-
Notifications
You must be signed in to change notification settings - Fork 22
/
train.py
231 lines (182 loc) · 9.31 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
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 argparse
import ast
from collections import deque
import numpy as np
import torch
from torch import nn
import torch.nn.functional as F
from models.model_factory import *
from optimizer.optimizer_helper import get_optim_and_scheduler
from data import *
from utils.Logger import Logger
from utils.tools import *
import warnings
warnings.filterwarnings("ignore")
from warnings import simplefilter
simplefilter(action='ignore', category=FutureWarning)
def get_args():
parser = argparse.ArgumentParser()
parser.add_argument("--source", choices=available_datasets, help="Source", nargs='+')
parser.add_argument("--target", choices=available_datasets, help="Target")
parser.add_argument("--input_dir", default=None, help="The directory of dataset lists")
parser.add_argument("--output_dir", default=None, help="The directory to save logs and models")
parser.add_argument("--config", default=None, help="Experiment configs")
parser.add_argument("--tf_logger", type=ast.literal_eval, default=True, help="If true will save tensorboard compatible logs")
args = parser.parse_args()
config_file = "config." + args.config.replace("/", ".")
print(f"\nLoading experiment {args.config}\n")
config = __import__(config_file, fromlist=[""]).config
return args, config
class Trainer:
def __init__(self, args, config, device):
self.args = args
self.config = config
self.device = device
self.global_step = 0
# networks
self.encoder = get_encoder_from_config(self.config["networks"]["encoder"]).to(device)
self.classifier = get_classifier_from_config(self.config["networks"]["classifier"]).to(device)
# teacher networks
self.encoder_teacher = get_encoder_from_config(self.config["networks"]["encoder"]).to(device)
self.classifier_teacher = get_classifier_from_config(self.config["networks"]["classifier"]).to(device)
preprocess_teacher(self.encoder, self.encoder_teacher)
preprocess_teacher(self.classifier, self.classifier_teacher)
# optimizers
self.encoder_optim, self.encoder_sched = \
get_optim_and_scheduler(self.encoder, self.config["optimizer"]["encoder_optimizer"])
self.classifier_optim, self.classifier_sched = \
get_optim_and_scheduler(self.classifier, self.config["optimizer"]["classifier_optimizer"])
# dataloaders
self.train_loader = get_fourier_train_dataloader(args=self.args, config=self.config)
self.val_loader = get_val_dataloader(args=self.args, config=self.config)
self.test_loader = get_test_loader(args=self.args, config=self.config)
self.eval_loader = {'val': self.val_loader, 'test': self.test_loader}
def _do_epoch(self):
criterion = nn.CrossEntropyLoss()
# turn on train mode
self.encoder.train()
self.classifier.train()
self.encoder_teacher.train()
self.classifier_teacher.train()
for it, (batch, label, domain) in enumerate(self.train_loader):
# preprocessing
batch = torch.cat(batch, dim=0).to(self.device)
label = torch.cat(label, dim=0).to(self.device)
# domain = torch.cat(domain, dim=0).to(self.device)
# zero grad
self.encoder_optim.zero_grad()
self.classifier_optim.zero_grad()
# forward
loss_dict = {}
correct_dict = {}
num_samples_dict = {}
total_loss = 0.0
features = self.encoder(batch)
scores = self.classifier(features)
with torch.no_grad():
features_teacher = self.encoder_teacher(batch)
scores_teacher = self.classifier_teacher(features_teacher)
assert batch.size(0) % 2 == 0
split_idx = int(batch.size(0) / 2)
scores_ori, scores_aug = torch.split(scores, split_idx)
scores_ori_tea, scores_aug_tea = torch.split(scores_teacher, split_idx)
scores_ori_tea, scores_aug_tea = scores_ori_tea.detach(), scores_aug_tea.detach()
labels_ori, labels_aug = torch.split(label, split_idx)
