-
Notifications
You must be signed in to change notification settings - Fork 35
/
train_stage1.py
190 lines (158 loc) · 7.84 KB
/
train_stage1.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
from __future__ import print_function, division
import logging
import numpy as np
import cv2
import os
from pathlib import Path
from tqdm import tqdm
from datetime import datetime
from lib.human_loader import StereoHumanDataset
from lib.network import RtStereoHumanModel
from config.stereo_human_config import ConfigStereoHuman as config
from lib.train_recoder import Logger, file_backup
from lib.utils import get_novel_calib_for_show as get_novel_calib
from lib.TaichiRender import TaichiRenderBatch
import torch
import torch.optim as optim
from torch.cuda.amp import GradScaler
from torch.utils.data import DataLoader
import warnings
warnings.filterwarnings("ignore", category=UserWarning)
class Trainer:
def __init__(self, cfg_file):
self.cfg = cfg_file
self.model = RtStereoHumanModel(self.cfg, with_gs_render=False)
self.train_set = StereoHumanDataset(self.cfg.dataset, phase='train')
self.train_loader = DataLoader(self.train_set, batch_size=self.cfg.batch_size, shuffle=True,
num_workers=self.cfg.batch_size*2, pin_memory=True)
self.train_iterator = iter(self.train_loader)
self.val_set = StereoHumanDataset(self.cfg.dataset, phase='val')
self.val_loader = DataLoader(self.val_set, batch_size=2, shuffle=False, num_workers=4, pin_memory=True)
self.len_val = int(len(self.val_loader) / self.val_set.val_boost) # real length of val set
self.val_iterator = iter(self.val_loader)
self.optimizer = optim.AdamW(self.model.parameters(), lr=self.cfg.lr, weight_decay=self.cfg.wdecay, eps=1e-8)
self.scheduler = optim.lr_scheduler.OneCycleLR(self.optimizer, self.cfg.lr, 100100, pct_start=0.01,
cycle_momentum=False, anneal_strategy='linear')
self.logger = Logger(self.scheduler, cfg.record)
self.total_steps = 0
self.model.cuda()
if self.cfg.restore_ckpt:
self.load_ckpt(self.cfg.restore_ckpt)
self.model.train()
self.model.raft_stereo.freeze_bn()
self.scaler = GradScaler(enabled=self.cfg.raft.mixed_precision)
self.render = TaichiRenderBatch(bs=1, res=self.cfg.dataset.src_res)
def train(self):
for _ in tqdm(range(self.total_steps, self.cfg.num_steps)):
self.optimizer.zero_grad()
data = self.fetch_data(phase='train')
# Raft Stereo
_, flow_loss, metrics = self.model(data)
loss = flow_loss
if self.total_steps and self.total_steps % self.cfg.record.loss_freq == 0:
self.logger.writer.add_scalar(f'lr', self.optimizer.param_groups[0]['lr'], self.total_steps)
self.save_ckpt(save_path=Path('%s/%s_latest.pth' % (cfg.record.ckpt_path, cfg.name)), show_log=False)
self.logger.push(metrics)
self.scaler.scale(loss).backward()
self.scaler.unscale_(self.optimizer)
torch.nn.utils.clip_grad_norm_(self.model.parameters(), 1.0)
self.scaler.step(self.optimizer)
self.scheduler.step()
self.scaler.update()
if self.total_steps and self.total_steps % self.cfg.record.eval_freq == 0:
self.model.eval()
self.run_eval()
self.model.train()
self.model.raft_stereo.freeze_bn()
self.total_steps += 1
print("FINISHED TRAINING")
self.logger.close()
self.save_ckpt(save_path=Path('%s/%s_final.pth' % (cfg.record.ckpt_path, cfg.name)))
def run_eval(self):
logging.info(f"Doing validation ...")
