-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathwarmup.py
238 lines (189 loc) · 8.51 KB
/
warmup.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
232
233
234
235
236
237
import os
import sys
import time
import tabulate
import torch
import torch.optim as optim
import torch.nn as nn
import numpy as np
from torch.utils.tensorboard import SummaryWriter
import pickle
import argparse
import random
from utils import (
save_checkpoint,
load_data,
train,
validate,
)
from models.cascade_net import CascadeNet
from torch.utils.data import DataLoader
from scheduler import PolynomialLR
import losses
from models.models import *
from data_loader import BraTSTrainDataset
#from apex import amp
from apex_dummy import amp
parser = argparse.ArgumentParser(description='Test for learning max and min values for cyclic learning rate.')
# In this directory is stored the script used to start the training,
# the most recent and best checkpoints, and a directory of logs.
parser.add_argument('--dir', type=str, required=True, metavar='PATH',
help='The directory to write all output to.')
parser.add_argument('--data_dir', type=str, required=True, metavar='PATH TO DATA',
help='Path to where the data is located.')
parser.add_argument('--model', type=str, default=None, required=True, metavar='MODEL',
help='model class (default: None)')
parser.add_argument('--device', type=int, required=True, metavar='N',
help='Which device to use for training.')
parser.add_argument('--upsampling', type=str, default='bilinear',
choices=['bilinear', 'deconv'],
help='upsampling algorithm to use in decoder (default: bilinear)')
parser.add_argument('--loss', type=str, default='avgdice',
choices=['dice', 'recon', 'avgdice', 'vae'],
help='which loss to use during training (default: avgdice)')
parser.add_argument('--data_par', action='store_true',
help='data parellelism flag (default: off)')
parser.add_argument('--mixed_precision', action='store_true',
help='mixed precision flag (default: off)')
parser.add_argument('--cross_val', action='store_true',
help='use train/val split of full dataset (default: off)')
parser.add_argument('--seed', type=int, default=1, metavar='S',
help='random seed (default: 1)')
parser.add_argument('--wd', type=float, default=1e-4,
help='weight decay (default: 1e-4)')
parser.add_argument('--resume', type=str, default=None, metavar='PATH',
help='checkpoint to resume training from (default: None)')
parser.add_argument('--epochs', type=int, default=300, metavar='N',
help='number of epochs to train (default: 300)')
parser.add_argument('--num_workers', type=int, default=4, metavar='N',
help='number of workers to assign to dataloader (default: 4)')
parser.add_argument('--batch_size', type=int, default=1, metavar='N',
help='batch_size (default: 1)')
parser.add_argument('--save_freq', type=int, default=25, metavar='N',
help='save frequency (default: 25)')
parser.add_argument('--eval_freq', type=int, default=5, metavar='N',
help='evaluation frequency (default: 25)')
parser.add_argument('--lr', type=float, default=1e-4, metavar='LR',
help='initial learning rate (default: 1e-4)')
parser.add_argument('--lr_add_cnst', type=float, default=1e-6, metavar='LR',
help='constant to add to learning rate each epoch (default: 1e-6)')
# Currently unused.
parser.add_argument('--momentum', type=float, default=0.9, metavar='M',
help='SGD momentum (default: 0.9)')
args = parser.parse_args()
if args.device >= 0:
device = torch.device(f'cuda:{args.device}')
else:
device = torch.device('cpu')
args.dir = f'{args.dir}'
os.makedirs(f'{args.dir}/logs', exist_ok=True)
os.makedirs(f'{args.dir}/checkpoints', exist_ok=True)
torch.manual_seed(args.seed)
torch.cuda.manual_seed(args.seed)
np.random.seed(args.seed)
random.seed(args.seed)
torch.manual_seed(args.seed)
torch.backends.cudnn.benchmark = False
torch.backends.cudnn.deterministic = True
with open(os.path.join(f'{args.dir}', 'command.sh'), 'w') as f:
f.write(' '.join(sys.argv))
f.write('\n')
#dims=[128, 128, 128]
dims=[160, 192, 128]
if args.cross_val:
(train_modes, train_segs), (val_modes, val_segs) = cross_val(args.data_dir)
train_data = BraTSTrainDataset(data_dir, dims=dims, augment_data=True,
modes=train_modes, segs=train_segs)
trainloader = DataLoader(train_data, batch_size=batch_size,
shuffle=True, num_workers=num_workers)
val_data = BraTSTrainDataset(data_dir, dims=dims, augment_data=False,
modes=val_modes, segs=val_segs)
valloader = DataLoader(val_data, batch_size=batch_size,
shuffle=True, num_workers=num_workers)
else:
# train without cross_val
train_data = BraTSTrainDataset(args.data_dir, dims=dims, augment_data=True)
trainloader = DataLoader(train_data, batch_size=args.batch_size,
shuffle=True, num_workers=args.num_workers)
val_data = BraTSTrainDataset(args.data_dir, dims=dims, augment_data=False)
valloader = DataLoader(val_data, batch_size=args.batch_size,
shuffle=True, num_workers=args.num_workers)
if args.model == 'MonoUNet':
model = MonoUNet()
loss = losses.AvgDiceLoss()
if args.model == 'CascadeNet':
model = CascadeNet()
loss = losses.CascadeAvgDiceLoss()
optimizer = \
optim.Adam(model.parameters(), lr=args.lr, weight_decay=args.wd)
model = model.to(device)
start_epoch = 0
if args.resume:
checkpoint = torch.load(args.resume)
start_epoch = checkpoint["epoch"]
model.load_state_dict(checkpoint["state_dict"])
optimizer.load_state_dict(checkpoint["optimizer"])
print(f'Resume training from {args.resume}, epoch {checkpoint["epoch"]}.')
