-
Notifications
You must be signed in to change notification settings - Fork 7
/
train_nuscenes.py
307 lines (273 loc) · 11 KB
/
train_nuscenes.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
#!/usr/bin/env python3
import os
import json
import argparse
import datetime
import random
import shutil
import torch
import torch.distributed as dist
import torch.multiprocessing as mp
from torch.nn.parallel import DistributedDataParallel as DDP
import numpy as np
from tensorboardX import SummaryWriter
from models.gfnet import GFNet
from libs.dataloader.nuScenes import Nuscenes
from libs.utils.training import train_epoch, validate
from libs.utils.sampler import DistributedEvalSampler
from libs.utils.cosine_schedule import CosineAnnealingWarmUpRestarts
from libs.utils.ohem import OhemCrossEntropy
from libs.utils.tools import (create_log, load_arch_cfg, load_data_cfg,
load_pretrained, recording_cfg,
get_weight_per_class, save_checkpoint,
resume_training, find_free_port)
seed = 6
random.seed(seed)
np.random.seed(seed)
torch.manual_seed(seed)
best_iou = 0.0
def parse_args():
parser = argparse.ArgumentParser(description='Geometric Flow Network for 3D Point Clouds Semantic Segmentation')
parser.add_argument(
'--dataset', '-d',
type=str,
required=False,
default='dataset/nuScenes/full/',
help='Dataset to train with.',
)
parser.add_argument(
'--pkl_train',
type=str,
required=False,
default='dataset/nuScenes/nuscenes_train.pkl',
help='pkl file containing train info',
)
parser.add_argument(
'--pkl_val',
type=str,
required=False,
default='dataset/nuScenes/nuscenes_val.pkl',
help='pkl file containing val info',
)
parser.add_argument(
'--arch_cfg', '-ac',
type=str,
required=False,
default='configs/resnet_nuscenes.yaml',
help='Architecture yaml cfg file. See /config/arch for sample.',
)
parser.add_argument(
'--data_cfg', '-dc',
type=str,
required=False,
default='configs/nuscenes.yaml',
help='Classification yaml cfg file. See /config/labels for sample.',
)
parser.add_argument(
'--log', '-l',
type=str,
default='logs/' +
datetime.datetime.now().strftime("%Y-%-m-%d-%H-%M-%S") + '/',
help='Directory to put the log data. Default: ~/logs/date+time'
)
parser.add_argument(
'--pretrained', '-p',
type=str,
required=False,
default=None,
help='Directory to get the pretrained model. If not passed, do from scratch!'
)
parser.add_argument(
'--resume', '-r',
type=str,
required=False,
default=None,
help='Directory to resume the checkpoint.'
)
parser.add_argument(
'--debug',
type=int,
required=False,
default=1,
help='whether debug'
)
parser.add_argument(
'--dist_backend',
type=str,
required=False,
default='nccl',
help='backend'
)
parser.add_argument(
'--dist_url',
type=str,
required=False,
default='tcp://127.0.0.1:8081',
help='init method'
)
parser.add_argument(
'--gpus', '-g',
type=str,
required=True,
default='0',
help='gpus to use'
)
FLAGS, unparsed = parser.parse_known_args()
return FLAGS
def main():
FLAGS = parse_args()
ARCH = load_arch_cfg(FLAGS.arch_cfg)
DATA = load_data_cfg(FLAGS.data_cfg)
os.environ['CUDA_VISIBLE_DEVICES'] = FLAGS.gpus
torch.backends.cudnn.benchmark = True
gpus = [int(i) for i in FLAGS.gpus.split(',')]
port = find_free_port()
FLAGS.dist_url = 'tcp://localhost:{}'.format(port)
try:
mp.set_start_method('spawn')
except RuntimeError:
pass
mp.set_sharing_strategy('file_system')
nprocs = len(gpus)
mp.spawn(main_worker, args=(FLAGS, ARCH, DATA, nprocs), nprocs=nprocs, join=True, daemon=False)
def main_worker(rank, FLAGS, ARCH, DATA, world_size):
global best_iou
dist.init_process_group(backend=FLAGS.dist_backend,
init_method=FLAGS.dist_url,
world_size=world_size,
rank=rank)
torch.cuda.set_device(rank)
writer = tb_dir = log_path = None
if rank == 0:
global logger
logger, log_path, tb_dir = create_log(FLAGS.log, FLAGS.data_cfg, FLAGS.debug)
recording_cfg(FLAGS.arch_cfg, FLAGS.data_cfg, log_path)
writer = SummaryWriter(log_dir=tb_dir, flush_secs=20)
# print summary of what we will do
logger.info("----------")
logger.info("INTERFACE:")
logger.info("pwd: {}".format(os.getcwd()))
logger.info("dataset: {}".format(FLAGS.dataset))
logger.info("arch_cfg: {}".format(FLAGS.arch_cfg))
logger.info("data_cfg: {}".format(FLAGS.data_cfg))
