-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathtrain_BT2.py
166 lines (136 loc) · 5.43 KB
/
train_BT2.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
#
# Copyright (C) 2022 Apple Inc. All rights reserved.
#
from accelerate import Accelerator
from utils.schedulers import get_policy
from utils.getters import get_model, get_optimizer
from dataset import SubImageCIFAR100
from trainers import TransferTrainer, build_feature_dict_transfer
from collections import Counter
import tqdm
import torch.nn as nn
import torch
from typing import Dict
from argparse import ArgumentParser
import yaml
from PIL import ImageFile
ImageFile.LOAD_TRUNCATED_IMAGES = True
def main(config: Dict) -> None:
"""Run training.
:param config: A dictionary with all configurations to run training.
:return:
"""
torch.backends.cudnn.benchmark = True
model = get_model(config.get('arch_params'))
accelerator = Accelerator()
device = accelerator.device
old_model = get_model(config.get('old_arch_params'))
old_model.load_state_dict(torch.load(config.get('old_model_path'))['model_state_dict'])
new_model = get_model(config.get('new_arch_params'))
new_model.load_state_dict(torch.load(config.get('new_model_path'))['model_state_dict'])
if torch.cuda.is_available():
model = torch.nn.DataParallel(model)
old_model = torch.nn.DataParallel(old_model)
new_model = torch.nn.DataParallel(new_model)
trainer = TransferTrainer()
optimizer = get_optimizer(model, **config.get('optimizer_params'))
data = SubImageCIFAR100(**config.get('dataset_params'))
lr_policy = get_policy(optimizer, **config.get('lr_policy_params'))
lambda_1 = float(config.get('lambda_1'))
lambda_2 = float(config.get('lambda_2'))
lambda_3 = float(config.get('lambda_3'))
train_loader = data.train_loader
val_loader = data.val_loader
optimizer, train_loader, val_loader =\
accelerator.prepare(optimizer, train_loader, val_loader)
criterion = nn.CosineSimilarity(dim=1)
print("==>Preparing pesudo classifier")
old_model = accelerator.prepare(old_model)
num_classes = int(config.get('arch_params')['num_classes'])
embedding_dim = int(config.get('arch_params')['embedding_dim_old'])
pseudo_classifier = torch.zeros(
num_classes, embedding_dim, requires_grad=False)
label_count = Counter()
for i, (paths, (images, target)) in tqdm.tqdm(
enumerate(data.train_loader), ascii=True, total=len(data.train_loader)
):
images = images.to(device, non_blocking=True)
target = target.cpu()
with torch.no_grad():
outputs = old_model(images)
features = outputs[-1]
for feature, label in zip(features, target):
pseudo_classifier[int(label)] += feature.flatten().cpu()
label_count.update([int(label)])
for i in range(num_classes):
pseudo_classifier[i] = pseudo_classifier[i]/label_count[i]
old_model = old_model.cpu()
pseudo_classifier = pseudo_classifier.to(device)
print(':====>BuildFeatureDict')
# build a dictionary of saved features, so that don't need to recompute
# old features at each iteration
old_feature_dict = {}
new_feature_dict = {}
new_model = accelerator.prepare(new_model)
new_feature_dict = build_feature_dict_transfer(
train_loader, new_model, device, new_feature_dict)
new_feature_dict = build_feature_dict_transfer(
val_loader, new_model, device, new_feature_dict)
new_model = new_model.cpu()
old_model = accelerator.prepare(old_model)
old_feature_dict = build_feature_dict_transfer(
train_loader, old_model, device, old_feature_dict)
old_feature_dict = build_feature_dict_transfer(
val_loader, old_model, device, old_feature_dict)
old_model = old_model.cpu()
model = accelerator.prepare(model)
print(':====>Training')
# Training loop
for epoch in range(config.get('epochs')):
lr_policy(epoch, iteration=None)
train_loss = trainer.train(
train_loader=train_loader,
model=model,
old_feature_dict=old_feature_dict,
new_feature_dict=new_feature_dict,
criterion=criterion,
optimizer=optimizer,
device=device,
accelerator=accelerator,
pseudo_classifier=pseudo_classifier,
lambda_1=lambda_1,
lambda_2=lambda_2,
lambda_3=lambda_3,
)
print(
"Train: epoch = {}, Loss = {}".format(
epoch, train_loss
))
test_loss, test_acc1 = trainer.validate(
val_loader=data.val_loader,
model=model,
old_feature_dict=old_feature_dict,
new_feature_dict=new_feature_dict,
criterion=criterion,
device=device
)
print(
"Test: epoch = {}, Loss = {}, Top1 = {}".format(
epoch, test_loss, test_acc1
))
# make checkpoints
if (epoch+1) % 5 == 0:
torch.save({
'epoch': epoch,
'model_state_dict': model.module.state_dict(),
'optimizer_state_dict': optimizer.state_dict(),
'loss': train_loss,
}, config.get('output_model_path'))
if __name__ == "__main__":
parser = ArgumentParser()
parser.add_argument('--config', type=str, required=True,
help='Path to config file for this pipeline.')
args = parser.parse_args()
with open(args.config) as f:
read_config = yaml.safe_load(f)
main(read_config)