PreTraM/train.py at master · chenfengxu714/PreTraM · GitHub

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
import os
import argparse
import tensorflow as tf
import torch
from torch.utils.data import DataLoader
import numpy as np
from data_util import dataset_factory, collate_factory
from model import model_factory, loss_factory
from model.dlow import dlow_loss_factory
from utils.config import Config
from utils.utils import create_logger
from utils.lr_scheduler import LR_Scheduler
from run_util import trainer_factory
import logging
import random

def prepare_seed(rand_seed):
    np.random.seed(rand_seed)
    random.seed(rand_seed)
    torch.manual_seed(rand_seed)
    torch.cuda.manual_seed_all(rand_seed)

def set_cuda_visible_devices(devs):
    gpus = []
    for dev in devs.split(','):
        dev = dev.strip().lower()
        if dev == 'cpu':
            continue
        if dev.startswith('gpu'):
            dev = dev[3:]
        if '-' in dev:
            l, r = map(int, dev.split('-'))
            gpus.extend(range(l, r + 1))
        else:
            gpus.append(int(dev))

    os.environ['CUDA_VISIBLE_DEVICES'] = ','.join([str(gpu) for gpu in gpus])
    return gpus

def train(num_epoch, trainer, log_dir, logger, model_save_freq=5, save_after_epoch=20):
	loss_best = 1e4
	for i in range(num_epoch):
		logger.info('------------Epoch {0}------------'.format(i))
		avg_loss = trainer.train_epoch()
		if avg_loss<loss_best:
			trainer.save_state(log_dir,i,best=True)
			loss_best = avg_loss
		elif i>save_after_epoch:
			trainer.save_state(log_dir,i)

if __name__ == '__main__':
	parser = argparse.ArgumentParser()
	parser.add_argument('--cfg', required=True)
	parser.add_argument('--ckpt_path', type=str, help='finetuning phase')
	parser.add_argument('--log_dir', required=True)
	parser.add_argument('--model_path', type=str, help='Only apply when training dlow,\
	 	load the state_dict per this path before training, overriding cfg')
	parser.add_argument('--save_after_epoch', type=int, default=90)
	parser.add_argument('--debug', action='store_true')

	args = parser.parse_args()

	cfg = Config(args.cfg, tmp=False, create_dirs=True)

	if args.debug:
		cfg.num_workers = 0
		cfg.debug = True
		cfg.batch_size = 2
	if cfg.model_id=='dlow':
		if os.path.isdir(args.model_path):
			model_path = os.path.join(args.model_path, 'model_best.pth')
			if os.path.exists(model_path):
				cfg.model_path = model_path
		elif os.path.exists(args.model_path):
			cfg.model_path = args.model_path

	seed = cfg.seed
	prepare_seed(seed)
	dataset_name = cfg.dataset
	args.log_dir = os.path.join('results',args.log_dir)
	logger = create_logger(args.log_dir)
	log_fd = open(os.path.join(args.log_dir, 'train_dataset_log.txt'),'w')
	train_dataset = dataset_factory[dataset_name](cfg, log_fd, split='train', phase='training')
	log_fd.close()
	log_fd = open(os.path.join(args.log_dir, 'val_dataset_log.txt'),'w')
	val_dataset = dataset_factory[dataset_name](cfg, log_fd, split='val', phase='training')
	log_fd.close()

	dynamic_padding = cfg.get('dynamic_padding', False)
	if dynamic_padding:
		collate_fn = collate_factory['dynamic_padding']
	else:
		collate_fn = None
	train_data_loader = DataLoader(
                        train_dataset, shuffle=True, batch_size=cfg.batch_size,
			num_workers=cfg.num_workers, pin_memory=True,
			worker_init_fn=lambda worker_id: np.random.seed(seed + worker_id),
			collate_fn=collate_fn
		)

	device = 'cuda'

	model = model_factory[cfg.model_id](cfg)
	print(model)
	if args.ckpt_path:
		print(args.ckpt_path)
		model_path = os.path.join(args.ckpt_path)
		model_dict = {}
		model_cp = torch.load(model_path)
		for key in model_cp['model'].keys():
			if 'moco' in key or "mask_" in key:
				continue
			model_dict[key] = model_cp['model'][key]
		model.load_state_dict(model_dict, strict=False)
		print('load success')

	if cfg.model_id == 'dlow':
		loss_factory = dlow_loss_factory

	loss_names = cfg.loss_cfg.keys()
	loss_criterions = {n:loss_factory[n](cfg) for n in loss_names}
	model.cuda()

	optimizer = torch.optim.Adam(model.parameters(), lr=cfg.lr, weight_decay=cfg.weight_decay)
	if 'clip' in cfg.loss_cfg and cfg.loss_cfg['clip']['learn_temperature'] and cfg.loss_cfg['clip'].do:
		optimizer.add_param_group({'params':[loss_criterions['clip'].temperature_clip], 'lr':1e-6})
	if 'moco' in cfg.loss_cfg and cfg.loss_cfg['moco']['learn_temperature'] and cfg.map_moco:
		optimizer.add_param_group({'params':[loss_criterions['moco'].temperature_moco], 'lr':1e-6})

	steps_per_epoch = len(train_data_loader)
	if cfg.lr_scheduler == 'warmup':
		steps_per_udpate = 1
	else:
		steps_per_udpate = len(train_data_loader)
	lr_scheduler = LR_Scheduler(cfg, optimizer, steps_per_udpate, steps_per_epoch)

	logger.info(cfg.yml_dict)

	trainer = trainer_factory[cfg.model_id](cfg, model, device, train_data_loader, \
			loss_criterions, optimizer, lr_scheduler, logger=logger, print_freq=100)

	train(cfg.num_epochs, trainer, args.log_dir, logger, model_save_freq=cfg.model_save_freq,
			save_after_epoch=args.save_after_epoch)