logger.info("dist_url: {}".format(FLAGS.dist_url))
logger.info("log: {}".format(log_path))
logger.info("pretrained: {}".format(FLAGS.pretrained))
logger.info("gpus: {}".format(FLAGS.gpus))
logger.info("debug: {}".format(FLAGS.debug))
logger.info('Configs: \n' + json.dumps(ARCH, indent=4, sort_keys=True))
logger.info('Args: \n' + json.dumps(vars(FLAGS), indent=4, sort_keys=True))
train_dataset = Nuscenes(pkl_path=FLAGS.pkl_train,
data_path=FLAGS.dataset,
labels=DATA['labels_16'],
range_cfg=ARCH['range'],
polar_cfg=ARCH['polar'],
dataset_cfg=ARCH['dataset'],
color_map=DATA['color_map'],
learning_map=DATA['learning_map'],
split='train')
train_sampler = torch.utils.data.distributed.DistributedSampler(train_dataset)
train_loader = torch.utils.data.DataLoader(train_dataset,
batch_size=ARCH["train"]["batch_size"],
num_workers=ARCH["train"]["workers"],
sampler=train_sampler,
shuffle=(train_sampler is None),
pin_memory=True,
drop_last=True)
val_dataset = Nuscenes(pkl_path=FLAGS.pkl_val,
data_path=FLAGS.dataset,
labels=DATA['labels_16'],
range_cfg=ARCH['range'],
polar_cfg=ARCH['polar'],
dataset_cfg=ARCH['dataset'],
color_map=DATA['color_map'],
learning_map=DATA['learning_map'],
split='val')
val_sampler = DistributedEvalSampler(val_dataset)
val_loader = torch.utils.data.DataLoader(val_dataset,
batch_size=ARCH["train"]["batch_size"],
num_workers=ARCH["train"]["workers"],
sampler=val_sampler,
shuffle=False,
pin_memory=True,
drop_last=False)
n_class = train_dataset.nclasses
model = GFNet(ARCH,
layers=ARCH["backbone"]["layers"],
n_class=n_class-1,
flow=ARCH["train"]["flow"],
data_type=torch.float32)
if rank == 0:
logger.info(model)
# add SyncBN
if ARCH["train"]["syncbn"]:
model = torch.nn.SyncBatchNorm.convert_sync_batchnorm(model)
criterion = torch.nn.CrossEntropyLoss(ignore_index=255, reduction='mean').cuda()
criterion_3d = OhemCrossEntropy(ignore_index=255, thresh=0.9, min_kept=3500, weight=None).cuda()
if ARCH["train"]["optimizer"] == 'sgd':
optimizer = torch.optim.SGD(
model.parameters(), lr=ARCH["train"]["min_lr"], momentum=0.9, weight_decay=1e-4
)
lr_scheduler = CosineAnnealingWarmUpRestarts(
optimizer, T_0=len(train_loader)*ARCH["train"]["max_epochs"], T_mult=10,
eta_max=ARCH["train"]["max_lr"], T_up=len(train_loader)*ARCH["train"]["wup_epochs"], gamma=0.5
)
elif ARCH["train"]["optimizer"] == 'adam':
optimizer = torch.optim.Adam(model.parameters(), lr=ARCH["train"]["lr"])
lr_step = []
for step in ARCH["train"]["lr_step"]:
lr_step.append(step * len(train_loader))
lr_scheduler = torch.optim.lr_scheduler.MultiStepLR(optimizer=optimizer,
milestones=lr_step,
gamma=ARCH["train"]["lr_factor"])
else:
raise ValueError('unsopported optimizer type!')
# when the weight for range or polar is 0, then the network has unused params
unused = not (ARCH["train"]["flow"])
model = DDP(model.cuda(), device_ids=[rank], find_unused_parameters=unused)
# load pretrained
start_epoch = 0
if FLAGS.pretrained:
model, start_epoch, best_iou = load_pretrained(FLAGS.pretrained, model)
if FLAGS.resume:
model, optimizer, start_epoch, best_iou = resume_training(FLAGS.resume, model, optimizer)
lr_scheduler.step(start_epoch*len(train_loader))
for epoch in range(start_epoch, ARCH["train"]["max_epochs"]):
train_sampler.set_epoch(epoch)
train_epoch(epoch, train_loader, model,
criterion, criterion_3d, optimizer,
lr_scheduler,
ARCH["loss"],
ARCH["train"]["report_batch"],
ARCH["train"]["max_epochs"],
writer)
if 'trainval' in FLAGS.pkl_train:
flag = (epoch == 0 or epoch+1 == ARCH["train"]["max_epochs"])
else:
if epoch < int(0.75*ARCH["train"]["max_epochs"]):
flag = (epoch % 50 == 0)
else:
flag = (epoch % 10 == 0 or epoch+1 == ARCH["train"]["max_epochs"])
if flag:
miou = validate(epoch, val_loader,
model, criterion,
ARCH["loss"],
writer)
is_best = miou > best_iou
if miou > best_iou:
best_iou = miou
if rank == 0:
save_checkpoint(
{
'epoch': epoch,
'model': model.state_dict(),
'optimizer': optimizer.state_dict(),
'best_iou': best_iou
}, log_path, is_best
)
if rank == 0:
src = os.path.join(log_path, 'model_best.pth.tar')
if os.path.isfile(src):
shutil.move(src, os.path.join(log_path, 'model_best_{:.4f}.pth.tar'.format(best_iou)))
if __name__ == '__main__':
main()