-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathshrinkage_phase.py
183 lines (155 loc) · 6.59 KB
/
shrinkage_phase.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
import os
import torch
import torch.nn as nn
import torch.optim as optim
import torch.distributed as dist
from utils import network_initialization, get_dataloader
from utils import (
get_m_s,
norm,
get_optim,
Loss,
set_seed,
get_center,
LargeMarginLoss,
)
from tqdm import tqdm
import datetime
import time
class Trainer:
def __init__(self, args):
set_seed(args.seed)
# dataloader
self.train_loader, self.dev_loader, _ = get_dataloader(args)
# model initialization
self.model = network_initialization(args)
# get mean and std for normalization
self.m, self.s = get_m_s(args)
self.save_path = os.path.join(args.save_path, args.dataset)
os.makedirs(self.save_path, exist_ok=True)
if args.resume:
pretrained_path = os.path.join(
self.save_path, f'{args.resume_model}.pt'
)
else:
pretrained_path = os.path.join(
self.save_path,
f'restricted_model_{args.model}.pt'
)
self.checkpoint = torch.load(pretrained_path)
self.model.module.load_state_dict(self.checkpoint["model_state_dict"])
try:
for param in self.model.module.fc.parameters():
param.requires_grad = False
except:
for param in self.model.module.classifier.parameters():
param.requires_grad = False
self.center = get_center(
self.model, self.train_loader, args.num_class, args.device, self.m, self.s
)
# set criterion
self.criterion_CE = nn.CrossEntropyLoss()
self.criterion = Loss(
args.num_class,
args.device,
pre_center=self.center,
phase=args.phase,
)
def training(self, args):
# set optimizer & scheduler
optimizer, scheduler = get_optim(
self.model, args.lr_intra, intra=True
)
# base model
model_name = f"intra_model_{args.model}.pt"
model_path = os.path.join(self.save_path, model_name)
best_loss = 1000
current_step = 0
dev_step = 0
if args.resume:
optimizer.load_state_dict(self.checkpoint["optimizer_state_dict"])
scheduler.load_state_dict(self.checkpoint["scheduler_state_dict"])
trained_epoch = self.checkpoint["trained_epoch"] + 1
best_loss = self.checkpoint["best_loss"]
else:
trained_epoch = 0
trn_loss_log = tqdm(total=0, position=2, bar_format='{desc}')
dev_loss_log = tqdm(total=0, position=4, bar_format='{desc}')
best_epoch_log = tqdm(total=0, position=5, bar_format='{desc}')
outer = tqdm(total=args.epochs-trained_epoch, desc="Epoch", position=0, leave=False)
os.makedirs('time_log', exist_ok=True)
f = open(f"time_log/{args.dataset}_{args.phase}.txt", 'w')
start_total = time.time()
# train target classifier
for epoch in range(trained_epoch, args.epochs):
start_epoch = time.time()
_dev_loss = 0.0
train = tqdm(total=len(self.train_loader), desc="Steps", position=1, leave=False)
dev = tqdm(total=len(self.dev_loader), desc="Steps", position=3, leave=False)
for step, (inputs, labels) in enumerate(self.train_loader):
self.model.train()
current_step += 1
if inputs.size(1) == 1:
inputs = inputs.repeat(1, 3, 1, 1)
inputs, labels = inputs.to(args.device), labels.to(args.device)
inputs = norm(inputs, self.m, self.s)
logit, features = self.model(inputs)
_, predict = torch.max(logit, 1)
correct_idx = predict.eq(labels)
ce_loss = self.criterion_CE(logit, labels)
intra_loss = self.criterion(features, labels, correct_idx)
loss = 0.*ce_loss + 10.0*intra_loss
optimizer.zero_grad()
loss.backward()
optimizer.step()
#################### Logging ###################
trn_loss_log.set_description_str(
f"[TRN] Total Loss: {loss.item():.4f}, CE Loss: {ce_loss.item():.4f}, Intra Loss: {intra_loss.item():.4f}"
)
train.update(1)
epoch_time = round(time.time() - start_epoch)
epoch_time = str(datetime.timedelta(seconds=epoch_time))
f.write(f"Epoch {epoch+1}: "+str(epoch_time)+'\n')
for idx, (inputs, labels) in enumerate(self.dev_loader):
self.model.eval()
dev_step += 1
if inputs.size(1) == 1:
inputs = inputs.repeat(1, 3, 1, 1)
inputs, labels = inputs.to(args.device), labels.to(args.device)
inputs = norm(inputs, self.m, self.s)
with torch.no_grad():
logit, features = self.model(inputs)
_, predict = torch.max(logit, 1)
correct_idx = predict.eq(labels)
ce_loss = self.criterion_CE(logit, labels)
intra_loss = self.criterion(features, labels, correct_idx)
loss = intra_loss
# Loss
_dev_loss += loss
dev_loss = _dev_loss / (idx + 1)
dev_loss_log.set_description_str(
f"[DEV] CE Loss: {ce_loss.item():.4f} Intra Loss: {intra_loss.item():.4f}"
)
dev.update(1)
if dev_loss > best_loss:
best_epoch_log.set_description_str(
f"Best Epoch: {epoch} / {args.epochs} | Best Loss: {dev_loss}"
)
best_loss = dev_loss
torch.save(
{
"model_state_dict": self.model.module.state_dict(),
"optimizer_state_dict": optimizer.state_dict(),
"scheduler_state_dict": scheduler.state_dict(),
"trained_epoch": epoch,
"best_loss": dev_loss,
"center": self.center
},
f"{model_path.split('_')[0]}_{str(epoch)}_{'_'.join(model_path.split('_')[1:])}"
)
scheduler.step(dev_loss)
outer.update(1)
total_time = round(time.time() - start_total)
total_time = str(datetime.timedelta(seconds=total_time))
f.write(f"Total: "+str(total_time)+'\n')
f.close()