-
Notifications
You must be signed in to change notification settings - Fork 5
/
train.py
327 lines (241 loc) · 11.4 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
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
import os, time, argparse
import cv2
import numpy as np
import random
import torch
import torch.nn as nn
import torch.backends.cudnn as cudnn
from torch.nn.parallel import DistributedDataParallel
import torch.optim as optim
import torch.utils.data
from basenet.model import Model_factory
from loader import ListDataset
from loss import SWM_FPEM_Loss
from utils.lr_scheduler import WarmupPolyLR
from utils.augmentations import Augmentation, Augmentation_test
cudnn.benchmark = True
def parse():
""" set argments """
parser = argparse.ArgumentParser()
parser.add_argument('--root', default='/data/DB/')
parser.add_argument('--batch_size', type=int, default=8, help='train batch size')
parser.add_argument('--input_size', type=int, default=1024, help='input size')
parser.add_argument('--workers', default=4, type=int, help='Number of workers')
parser.add_argument('--backbone', type=str, default='hourglass104_MRCB_cascade',
help='[hourglass104_MRCB_cascade, hourglass104_MRCB, hhrnet48, DLA_dcn, uesnet101_dcn]')
parser.add_argument('--dataset', type=str, default='DOTA', help='training dataset')
parser.add_argument('--epochs', type=int, default=120, help='number of train epochs')
parser.add_argument('--lr', type=float, default=2.5e-4, help='learning rate')
parser.add_argument('--print_freq', default=100, type=int, help='interval of showing training conditions')
parser.add_argument('--train_iter', default=0, type=int, help='number of total iterations for training')
parser.add_argument('--curr_iter', default=0, type=int, help='current iteration')
parser.add_argument('--save_path', type=str, default='./weight', help='Model save path')
parser.add_argument('--resume', default=None, type=str, help='training restore')
parser.add_argument('--data_split', default='1024_single', type=str, help='data split for DOTA')
parser.add_argument('--alpha', type=float, default=10, help='weight for positive loss, default=10')
parser.add_argument('--gamma', type=float, default=0.1, help='gamma for learninf rate decay')
parser.add_argument("--random_seed", type=int, default=1)
parser.add_argument('--sync_bn', action='store_true', help='enabling apex sync BN.')
parser.add_argument('--local_rank', type=int, default=0)
parser.add_argument('--amp', action='store_true', help='half precision')
args = parser.parse_args()
return args
def main():
args = parse()
# fixed seed
torch.manual_seed(args.random_seed)
np.random.seed(args.random_seed)
random.seed(args.random_seed)
if type(args.input_size) == int:
args.input_size = (args.input_size, args.input_size)
out_size = (args.input_size[0]//2, args.input_size[1]//2)
mean=(0.485,0.456,0.406)
var=(0.229,0.224,0.225)
""" initial parameters for training """
NUM_CLASSES = {'DOTA' : 18, 'HRSC2016' : 1}
num_classes = NUM_CLASSES[args.dataset]
""" cuda & distributed """
ngpus = torch.cuda.device_count()
if args.local_rank==0: print("ngpus : ", ngpus)
distributed = ngpus > 1
device = torch.device('cuda:{}'.format(args.local_rank))
if distributed:
#os.environ['MASTER_ADDR'] = '127.0.0.1'
#os.environ['MASTER_PORT'] = '99999'
torch.cuda.set_device(args.local_rank)
torch.distributed.init_process_group(
backend="nccl", init_method="env://",
rank=args.local_rank, world_size=torch.cuda.device_count()
)
model = Model_factory(args.backbone, num_classes)
if distributed and args.sync_bn:
model = nn.SyncBatchNorm.convert_sync_batchnorm(model)
if args.local_rank == 0: print("using synced BN")
torch.cuda.set_device(device)
model = model.to(device)
# Scale learning rate based on global batch size
optimizer = optim.Adam(model.parameters(), lr=args.lr)
if torch.distributed.get_world_size() > 1:
model = DistributedDataParallel(
model,
device_ids=[args.local_rank],
output_device=args.local_rank,
)
# define loss function (criterion) and optimizer
criterion = SWM_FPEM_Loss(num_classes=num_classes, alpha=args.alpha)
# Optionally resume from a checkpoint
if args.resume:
# Use a local scope to avoid dangling references
def resume():
if os.path.isfile(args.resume):
if args.local_rank == 0: print("=> loading checkpoint '{}'".format(args.resume))
state = torch.load(args.resume, map_location='cpu')
#model.module.load_state_dict(state['state_dict'], strict=True)
model.load_state_dict(state['model'], strict=True)
#optimizer.load_state_dict(state["optimizer"])
if args.local_rank == 0: print("=> loaded checkpoint ", args.resume)
else:
if args.local_rank == 0: print("=> no checkpoint found at '{}'".format(args.resume))
resume()
""" Get data loader """
transform_train = Augmentation(args.input_size, mean, var)
transform_valid = Augmentation_test(args.input_size, mean, var)
train_dataset = ListDataset(root=args.root, dataset=args.dataset, mode='train', split=args.data_split,
