-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathtrain_final.py
210 lines (190 loc) · 9.06 KB
/
train_final.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
"""
"""
import json
import os
import torch.backends.cudnn as cudnn
import torch.nn.parallel
from dataloader.data_utils import get_class_weights
from functions import train_one_epoch, train_out_one_epoch, test, set_optimizer, save_checkpoint
from models.loss import WL2SP, UMSE
from utils import Logger, mkdir_p
from utils.args import args_for_train_tl
from noisystudent.ensemble import create_ensemble_noisy_labels
# get all the input variables
args = args_for_train_tl()
state = {k: v for k, v in args._get_kwargs()}
if not os.path.isdir(args.checkpoint):
print("==> Creating checkpoint folder")
mkdir_p(args.checkpoint)
# save all the arguments
args_save_path = args.checkpoint if args.resume == '' else os.path.dirname(args.resume)
with open(os.path.join(args_save_path, 'experiment_args.txt'), 'w') as f:
json.dump(args.__dict__, f, indent=2)
# writer = SummaryWriter(os.path.join(args_save_path, "tensorboard"))
# Use CUDA
os.environ['CUDA_VISIBLE_DEVICES'] = args.gpu_id
use_cuda = torch.cuda.is_available()
# set parameters
best_acc = 0 # best test accuracy
do_save_checkpoint = True
start_epoch = args.start_epoch # start from epoch 0 or last checkpoint epoch
############################################
# Create and Load Model
############################################
from models.models import get_model
model, image_size = get_model(args, num_classes=200, use_pretrained=args.pre_trained, train=True)
image_size = image_size
model = model.cuda()
model = torch.nn.DataParallel(model)
cudnn.benchmark = True
print(' Total params: %.4fM' % (sum(p.numel() for p in model.parameters()) / 1000000.0))
optimizer = set_optimizer(model, args)
scheduler = torch.optim.lr_scheduler.ExponentialLR(optimizer, 0.94, last_epoch=-1)
# pre-trained weights for loss function
pretrained_weights = {}
for n, m in model.named_modules():
if "conv" in n:
try:
pretrained_weights[n] = m.weight
except:
pass
cls_num_list, per_cls_weights = get_class_weights(0.99)
criterion = WL2SP(class_weights=per_cls_weights, pretrained_weights=pretrained_weights)
mse_criterion = UMSE()
##############################################
# Load Dataset
##############################################
print('==> Preparing dataset')
from dataloader.data import get_trainval_data, get_test_data, get_noisy_student_data, get_outdist_data, get_final_train_data, get_indist_data
trainloader, validloader = get_trainval_data(batch_size=args.train_batch, val_batch_size=32, image_size=image_size)
# Noisy Data
# noisy annots means labels generated by teacher model
indist_noisy_annots = "/root/volume/cvpr/results/noisy_labels/inceptionresnetv2_rbn_bs32_baseline.csv"
indist_noisy_data_path = "u_train_in"
test_noisy_annots = "/root/volume/cvpr/results/noisy_labels/inceptionresnetv2_rbn_bs32_test_baseline.csv"
test_noisy_data_path = "test"
noisy_trainloader = get_noisy_student_data(indist_noisy_data_path, indist_noisy_annots, batch_size=32, image_size=image_size)
test_noisy_trainloader = get_noisy_student_data(test_noisy_data_path, test_noisy_annots, batch_size=32, image_size=image_size)
#
ood_trainloader = get_outdist_data(image_size=image_size)
##############################################
# Initialize Model
#############################################
title = '{}-{}-{}-{}'.format(args.dataset, args.model, args.depth, args.norm)
# Resume
if args.resume:
# Load checkpoint.
print('==> Resuming from checkpoint..')
assert os.path.isfile(args.resume), 'Error: no checkpoint directory found!'
args.checkpoint = os.path.dirname(args.resume)
checkpoint = torch.load(args.resume)
best_acc = checkpoint['best_acc']
start_epoch = checkpoint['epoch']
if start_epoch==121:
best_acc = 0
model.load_state_dict(checkpoint['state_dict'])
optimizer.load_state_dict(checkpoint['optimizer'])
logger = Logger(os.path.join(args.checkpoint, 'log.txt'), title=title, resume=True)
else:
logger = Logger(os.path.join(args.checkpoint, 'log.txt'), title=title)
logger.set_names(['Type', 'Learning Rate', 'Train Loss', 'Valid Loss', 'Train Acc.', 'Valid Acc.', 'Train Acc.5',
'Valid Acc.5'])
############################################
# Train and Validate
############################################
test_flag = False
for epoch in range(start_epoch, args.epochs - 20):
print('\nEpoch: [%d | %d] LR: %f' % (epoch + 1, args.epochs, scheduler.get_lr()[0]))
################################################
# Different Training Regimes
################################################
learning_rate = scheduler.get_lr()[0]
# training on train set, supervised
train_loss, train_acc, train_acc5 = train_one_epoch(trainloader, model, criterion, optimizer, use_cuda=use_cuda, test_flag=test_flag)
valid_loss, valid_acc, valid_acc5 = test(validloader, model, criterion, use_cuda=use_cuda, test_flag=test_flag)
logger.append(
[0, args.lr, train_loss, valid_loss, train_acc, valid_acc, train_acc5, valid_acc5])
