard':
model = self.standard_model
optimizer = self.standard_optimizer
elif save_flag == 'learning':
model = self.model
optimizer = self.optimizer
else:
raise Exception('save_flag should be one of standard or learning')
state = {
'epoch': epoch,
'state_dict': model.state_dict(),
'best_prec1': self.best_prec1,
'optimizer' : optimizer.state_dict(),
}
fname = filename or 'checkpoint' + '.pth.tar'
self.training_aux.save_checkpoint(state = state, is_best = self.is_best, filename=fname)
return
def fcConvWeightReguViaSVB(self):
for m in self.model.modules():
#svb
if self.svb_flag == True:
if isinstance(m,nn.Conv2d):
tmpbatchM = m.weight.data.view(m.weight.data.size(0), -1).t().clone()
try:
tmpU, tmpS, tmpV = torch.svd(tmpbatchM)
except:
tmpbatchM = tmpbatchM[np.logical_not(np.isnan(tmpbatchM))]
tmpbatchM = tmpbatchM.view(m.weight.data.size(0), -1).t()
tmpU, tmpS, tmpV = np.linalg.svd(tmpbatchM.cpu().numpy())
tmpU = torch.from_numpy(tmpU).cuda()
tmpS = torch.from_numpy(tmpS).cuda()
tmpV = torch.from_numpy(tmpV).cuda()
for idx in range(0, tmpS.size(0)):
if tmpS[idx] > (1+self.svb_factor):
tmpS[idx] = 1+self.svb_factor
elif tmpS[idx] < 1/(1+self.svb_factor):
tmpS[idx] = 1/(1+self.svb_factor)
tmpbatchM = torch.mm(torch.mm(tmpU, torch.diag(tmpS.cuda())), tmpV.t()).t().contiguous()
m.weight.data.copy_(tmpbatchM.view_as(m.weight.data))
elif isinstance(m, nn.Linear):
tmpbatchM = m.weight.data.t().clone()
tmpU, tmpS, tmpV = torch.svd(tmpbatchM)
for idx in range(0, tmpS.size(0)):
if tmpS[idx] > (1+self.svb_factor):
tmpS[idx] = 1+self.svb_factor
elif tmpS[idx] < 1/(1+self.svb_factor):
tmpS[idx] = 1/(1+self.svb_factor)
tmpbatchM = torch.mm(torch.mm(tmpU, torch.diag(tmpS.cuda())), tmpV.t()).t().contiguous()
m.weight.data.copy_(tmpbatchM.view_as(m.weight.data))
# bbn
if self.bbn_flag == True:
if isinstance(m, nn.BatchNorm2d):
tmpbatchM = m.weight.data
if self.bbn_type == 'abs':
for idx in range(0, tmpbatchM.size(0)):
if tmpbatchM[idx] > (1+self.bbn_factor):
tmpbatchM[idx] = (1+self.bbn_factor)
elif tmpbatchM[idx] < 1/(1+self.bbn_factor):
tmpbatchM[idx] = 1/(1+self.bbn_factor)
elif self.bbn_type == 'rel':
mean = torch.mean(tmpbatchM)
relVec = torch.div(tmpbatchM, mean)
for idx in range(0, tmpbatchM.size(0)):
if relVec[idx] > (1+self.bbn_factor):
tmpbatchM[idx] = mean * (1+self.bbn_factor)
elif relVec[idx] < 1/(1+self.bbn_factor):
tmpbatchM[idx] = mean/(1+self.bbn_factor)
elif self.bbn_type == 'bbn':
running_var = m.running_var
eps = m.eps
running_std = torch.sqrt(torch.add(running_var, eps))
mean = torch.mean(tmpbatchM/running_std)
for idx in range(0, tmpbatchM.size(0)):
if tmpbatchM[idx]/(running_std[idx]*mean) > 1+self.bbn_factor:
tmpbatchM[idx] = running_std[idx] * mean * (1+self.bbn_factor)
elif tmpbatchM[idx]/(running_std[idx]*mean) < 1/(1+self.bbn_factor):
tmpbatchM[idx] = running_std[idx] * mean / (1+self.bbn_factor)
m.weight.data.copy_(tmpbatchM)