/
find_lr.py
199 lines (161 loc) · 7.04 KB
/
find_lr.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
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import os
import math
import argparse
import pprint
import tqdm
import pandas as pd
import numpy as np
import torch
import torch.nn.functional as F
import matplotlib.pyplot as plt
from datasets import get_dataloader
from models import get_model
from losses import get_loss
from optimizers import get_optimizer
from schedulers import CLR
import utils
from utils import checkpoint
from utils.utils import Logger, seed_everything
def update_lr(optimizer, lr):
for g in optimizer.param_groups:
g['lr'] = lr
def update_mom(optimizer, mom):
for g in optimizer.param_groups:
g['momentum'] = mom
def train(args, model, dataloader, criterion, optimizer, clr):
running_loss = 0.
avg_beta = 0.98
model.train()
tbar = tqdm.tqdm(dataloader, total=len(dataloader))
for batch_idx, data in enumerate(tbar):
images = data['image'].cuda()
labels = data['mask'].cuda()
masks, cls_logits = model(images)
B,C,H,W = labels.size()
cls_labels = labels.view(B, C, H*W)
cls_labels = torch.sum(cls_labels, dim=2)
cls_labels = (cls_labels > 0).float()
if args.mode == 'classification':
loss = criterion(cls_logits, cls_labels)
elif args.mode == 'segmentation':
loss = criterion(masks, labels)
elif args.pretrain:
loss = criterion(masks, labels)
else:
loss = criterion(masks, cls_logits, labels, cls_labels)
running_loss = avg_beta * running_loss + (1-avg_beta) *loss.data
smoothed_loss = running_loss / (1 - avg_beta**(batch_idx+1))
tbar.set_description('loss: %.5f' % (smoothed_loss))
lr = clr.calc_lr(smoothed_loss)
if lr == -1 :
clr.plot()
plt.show()
plt.savefig('lr_find.png')
break
update_lr(optimizer, lr)
# compute gradient and do SGD step
optimizer.zero_grad()
loss.backward()
optimizer.step()
def run(args, log):
df = pd.read_csv(args.df_path)
df_train = df[df['Fold']!=args.fold]
df_valid = df[df['Fold']==args.fold]
dfs = {}
dfs['train'] = df_train
dfs['val'] = df_valid
model = get_model(args).cuda()
if args.mode != 'segmentation':
for param in model.model.encoder.parameters():
param.requires_grad = True
for param in model.model.decoder.parameters():
param.requires_grad = True
for params in model.model.classification_head.parameters():
params.requires_grad = False
elif args.mode == 'classification':
for param in model.model.encoder.parameters():
param.requires_grad = False
for param in model.model.decoder.parameters():
param.requires_grad = False
for param in model.classification_head.parameters():
param.requires_grad = True
criterion = get_loss(args)
optimizer = get_optimizer(args, model)
if args.initial_ckpt is not None:
last_epoch, step = checkpoint.load_checkpoint(args, model, checkpoint=args.initial_ckpt)
log.write(f'Resume training from {args.initial_ckpt} @ {last_epoch}\n')
else:
last_epoch, step = -1, -1
dataloaders = {mode:get_dataloader(args.data_dir, dfs[mode], mode, args.pretrain, args.batch_size) for mode in ['train', 'val']}
seed_everything(seed=123)
clr = CLR(optimizer, len(dataloaders['train']))
train(args, model, dataloaders['train'], criterion, optimizer, clr)
def parse_args():
parser = argparse.ArgumentParser(description='Kaggle Cloud Competition')
parser.add_argument('--gpu', type=int, default=0,
help='Choose GPU to use. This only support single GPU')
parser.add_argument('--data_dir', default='./data/train_images',
help='datasest directory')
parser.add_argument('--df_path', default='./data/train_splits.csv',
help='df_path')
parser.add_argument('--fold', type=int, default=0,
help='which fold to use for training')
parser.add_argument('--model_name', type=str, default='FPN_effb4',
help='model_name as exp_name')
parser.add_argument('--batch_size', type=int, default=16,
help='batch size')
parser.add_argument('--num_epochs', type=int, default=30,
help='num of epochs to train')
parser.add_argument('--mode', type=str, default=None,
help='mode: segmentation or classification or both')
parser.add_argument('--pretrain', type=bool, default=False,
help='number of classes')
parser.add_argument('--use_compressor', type=bool, default=False,
help='whether to use small network for extraction of features')
parser.add_argument('--optimizer_name', type=str, default='adamW',
help='name of optimizer to use')
parser.add_argument('--lr', type=float, default=0.0005,
help='minimum learning rate for scheduler')
parser.add_argument('--max_lr', type=float, default=0.0001,
help='maximum learning rate for scheduler')
parser.add_argument('--weight_decay', type=float, default=1e-4,
help='weight decay for optimizer')
parser.add_argument('--no_bias_decay', type=bool, default=True,
help='wheter to apply weight decay to bias or not?')
parser.add_argument('--encoder_name', type=str, default='efficientnet-b4',
help='which encode to use for model')
parser.add_argument('--encoder_lr_ratio', type=float, default=0.1,
help='relative learning rate-ratio for encode')
parser.add_argument('--decoder_name', type=str, default='FPN',
help='type of decoder to use for model')
parser.add_argument('--num_class', type=int, default=4,
help='number of classes')
parser.add_argument('--initial_ckpt', type=str, default=None,
help='inital checkpoint to resume training')
parser.add_argument('--log_dir', type=str, default='runs',
help='logging directory')
return parser.parse_args()
def main():
import warnings
warnings.filterwarnings("ignore")
args = parse_args()
os.environ['CUDA_VISIBLE_DEVICES']= f'{args.gpu}'
utils.prepare_train_directories(args)
log = Logger()
log.open(args.log_dir + '/' + args.model_name + f'/fold_{args.fold}' + '/findlr_log.txt', mode='a')
log.write('*'*30)
log.write('\n')
log.write('Logging arguments!!\n')
log.write('*'*30)
log.write('\n')
for arg, value in sorted(vars(args).items()):
log.write(f'{arg}: {value}\n')
log.write('*'*30)
log.write('\n')
run(args, log)
print('success!')
if __name__ == '__main__':
main()