torch.cuda.empty_cache()
epe_list, one_pix_list = [], []
show_idx = np.random.choice(list(range(self.len_val)), 1)
for idx in range(self.len_val):
data = self.fetch_data(phase='val')
with torch.no_grad():
data, _, _ = self.model(data, is_train=False)
if idx == show_idx:
data = get_novel_calib(data, ratio=0.5)
data = self.render.flow2render(data)
tmp_novel = data['novel_view']['img_pred'][0].detach()
tmp_novel = (tmp_novel / 2.0 + 0.5) * 255
tmp_novel = tmp_novel.permute(1, 2, 0).cpu().numpy()
tmp_img_name = '%s/%s.jpg' % (cfg.record.show_path, self.total_steps)
cv2.imwrite(tmp_img_name, tmp_novel[:, :, ::-1].astype(np.uint8))
for view in ['lmain', 'rmain']:
valid = (data[view]['valid'] >= 0.5)
epe = torch.sum((data[view]['flow'] - data[view]['flow_pred']) ** 2, dim=1).sqrt()
epe = epe.view(-1)[valid.view(-1)]
one_pix = (epe < 1)
epe_list.append(epe.mean().item())
one_pix_list.append(one_pix.float().mean().item())
val_epe = np.round(np.mean(np.array(epe_list)), 4)
val_one_pix = np.round(np.mean(np.array(one_pix_list)), 4)
logging.info(f"Validation Metrics ({self.total_steps}): epe {val_epe}, 1pix {val_one_pix}")
self.logger.write_dict({'val_epe': val_epe, 'val_1pix': val_one_pix}, write_step=self.total_steps)
torch.cuda.empty_cache()
def fetch_data(self, phase):
if phase == 'train':
try:
data = next(self.train_iterator)
except:
self.train_iterator = iter(self.train_loader)
data = next(self.train_iterator)
elif phase == 'val':
try:
data = next(self.val_iterator)
except:
self.val_iterator = iter(self.val_loader)
data = next(self.val_iterator)
for view in ['lmain', 'rmain']:
for item in data[view].keys():
data[view][item] = data[view][item].cuda()
return data
def load_ckpt(self, load_path, load_optimizer=True, strict=True):
assert os.path.exists(load_path)
logging.info(f"Loading checkpoint from {load_path} ...")
ckpt = torch.load(load_path, map_location='cuda')
self.model.load_state_dict(ckpt['network'], strict=strict)
logging.info(f"Parameter loading done")
if load_optimizer:
self.total_steps = ckpt['total_steps'] + 1
self.logger.total_steps = self.total_steps
self.optimizer.load_state_dict(ckpt['optimizer'])
self.scheduler.load_state_dict(ckpt['scheduler'])
logging.info(f"Optimizer loading done")
def save_ckpt(self, save_path, show_log=True):
if show_log:
logging.info(f"Save checkpoint to {save_path} ...")
torch.save({
'total_steps': self.total_steps,
'network': self.model.state_dict(),
'optimizer': self.optimizer.state_dict(),
'scheduler': self.scheduler.state_dict()
}, save_path)
if __name__ == '__main__':
logging.basicConfig(level=logging.INFO,
format='%(asctime)s %(levelname)-8s [%(filename)s:%(lineno)d] %(message)s')
cfg = config()
cfg.load("config/stage1.yaml")
cfg = cfg.get_cfg()
cfg.defrost()
dt = datetime.today()
cfg.exp_name = '%s_%s%s' % (cfg.name, str(dt.month).zfill(2), str(dt.day).zfill(2))
cfg.record.ckpt_path = "experiments/%s/ckpt" % cfg.exp_name
cfg.record.show_path = "experiments/%s/show" % cfg.exp_name
cfg.record.logs_path = "experiments/%s/logs" % cfg.exp_name
cfg.record.file_path = "experiments/%s/file" % cfg.exp_name
cfg.freeze()
for path in [cfg.record.ckpt_path, cfg.record.show_path, cfg.record.logs_path, cfg.record.file_path]:
Path(path).mkdir(exist_ok=True, parents=True)
file_backup(cfg.record.file_path, cfg, train_script=os.path.basename(__file__))
torch.manual_seed(1314)
np.random.seed(1314)
trainer = Trainer(cfg)
trainer.train()