-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathtrain.py
202 lines (165 loc) · 7.78 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
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
199
200
201
202
import numpy as np
import os
import time
import torch
import torch.nn as nn
from argparse import ArgumentParser
from datasets import Train_Dataset, Eval_Dataset
from models import HDRnetModel
from torch.optim import lr_scheduler, AdamW
from torch.utils.data import DataLoader
from torchvision.utils import save_image
from utils import psnr, print_params, load_train_ckpt, save_model_stats, AvgMeter
import random
from tqdm import tqdm
def train(params, train_loader, valid_loader, model, start_epoch=0):
nodecay = 20
# Optimization
optimizer = AdamW(model.parameters(), params['learning_rate'], weight_decay=1e-8)
# # Learning rate adjustment
scheduler = lr_scheduler.LinearLR(optimizer,
start_factor=1.0, end_factor=0.0, total_iters=params['epochs'] - nodecay, verbose=True)
# scheduler = lr_scheduler.CosineAnnealingLR(optimizer, T_max=params['epochs'] - nodecay, eta_min=1e-6, verbose=True)
# Loss function
criterion_l1 = nn.L1Loss()
# Training
train_loss_meter = AvgMeter()
train_psnr_meter = AvgMeter()
stats = {'train_loss': [],
'train_psnr': [],
'valid_psnr': []}
iteration = 0
old_time = time.time()
for epoch in range(start_epoch, params['epochs']):
for batch_idx, (low, full, target) in enumerate(train_loader):
iteration += 1
model.train()
low = low.to(device)
full = full.to(device)
target = target.to(device)
# Normalize to [0, 1] on GPU
if params['hdr']:
low = torch.div(low, 65535.0)
full = torch.div(full, 65535.0)
if params['ppr']:
low = torch.div(low, 255.0)
full = torch.div(full, 65535.0)
else:
low = torch.div(low, 255.0)
full = torch.div(full, 255.0)
target = torch.div(target, 255.0)
output = model(low, full)
loss = 20 * criterion_l1(output, target)
optimizer.zero_grad()
loss.backward()
optimizer.step()
if iteration % params['summary_interval'] == 0:
save_image(target, 'target.jpg')
save_image(full, 'full.jpg')
save_image(output, 'output.jpg')
train_loss_meter.update(loss.item())
train_psnr = psnr(output, target).item()
train_psnr_meter.update(train_psnr)
new_time = time.time()
print('[%d/%d] Iteration: %d | Loss: %.4f | PSNR: %.4f | Time: %.2fs' %
(epoch+1, params['epochs'], iteration, loss, train_psnr, new_time-old_time))
old_time = new_time
if iteration % params['ckpt_interval'] == 0:
stats['train_loss'].append(train_loss_meter.avg)
train_loss_meter.reset()
stats['train_psnr'].append(train_psnr_meter.avg)
train_psnr_meter.reset()
valid_psnr = eval(params, valid_loader, model, device)
stats['valid_psnr'].append(valid_psnr)
ckpt_fname = "epoch_" + str(epoch+1)+'_iter_' + str(iteration) + ".pth"
save_model_stats(model, params, ckpt_fname, stats)
if epoch > nodecay:
scheduler.step()
def eval(params, valid_loader, model, device):
model.eval()
psnr_meter = AvgMeter()
with torch.no_grad():
for (low, full, target, fname) in tqdm(valid_loader):
low = low.to(device)
full = full.to(device)
target = target.to(device)
# Normalize to [0, 1] on GPU
if params['hdr']:
low = torch.div(low, 65535.0)
full = torch.div(full, 65535.0)
if params['ppr']:
low = torch.div(low, 255.0)
full = torch.div(full, 65535.0)
else:
low = torch.div(low, 255.0)
full = torch.div(full, 255.0)
target = torch.div(target, 255.0)
output = model(low, full)
# save_image(output, os.path.join(params['eval_out'], fname[0]))
eval_psnr = psnr(output, target).item()
psnr_meter.update(eval_psnr)
print ("Validation PSNR: ", psnr_meter.avg)
return psnr_meter.avg
def parse_args():
parser = ArgumentParser(description='HDRnet training')
# Training, logging and checkpointing parameters
parser.add_argument('--cuda', action='store_true', help='Use CUDA')
parser.add_argument('--ckpt_interval', default=int(8876/4*10), type=int, help='Interval for saving checkpoints, unit is iteration')
parser.add_argument('--ckpt_dir', default='./ckpts', type=str, help='Checkpoint directory')
parser.add_argument('--stats_dir', default='./stats', type=str, help='Statistics directory')
parser.add_argument('--epochs', default=100, type=int)
parser.add_argument('-lr', '--learning_rate', default=1e-4, type=float)
parser.add_argument('--summary_interval', default=100, type=int)
parser.add_argument('--seed', default=0, type=int)
# Data pipeline and data augmentation
parser.add_argument('--batch_size', default=4, type=int, help='Size of a mini-batch')
parser.add_argument('--train_data_dir', type=str, required=True, help='Dataset path')
parser.add_argument('--eval_data_dir', default=None, type=str, help='Directory with the validation data.')
parser.add_argument('--eval_out', default='./outputs', type=str, help='Validation output path')
parser.add_argument('--hdr', action='store_true', help='Handle HDR image')
parser.add_argument('--ppr', action='store_true', help='Handle ppr image')
# Model parameters
parser.add_argument('--batch_norm', action='store_true', help='Use batch normalization')
parser.add_argument('--input_res', default=256, type=int, help='Resolution of the down-sampled input')
parser.add_argument('--output_res', default=(512, 512), type=int, nargs=2, help='Resolution of the guidemap/final output')
parser.add_argument('--gpu_ids', type=str, default='0', help='gpu ids: e.g. 0 0,1,2, 0,2. use -1 for CPU')
parser.add_argument('--init_type', type=str, default='normal', help='network initialization [normal|xavier|kaiming|orthogonal]')
parser.add_argument('--init_gain', type=float, default=0.02, help='scaling factor for normal, xavier and orthogonal.')
return parser.parse_args()
def set_seed(seed):
torch.manual_seed(seed)
torch.cuda.manual_seed_all(seed)
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False
np.random.seed(seed)
random.seed(seed)
os.environ['PYTHONHASHSEED'] = str(seed)
if __name__ == '__main__':
# Parse training parameters
params = vars(parse_args())
print_params(params)
# Random seeds
if params['seed'] > 0:
set_seed(params['seed'])
print('Set random seed %s' % params['seed'])
# Folders
os.makedirs(params['ckpt_dir'], exist_ok=True)
os.makedirs(params['stats_dir'], exist_ok=True)
os.makedirs(params['eval_out'], exist_ok=True)
if params['gpu_ids']:
device = torch.device('cuda:{}'.format(params['gpu_ids'][0]))
else:
device = torch.device("cpu")
# Dataloader for training
train_dataset = Train_Dataset(params)
train_loader = DataLoader(train_dataset, batch_size=params['batch_size'], shuffle=True, num_workers=8, pin_memory=True)
# Dataloader for validation
valid_dataset = Eval_Dataset(params)
valid_loader = DataLoader(valid_dataset, batch_size=1, num_workers=8, pin_memory=True)
# Model for training
model = HDRnetModel(params)
print(model)
start = load_train_ckpt(model, params['ckpt_dir'])
start = start if start else 0
model.to(device)
train(params, train_loader, valid_loader, model, start_epoch=start)