writer = SummaryWriter(log_dir=f'{args.dir}/logs')
# model has to be on device before passing to amp
if args.mixed_precision:
# Allow Amp to perform casts as required by the opt_level
model, optimizer = amp.initialize(model, optimizer, opt_level="O1")
# TODO: optimizer factory, allow for SGD with momentum etx.
columns = ['set', 'ep', 'lr', 'loss', 'dice_et', 'dice_wt','dice_tc', \
'time', 'mem_usage']
lr = args.lr
for epoch in range(start_epoch, args.epochs):
time_ep = time.time()
model.train()
train(model,
loss,
optimizer,
trainloader,
device,
mixed_precision=args.mixed_precision)
if (epoch + 1) % args.save_freq == 0:
save_checkpoint(
f'{args.dir}/checkpoints',
epoch + 1,
state_dict=model.state_dict(),
optimizer=optimizer.state_dict()
)
if (epoch + 1) % args.eval_freq == 0:
# Evaluate on training data
model.eval()
train_val = validate(model, loss, trainloader, device)
eval_val = validate(model, loss, valloader, device)
time_ep = time.time() - time_ep
memory_usage = torch.cuda.memory_allocated() / (1024.0 ** 3)
train_values = ['train', epoch + 1, lr*1000, train_val['loss'].data] \
+ train_val['dice'].tolist()\
+ [ time_ep, memory_usage]
eval_values = ['eval', epoch + 1, lr*1000, eval_val['loss'].data] \
+ eval_val['dice'].tolist()\
+ [ time_ep, memory_usage]
table = tabulate.tabulate([train_values, eval_values],
columns, tablefmt="simple", floatfmt="8.4f")
print(table)
# Log validation
writer.add_scalar(f'{args.dir}/logs/loss/train', train_val['loss'], epoch)
et, wt, tc = train_val['dice']
writer.add_scalar(f'{args.dir}/logs/dice/train/et', et, epoch)
writer.add_scalar(f'{args.dir}/logs/dice/train/wt', wt, epoch)
writer.add_scalar(f'{args.dir}/logs/dice/train/tc', tc, epoch)
writer.add_scalar(f'{args.dir}/logs/dice/train/et_lr', et, lr)
writer.add_scalar(f'{args.dir}/logs/dice/train/wt_lr', wt, lr)
writer.add_scalar(f'{args.dir}/logs/dice/train/tc_lr', tc, lr)
writer.add_scalar(f'{args.dir}/logs/loss/eval', eval_val['loss'], epoch)
et, wt, tc = eval_val['dice']
writer.add_scalar(f'{args.dir}/logs/dice/eval/et', et, epoch)
writer.add_scalar(f'{args.dir}/logs/dice/eval/wt', wt, epoch)
writer.add_scalar(f'{args.dir}/logs/dice/eval/tc', tc, epoch)
writer.add_scalar(f'{args.dir}/logs/dice/eval/et_lr', et, lr)
writer.add_scalar(f'{args.dir}/logs/dice/eval/wt_lr', wt, lr)
writer.add_scalar(f'{args.dir}/logs/dice/eval/tc_lr', tc, lr)
writer.flush()
# equivalent: scheduler.step()
lr = (epoch + 1)*args.lr_add_cnst
for param_group in optimizer.param_groups:
param_group['lr'] = lr