-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathtrainer_pointgroup.py
198 lines (152 loc) · 6.83 KB
/
trainer_pointgroup.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
import os
import time
import argparse
import random
import numpy as np
import torch
import torch.nn as nn
from torch.utils.data import DataLoader
from torch.utils.tensorboard import SummaryWriter
from PointGroup.data.dataset_seg import Dataset
from PointGroup.model.pointgroup.pointgroup import PointGroup, model_fn_decorator
from PointGroup.util.config import get_parser
from PointGroup.util.utils import AverageMeter
def init():
os.makedirs(cfg.exp_dir, exist_ok=True)
os.system('cp {} {}'.format(cfg.config_seg, cfg.exp_dir))
# summary writer
global writer
writer = SummaryWriter(cfg.exp_dir)
# random seed
random.seed(cfg.manual_seed)
np.random.seed(cfg.manual_seed)
torch.manual_seed(cfg.manual_seed)
torch.cuda.manual_seed_all(cfg.manual_seed)
def worker_init_fn(worker_id):
np.random.seed(np.random.get_state()[1][0] + worker_id)
class TrainerPointGroup:
def __init__(self):
self.epoch = 1
self.best_train = 1e9
self.best_val = 1e9
self.train_data = Dataset(cfg, phase='train')
self.val_data = Dataset(cfg, phase='val')
self.train_loader = torch.utils.data.DataLoader(self.train_data, batch_size=cfg.batch_size,
shuffle=True, num_workers=cfg.n_workers,
pin_memory=False, drop_last=True,
worker_init_fn=worker_init_fn,
collate_fn=self.train_data.merge)
self.val_loader = torch.utils.data.DataLoader(self.val_data, batch_size=cfg.batch_size,
shuffle=True, num_workers=cfg.n_workers,
pin_memory=False, drop_last=False,
worker_init_fn=worker_init_fn,
collate_fn=self.val_data.merge)
self.model = PointGroup(cfg)
# self.model = nn.DataParallel(self.model)
self.model.cuda()
if cfg.optim == 'Adam':
self.optimizer = torch.optim.Adam(filter(lambda p: p.requires_grad, self.model.parameters()), lr=cfg.lr)
elif cfg.optim == 'SGD':
self.optimizer = torch.optim.SGD(filter(lambda p: p.requires_grad, self.model.parameters()), lr=cfg.lr,
weight_decay=cfg.weight_decay, momentum=cfg.momentum)
self.scheduler = torch.optim.lr_scheduler.MultiStepLR(self.optimizer, milestones=cfg.lr_milestones, gamma=0.1)
def train_loop(self):
iter_time = AverageMeter()
data_time = AverageMeter()
avg_loss = {}
model_fn = model_fn_decorator(cfg, test=False)
self.model.train()
start_epoch = time.time()
end = time.time()
for iter, batch in enumerate(self.train_loader):
data_time.update(time.time() - end)
loss, preds, visual_dict, meter_dict = model_fn(batch, self.model, self.epoch)
for k, v in visual_dict.items():
if k not in avg_loss.keys():
avg_loss[k] = AverageMeter()
avg_loss[k].update(v)
self.optimizer.zero_grad()
loss.backward()
self.optimizer.step()
iter_time.update(time.time() - end)
end = time.time()
current_iter = (self.epoch - 1) * len(self.train_loader) + iter + 1
max_iter = cfg.epochs * len(self.train_loader)
remain_iter = max_iter - current_iter
remain_time = remain_iter * iter_time.avg
t_m, t_s = divmod(remain_time, 60)
t_h, t_m = divmod(t_m, 60)
remain_time = '{:02d}:{:02d}:{:02d}'.format(int(t_h), int(t_m), int(t_s))