assert scores_ori.size(0) == scores_aug.size(0)
# classification loss for original data
loss_cls = criterion(scores_ori, labels_ori)
loss_dict["main"] = loss_cls.item()
correct_dict["main"] = calculate_correct(scores_ori, labels_ori)
num_samples_dict["main"] = int(scores.size(0) / 2)
# classification loss for augmented data
loss_aug = criterion(scores_aug, labels_aug)
loss_dict["aug"] = loss_aug.item()
correct_dict["aug"] = calculate_correct(scores_aug, labels_aug)
num_samples_dict["aug"] = int(scores.size(0) / 2)
# calculate probability
p_ori, p_aug = F.softmax(scores_ori / self.config["T"], dim=1), F.softmax(scores_aug / self.config["T"], dim=1)
p_ori_tea, p_aug_tea = F.softmax(scores_ori_tea / self.config["T"], dim=1), F.softmax(scores_aug_tea / self.config["T"], dim=1)
# use KLD for consistency loss
loss_ori_tea = F.kl_div(p_aug.log(), p_ori_tea, reduction='batchmean')
loss_aug_tea = F.kl_div(p_ori.log(), p_aug_tea, reduction='batchmean')
# get consistency weight
const_weight = get_current_consistency_weight(epoch=self.current_epoch,
weight=self.config["lam_const"],
rampup_length=self.config["warmup_epoch"],
rampup_type=self.config["warmup_type"])
# calculate total loss
total_loss = 0.5 * loss_cls + 0.5 * loss_aug + \
const_weight * loss_ori_tea + const_weight * loss_aug_tea
loss_dict["ori_tea"] = loss_ori_tea.item()
loss_dict["aug_tea"] = loss_aug_tea.item()
loss_dict["total"] = total_loss.item()
# backward
total_loss.backward()
# update
self.encoder_optim.step()
self.classifier_optim.step()
self.global_step += 1
# update teachers
warm_update_teacher(self.encoder, self.encoder_teacher, self.config["teacher_momentum"], self.global_step)
warm_update_teacher(self.classifier, self.classifier_teacher, self.config["teacher_momentum"], self.global_step)
# record
self.logger.log(
it=it,
iters=len(self.train_loader),
losses=loss_dict,
samples_right=correct_dict,
total_samples=num_samples_dict
)
# turn on eval mode
self.encoder.eval()
self.classifier.eval()
self.encoder_teacher.eval()
self.classifier_teacher.eval()
# evaluation
with torch.no_grad():
for phase, loader in self.eval_loader.items():
total = len(loader.dataset)
class_correct = self.do_eval(loader)
class_acc = float(class_correct) / total
self.logger.log_test(phase, {'class': class_acc})
self.results[phase][self.current_epoch] = class_acc
# save from best val
if self.results['val'][self.current_epoch] >= self.best_val_acc:
self.best_val_acc = self.results['val'][self.current_epoch]
self.best_val_epoch = self.current_epoch + 1
self.logger.save_best_model(self.encoder, self.classifier, self.best_val_acc)
def do_eval(self, loader):
correct = 0
for it, (batch, domain) in enumerate(loader):
data, labels, domains = batch[0].to(self.device), batch[1].to(self.device), domain.to(self.device)
features = self.encoder(data)
scores = self.classifier(features)
correct += calculate_correct(scores, labels)
return correct
def do_training(self):
self.logger = Logger(self.args, self.config, update_frequency=30)
self.logger.save_config()
self.epochs = self.config["epoch"]
self.results = {"val": torch.zeros(self.epochs), "test": torch.zeros(self.epochs)}
self.best_val_acc = 0
self.best_val_epoch = 0
for self.current_epoch in range(self.epochs):
# step schedulers
self.encoder_sched.step()
self.classifier_sched.step()
self.logger.new_epoch([group["lr"] for group in self.encoder_optim.param_groups])
self._do_epoch()
self.logger.finish_epoch()
# save from best val
val_res = self.results['val']
test_res = self.results['test']
self.logger.save_best_acc(val_res, test_res, self.best_val_acc, self.best_val_epoch - 1)
return self.logger
def main():
args, config = get_args()
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
trainer = Trainer(args, config, device)
trainer.do_training()
if __name__ == "__main__":
torch.backends.cudnn.benchmark = True
main()