-
Notifications
You must be signed in to change notification settings - Fork 17
/
train.py
executable file
·141 lines (117 loc) · 4.85 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
import torch
import torch.optim as optim
from tensorboardX import SummaryWriter
import numpy as np
import os
import sys, importlib
import argparse
import time, datetime
import matplotlib; matplotlib.use('Agg')
from src import config, data
from src.checkpoints import CheckpointIO
from collections import defaultdict
import shutil
# Arguments
parser = argparse.ArgumentParser(
description='Train Implicit Autoencoder.'
)
parser.add_argument('config', type=str, help='Path to config file.')
args = parser.parse_args()
cfg = config.load_config(args.config, 'configs/default.yaml')
is_cuda = torch.cuda.is_available()
device = torch.device("cuda" if is_cuda else "cpu")
# Set t0
t0 = time.time()
# Shorthands
out_dir = cfg['training']['out_dir']
batch_size = cfg['training']['batch_size']
backup_every = cfg['training']['backup_every']
model_selection_metric = cfg['training']['model_selection_metric']
if cfg['training']['model_selection_mode'] == 'maximize':
model_selection_sign = 1
elif cfg['training']['model_selection_mode'] == 'minimize':
model_selection_sign = -1
else:
raise ValueError('model_selection_mode must be '
'either maximize or minimize.')
# Output directory
if not os.path.exists(out_dir):
os.makedirs(out_dir)
shutil.copyfile(args.config, os.path.join(out_dir, 'config.yaml'))
# Dataset
train_dataset = config.get_dataset('train', cfg)
val_dataset = config.get_dataset('val', cfg, return_idx=True)
train_loader = torch.utils.data.DataLoader(
train_dataset, batch_size=batch_size, num_workers=cfg['training']['n_workers'], shuffle=True,
collate_fn=data.collate_remove_none,
worker_init_fn=data.worker_init_fn)
val_loader = torch.utils.data.DataLoader(
val_dataset, batch_size=1, num_workers=cfg['training']['n_workers_val'], shuffle=False,
collate_fn=data.collate_remove_none,
worker_init_fn=data.worker_init_fn)
print('Loaded data.')
# Model
model = config.get_model(cfg, device=device, dataset=train_dataset)
# Intialize training
optimizer = optim.Adam(model.parameters(), lr=1e-4)
# optimizer = optim.SGD(model.parameters(), lr=1e-4, momentum=0.9)
trainer = config.get_trainer(model, optimizer, cfg, device=device)
checkpoint_io = CheckpointIO(out_dir, model=model, optimizer=optimizer)
try:
if cfg['training']['model_file']:
load_dict = checkpoint_io.load(cfg['training']['model_file'])
else:
load_dict = checkpoint_io.load('model.pt')
except FileExistsError:
print('No file load...')
load_dict = dict()
epoch_it = load_dict.get('epoch_it', 0)
it = load_dict.get('it', 0)
metric_val_best = load_dict.get(
'loss_val_best', -model_selection_sign * np.inf)
if metric_val_best == np.inf or metric_val_best == -np.inf:
metric_val_best = -model_selection_sign * np.inf
print('Current best validation metric (%s): %.8f'
% (model_selection_metric, metric_val_best))
logger = SummaryWriter(os.path.join(out_dir, 'logs'))
# Shorthands
print_every = cfg['training']['print_every']
checkpoint_every = cfg['training']['checkpoint_every']
validate_every = cfg['training']['validate_every']
print('output path: ', cfg['training']['out_dir'])
while True:
epoch_it += 1
for batch in train_loader:
it += 1
loss = trainer.train_step(batch)
logger.add_scalar('train/loss', loss, it)
# Print output
if print_every > 0 and (it % print_every) == 0:
t = datetime.datetime.now()
print('[Epoch %02d] it=%03d, loss=%.4f, time: %.2fs, %02d:%02d'
% (epoch_it, it, loss, time.time() - t0, t.hour, t.minute))
# Save checkpoint
if (checkpoint_every > 0 and (it % checkpoint_every) == 0):
print('Saving checkpoint')
checkpoint_io.save('model.pt', epoch_it=epoch_it, it=it,
loss_val_best=metric_val_best)
# Backup if necessary
if (backup_every > 0 and (it % backup_every) == 0):
print('Backup checkpoint')
checkpoint_io.save('model_%d.pt' % it, epoch_it=epoch_it, it=it,
loss_val_best=metric_val_best)
# Run validation
if validate_every > 0 and (it % validate_every) == 0:
#if 1:
eval_dict = trainer.evaluate(val_loader)
metric_val = eval_dict[model_selection_metric]
print('Validation metric (%s): %.4f'
% (model_selection_metric, metric_val))
for k, v in eval_dict.items():
logger.add_scalar('val/%s' % k, v, it)
if model_selection_sign * (metric_val - metric_val_best) > 0:
metric_val_best = metric_val
print('New best model (loss %.4f)' % metric_val_best)
checkpoint_io.save('model_best.pt', epoch_it=epoch_it, it=it,
loss_val_best=metric_val_best)
#import ipdb; ipdb.set_trace()