transform=transform_train, out_size=out_size)
valid_dataset = ListDataset(root=args.root, dataset=args.dataset, mode='val', split=args.data_split,
transform=transform_valid, out_size=out_size)
if args.local_rank == 0: print("number of train = %d / valid = %d" % (len(train_dataset), len(valid_dataset)))
train_sampler = torch.utils.data.distributed.DistributedSampler(train_dataset)
valid_sampler = torch.utils.data.distributed.DistributedSampler(valid_dataset)
train_loader = torch.utils.data.DataLoader(
train_dataset, batch_size=args.batch_size, shuffle=(train_sampler is None),
num_workers=args.workers, pin_memory=True, sampler=train_sampler, drop_last=True)
valid_loader = torch.utils.data.DataLoader(
valid_dataset, batch_size=args.batch_size, shuffle=False,
num_workers=args.workers, pin_memory=True, sampler=valid_sampler, drop_last=False)
""" lr scheduler """
args.train_iter = len(train_loader) * args.epochs
scheduler = WarmupPolyLR(
optimizer,
args.train_iter,
warmup_iters=1000,
power=0.90
)
if args.local_rank == 0: print(args)
best_loss = 999999999
best_dist = 999999999
start = time.time()
for epoch in range(0, args.epochs):
if distributed: train_sampler.set_epoch(epoch)
# train for one epoch
train(train_loader, model, criterion, optimizer, scheduler, device, start, epoch, args)
# evaluate on validation set
val_loss, val_dist = validate(valid_loader, model, criterion, device, epoch, args)
# save checkpoint
if args.local_rank == 0:
if best_loss >= val_loss:
best_loss = val_loss
save_checkpoint(model, optimizer, epoch, "best_loss", args.save_path)
if best_dist >= val_dist:
best_dist = val_dist
save_checkpoint(model, optimizer, epoch, "best_dist", args.save_path)
def train(train_loader, model, criterion, optimizer, scheduler, device, start, epoch, args):
batch_time = AverageMeter()
losses = AverageMeter()
# switch to train mode
model.train()
end = time.time()
world_size = torch.distributed.get_world_size()
scaler = torch.cuda.amp.GradScaler(enabled=args.amp)
for x, y, w, s in train_loader:
args.curr_iter += 1
x = x.to(device, non_blocking=True)
y = y.to(device, non_blocking=True)
w = w.to(device, non_blocking=True)
s = s.to(device, non_blocking=True)
with torch.cuda.amp.autocast(enabled=args.amp):
outs = model(x)
if type(outs) == list:
loss = 0
for out in outs:
loss += criterion(y, out, w, s)
loss /= len(outs)
outs = outs[-1]
else:
loss = criterion(y, outs, w, s)
# compute gradient and backward
optimizer.zero_grad()
scaler.scale(loss).backward()
scaler.step(optimizer)
scaler.update()
scheduler.step()
reduced_loss = reduce_tensor(loss.data, world_size)
losses.update(reduced_loss.item())
batch_time.update(time.time() - end)
end = time.time()
if args.local_rank == 0 and args.curr_iter % args.print_freq == 0:
train_log = "Epoch: [%d/%d][%d/%d] " % (epoch, args.epochs, args.curr_iter, args.train_iter)
train_log += "({0:.1f}%, {1:.1f} min) | ".format(args.curr_iter/args.train_iter*100, (end-start) / 60)
train_log += "Time %.1f ms | Left %.1f min | " % (batch_time.avg * 1000, (args.train_iter - args.curr_iter) * batch_time.avg / 60)
train_log += "Loss %.6f " % (losses.avg)
print(train_log)
def validate(valid_loader, model, criterion, device, epoch, args):
losses = AverageMeter()
distances = AverageMeter()
# switch to evaluate mode
model.eval()
world_size = torch.distributed.get_world_size()
end = time.time()
for x, y, w, s in valid_loader:
x = x.to(device, non_blocking=True)
y = y.to(device, non_blocking=True)
w = w.to(device, non_blocking=True)
s = s.to(device, non_blocking=True)
# compute output
with torch.no_grad():
outs = model(x)
if type(outs) == list:
outs = outs[-1]
loss = criterion(y, outs, w, s)
# measure accuracy and record loss
dist = torch.sqrt((y - outs)**2).mean()
reduced_loss = reduce_tensor(loss.data, world_size)
reduced_dist = reduce_tensor(dist.data, world_size)
losses.update(reduced_loss.item())
distances.update(reduced_dist.item())
if args.local_rank == 0:
valid_log = "\n============== validation ==============\n"
valid_log += "valid time : %.1f s | " % (time.time() - end)
valid_log += "valid loss : %.6f | " % (losses.avg)
valid_log += "valid dist : %.6f \n" % (distances.avg)
print(valid_log)
return losses.avg, distances.avg
def save_checkpoint(model, optimizer, epoch, name, save_path):
state = {
'model': model.module.state_dict(),
"optimizer": optimizer.state_dict(),
"epoch": epoch,
}
model_file = os.path.join(save_path, f"{name}.pt")
torch.save(state, model_file)
class AverageMeter(object):
"""Computes and stores the average and current value"""
def __init__(self):
self.reset()
def reset(self):
self.val = 0
self.avg = 0
self.sum = 0
self.count = 0
def update(self, val, n=1):
self.val = val
self.sum += val * n
self.count += n
self.avg = self.sum / self.count
def reduce_tensor(tensor, world_size):
rt = tensor.clone()
torch.distributed.all_reduce(rt, op=torch.distributed.ReduceOp.SUM)
rt /= world_size
return rt
if __name__ == '__main__':
main()