print("epoch: {}/{} iter: {}/{} loss: {:.6f} semantic_loss: {:.6f} offset_norm_loss: {:.6f} data_time: {:.2f}({:.2f}) iter_time: {:.2f}({:.2f}) "
"remain_time: {remain_time}".format(
self.epoch, cfg.epochs, iter+1, len(self.train_loader), avg_loss['loss'].val, avg_loss['semantic_loss'].val,
avg_loss['offset_norm_loss'].val, data_time.val, data_time.avg, iter_time.val,
iter_time.avg, remain_time=remain_time))
if (iter == len(self.train_loader) - 1): print()
print("epoch: {}/{}, train loss: {:.6f}, time: {}s".format(self.epoch, cfg.epochs, avg_loss['loss'].avg,
time.time() - start_epoch))
if self.epoch % cfg.save_freq == 0:
self.checkpoint_save(phase='train')
if avg_loss['loss'].avg < self.best_train:
self.best_train = avg_loss['loss'].avg
self.checkpoint_save(phase='train', f='best_train')
for k in avg_loss.keys():
writer.add_scalar(f'Train/{k}', avg_loss[k].avg, self.epoch)
def checkpoint_save(self, phase, f=''):
assert phase in ['train', 'val']
checkpoint_data = {
'epoch': self.epoch,
'state_dict': self.model.state_dict(),
'best_res': self.best_train if phase == 'train' else self.best_val
}
if len(f) == 0:
save_path = f'{cfg.exp_dir}/pointgroup_{self.epoch:03d}.pth.tar'
else:
save_path = f'{cfg.exp_dir}/pointgroup_{f}.pth.tar'
torch.save(checkpoint_data, save_path, _use_new_zipfile_serialization=False)
def val_loop(self):
model_fn = model_fn_decorator(cfg, test=False)
avg_loss = {}
with torch.no_grad():
self.model.eval()
start_epoch = time.time()
for iter, batch in enumerate(self.val_loader):
loss, preds, visual_dict, meter_dict = model_fn(batch, self.model, self.epoch)
for k, v in visual_dict.items():
if k not in avg_loss.keys():
avg_loss[k] = AverageMeter()
avg_loss[k].update(v)
# if iter % max(1, len(self.val_loader) // 10) == 0:
# print('epoch={}, {}/{}, val_loss={}'.format(self.epoch, iter, len(self.val_loader), loss.item()))
print("iter: {}/{} loss: {:.6f}({:.6f})".format(iter+1, len(self.val_loader), avg_loss['loss'].val,
avg_loss['loss'].avg))
if (iter == len(self.val_loader) - 1): print()
print("epoch: {}/{}, val loss: {:.6f}, time: {}s".format(self.epoch, cfg.epochs, avg_loss['loss'].avg,
time.time() - start_epoch))
if avg_loss['loss'].avg < self.best_val:
self.best_val = avg_loss['loss'].avg
self.checkpoint_save(phase='val', f='best_val')
for k in avg_loss.keys():
writer.add_scalar(f'Val/{k}', avg_loss[k].avg, self.epoch)
def train(self):
for self.epoch in range(cfg.epochs+1):
print(f'>>>>>>>>>> Training Epoch: {self.epoch}/{cfg.epochs} <<<<<<<<<<')
self.train_loop()
print(">>>>>>>>>> End <<<<<<<<<<")
print(f'>>>>>>>>>> Start Evaluation <<<<<<<<<<')
self.val_loop()
print(">>>>>>>>>> End <<<<<<<<<<")
self.scheduler.step()
torch.cuda.empty_cache()
if __name__ == "__main__":
# parser = argparse.ArgumentParser()
# parser.add_argument('--dataset', type=str, default='Sileance_Dataset', help='indicate dataset name')
# parser.add_argument('--obj_name', type=str, default='gear', help='indicate object name')
# args = parser.parse_args()
# dataset = args.dataset
# obj_name = args.obj_name
cfg = get_parser()
init()
print(cfg)
trainer = TrainerPointGroup()
trainer.train()