-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathtrain.py
160 lines (135 loc) · 6.5 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
# Licensed under the CC BY-NC 4.0 license (https://creativecommons.org/licenses/by-nc/4.0/)
import os
import yaml
import argparse
import pathlib
from tqdm import tqdm
import math
import random
import numpy as np
import torch
from torch.utils.data import DataLoader
from torch.optim import AdamW
from network.concretenet import ConcreteNet
from dataloader.factory import DatasetFactory
from utils.losses import JointLoss
from utils.loss_tracker import LossTracker
from utils.evaluator import Evaluator
parser = argparse.ArgumentParser()
parser.add_argument('--config_path', default='config/concretenet.yaml')
parser.add_argument('--evaluate', action='store_true')
parser.add_argument('--resume', action='store_true')
parser.add_argument('--ckpt_path', default=None)
parser.add_argument('--random_seed', default=None)
args = parser.parse_args()
if not args.random_seed:
random.seed(42)
np.random.seed(42)
torch.manual_seed(42)
torch.cuda.manual_seed_all(42)
class Trainer:
def __init__(self):
super().__init__()
config = yaml.safe_load(open(args.config_path, 'r'))
self.config = config
# Create dataloaders
train_dataset = DatasetFactory(split='train', **config['dataset'])
self.train_dataloader = DataLoader(train_dataset, collate_fn=train_dataset.collate_fn,
**config['train_dataloader'])
val_dataset = DatasetFactory(split='val', **config['dataset'])
self.val_dataloader = DataLoader(val_dataset, collate_fn=val_dataset.collate_fn,
**config['val_dataloader'])
# Build model, loss, optimizer and evaluator
self.model = ConcreteNet(**config['model']).cuda()
self.criterion = JointLoss(**config['loss']).cuda()
self.optimizer = AdamW(filter(lambda p: p.requires_grad, self.model.parameters()),
**config['optimizer'])
self.loss_tracker = LossTracker()
self.evaluator = Evaluator(**config['loss']['visual_config'])
self.start_epoch = 1 # 1-indexing epoch count
self.num_epochs = config['trainer']['num_epochs']
self.save_freq = config['trainer']['save_freq']
self.load_checkpoint(args.ckpt_path, args.resume)
def cosine_scheduler_step(self, epoch, epsilon=1e-6):
initial_lr = self.config['optimizer']['lr']
warmup_epochs = self.config['scheduler']['warmup_epochs']
lr = initial_lr if epoch < warmup_epochs else epsilon + 0.5 * (initial_lr - epsilon) * \
(1 + math.cos(math.pi * ((epoch - warmup_epochs) / (self.num_epochs - warmup_epochs))))
for param_group in self.optimizer.param_groups:
param_group['lr'] = lr
def decipher_stage(self, epoch):
epoch_thresholds = self.config['trainer']['pretraining']
stage = 0
if epoch > epoch_thresholds['semantic']: stage += 1
if epoch > epoch_thresholds['candidate']: stage += 1
if epoch > epoch_thresholds['instance']: stage += 1
return stage
def train_epoch(self, epoch, stage=None):
stage = stage if stage is not None else self.decipher_stage(epoch)
self.model.train()
self.loss_tracker.reset()
for batch in (pbar := tqdm(self.train_dataloader, total=len(self.train_dataloader))):
self.optimizer.zero_grad()
self.cosine_scheduler_step(epoch - 1)
batch = {k: v.cuda() if hasattr(v, 'cuda') else v for k, v in batch.items()}
stage = self.decipher_stage(epoch)
ret = self.model(batch, stage=stage)
loss_dict = self.criterion(batch, ret)
loss_dict['loss_total'].backward()
self.optimizer.step()
self.loss_tracker.update(loss_dict)
pbar.set_description(f'train:{epoch}/{self.num_epochs}({stage})' + self.loss_tracker.track())
self.save_checkpoint(epoch)
def eval_epoch(self, epoch):
self.model.eval()
self.loss_tracker.reset()
self.evaluator.reset()
for batch in (pbar := tqdm(self.val_dataloader, total=len(self.val_dataloader))):
batch = {k: v.cuda() if hasattr(v, 'cuda') else v for k, v in batch.items()}
with torch.no_grad():
ret = self.model(batch, stage=-1)
loss_dict = self.criterion(batch, ret)
decode_ret = self.model.decode_output(ret)
self.loss_tracker.update(loss_dict)
pbar.set_description(f'eval:{epoch}' + self.loss_tracker.track())
self.evaluator.update(batch, decode_ret)
self.evaluator.evaluate()
def train(self):
for epoch in range(self.start_epoch, self.num_epochs + 1):
stage = self.decipher_stage(epoch)
self.train_epoch(epoch, stage)
# if not epoch % self.save_freq:
# self.eval_epoch(epoch)
def evaluate(self):
self.eval_epoch(self.start_epoch - 1)
def save_checkpoint(self, epoch):
save_dir = self.config['trainer']['save_dir']
pathlib.Path(save_dir).mkdir(parents=True, exist_ok=True)
save_path = os.path.join(save_dir, f'ckpt_{epoch}.pth')
torch.save({'model': self.model.state_dict(),
'optimizer': self.optimizer.state_dict(),
'epoch': epoch}, save_path)
# Remove previous checkpoint if not within save frequency
prev_path = os.path.join(save_dir, f'ckpt_{epoch - 1}.pth')
if os.path.isfile(prev_path) and (epoch - 1) % self.save_freq:
os.remove(prev_path)
def load_checkpoint(self, load_path=None, resume=False):
# If load path isn't provided, find the last epoch in save_dir
if load_path is None and resume:
load_dir = self.config['trainer']['save_dir']
ckpt_epochs = [int(fn.split('_')[-1].split('.')[0]) for fn in
os.listdir(load_dir) if fn.endswith('.pth')]
if not ckpt_epochs:
return
load_path = os.path.join(load_dir, f'ckpt_{max(ckpt_epochs)}.pth')
assert os.path.exists(load_path)
if load_path is None:
return
print(f'Loading checkpoint at {load_path}.')
ckpt = torch.load(load_path, map_location=torch.device('cpu'))
self.model.load_state_dict(ckpt['model'], strict=False)
if resume:
self.optimizer.load_state_dict(ckpt['optimizer'])
self.start_epoch = ckpt['epoch'] + 1
if __name__=='__main__':
Trainer().evaluate() if args.evaluate else Trainer().train()