# # training with noisy student.
# if epoch % 5 == 0 and epoch - 50 != 0 and epoch>15:
# train_loss, train_acc, train_acc5 = train_one_epoch(noisy_trainloader, model, criterion, optimizer, use_cuda=use_cuda, test_flag=test_flag)
# valid_loss, valid_acc, valid_acc5 = test(validloader, model, criterion, use_cuda=use_cuda, test_flag=test_flag)
# logger.append(
# [1, learning_rate, train_loss, valid_loss, train_acc, valid_acc, train_acc5, valid_acc5])
#
# if epoch % 5 == 0 and epoch - 50 != 0and epoch>15:
# train_loss, train_acc, train_acc5 = train_one_epoch(test_noisy_trainloader, model, criterion, optimizer, use_cuda=use_cuda, test_flag=test_flag)
# valid_loss, valid_acc, valid_acc5 = test(validloader, model, criterion, use_cuda=use_cuda, test_flag=test_flag)
# logger.append(
# [2, learning_rate, train_loss, valid_loss, train_acc, valid_acc, train_acc5, valid_acc5])
#
# if epoch % 10 == 0 and epoch - 50 != 0:
# # ToDo: Train on Out Of distribution data
# out_loss = train_out_one_epoch(ood_trainloader, model, mse_criterion, optimizer, use_cuda=use_cuda, test_flag=test_flag)
# valid_loss, valid_acc, valid_acc5 = test(validloader, model, criterion, use_cuda=use_cuda, test_flag=test_flag)
# logger.append(
# [3, learning_rate, out_loss, valid_loss, 0, valid_acc, 0, valid_acc5])
# if epoch % 100 == 0 and epoch - 50 != 0:
# # ToDo: recreate noisy students labels via 10 ensemble for In distribution Data
# indist_dataloader = None
# create_ensemble_noisy_labels(model, indist_dataloader, indist_noisy_annots, ensemble_epochs=10, tolerance=0.0, test_flag=test_flag)
# # ToDo: create nosiy studwent labels via 1- ensemble for test data
# # ToDo: load In distribution data again
# # ToDo: Load TestSet data again.
# test_dataloader = None
# create_ensemble_noisy_labels(model, test_dataloader, indist_noisy_annots, ensemble_epochs=10, tolerance=0.0, test_flag=test_flag)
if epoch%4==0:
scheduler.step()
# save model ap
is_best = valid_acc > best_acc
best_loss = max(valid_acc, best_acc)
best_acc = max(valid_acc, best_acc)
if do_save_checkpoint:
save_checkpoint({
'epoch': epoch + 1,
'state_dict': model.state_dict(),
'acc': valid_acc,
'best_acc': best_acc,
'optimizer': optimizer.state_dict(),
}, is_best, checkpoint=args.checkpoint)
# ToDo: Save separate model before this step
# ToDO: FineTune on train+validation data
###################################
# FineTune on Validation Data ###
###################################
# print("==> Loading best model for fine tuning...")
# checkpoint = torch.load(os.path.join(args.checkpoint, "model_best.pth.tar"))
# model.load_state_dict(checkpoint['state_dict'])
#
# # freeze feature extraction layers
# for n, m in model.named_modules():
# if "conv" in n:
# m.requires_grad = False
#
#
#
# print("==> Fine Tuning on Validation Data")
# for fine_epoch in range(args.epochs-20, args.epochs):
# # Deferred Re-balancing Optimization Schedule (https://arxiv.org/pdf/1906.07413.pdf)
# cls_num_list, per_cls_weights = get_class_weights(0)
# criterion = WL2SP(class_weights=per_cls_weights, pretrained_weights=pretrained_weights)
# scheduler.get_lr()[0] = args.lr
# print('\nEpoch: [%d | %d] LR: %f' % (fine_epoch + 1, args.epochs, scheduler.get_lr()[0]))
# train_loss, train_acc, train_acc5 = train_one_epoch(validloader, model, criterion, optimizer, use_cuda=use_cuda)
# valid_loss, valid_acc, valid_acc5 = train_loss, train_acc, train_acc5#test(validloader, model, criterion, use_cuda=use_cuda)
# logger.append([scheduler.get_lr()[0], train_loss, valid_loss, train_acc, valid_acc, train_acc5, valid_acc5])
# # log_tensorboard(writer, fine_epoch, train_loss, train_acc, train_acc5, valid_loss, valid_acc, valid_acc5)
#
# # save model ap
# is_best = valid_acc > best_acc
# best_loss = max(valid_acc, best_acc)
# best_acc = max(valid_acc, best_acc)
# if do_save_checkpoint:
# save_checkpoint({
# 'epoch': fine_epoch + 1,
# 'state_dict': model.state_dict(),
# 'acc': valid_acc,
# 'best_acc': best_acc,
# 'optimizer' : optimizer.state_dict(),
# }, is_best, checkpoint=args.checkpoint)
logger.close()
print('Best acc:', best_acc)