-
Notifications
You must be signed in to change notification settings - Fork 9
/
train_OneRestore_multi-gpu.py
153 lines (121 loc) · 7.6 KB
/
train_OneRestore_multi-gpu.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
import os, time, torch, argparse
import torch.nn.functional as F
from torch.utils.data import DataLoader
from torchvision.utils import save_image as imwrite
import numpy as np
from torchvision import transforms
from makedataset import Dataset
from utils.utils import print_args, load_restore_ckpt_with_optim, load_embedder_ckpt, adjust_learning_rate, data_process, tensor_metric, load_excel, save_checkpoint
from model.loss import Total_loss
from model.Embedder import Embedder
from model.OneRestore import OneRestore
from torch.utils.data.distributed import DistributedSampler
from PIL import Image
torch.distributed.init_process_group(backend="nccl")
local_rank = torch.distributed.get_rank()
torch.cuda.set_device(local_rank)
device = torch.device("cuda", local_rank)
transform_resize = transforms.Compose([
transforms.Resize([224,224]),
transforms.ToTensor()
])
def main(args):
print('> Model Initialization...')
embedder = load_embedder_ckpt(device, freeze_model=True, ckpt_name=args.embedder_model_path)
restorer, optimizer, cur_epoch = load_restore_ckpt_with_optim(device, local_rank=local_rank, freeze_model=False, ckpt_name=args.restore_model_path, lr=args.lr)
loss = Total_loss(args)
print('> Loading dataset...')
data = Dataset(args.train_input)
dataset = DataLoader(dataset=data, batch_size=args.bs,
shuffle=False,
num_workers=args.num_works,
pin_memory=True,drop_last=False,
sampler=DistributedSampler(data,shuffle=True))
print('> Start training...')
start_all = time.time()
train(restorer, embedder, optimizer, loss, cur_epoch, args, dataset, device)
end_all = time.time()
print('Whloe Training Time:' +str(end_all-start_all)+'s.')
def train(restorer, embedder, optimizer, loss, cur_epoch, args, dataset, device):
metric = []
for epoch in range(cur_epoch, args.epoch):
optimizer = adjust_learning_rate(optimizer, epoch, args.adjust_lr)
learnrate = optimizer.param_groups[-1]['lr']
restorer.train()
for i, data in enumerate(dataset,0):
pos, inp, neg = data_process(data, args, device)
text_embedding,_,_ = embedder(inp[1],'text_encoder')
out = restorer(inp[0], text_embedding)
restorer.zero_grad()
total_loss = loss(inp, pos, neg, out)
total_loss.backward()
optimizer.step()
mse = tensor_metric(pos,out, 'MSE', data_range=1)
psnr = tensor_metric(pos,out, 'PSNR', data_range=1)
ssim = tensor_metric(pos,out, 'SSIM', data_range=1)
print("[epoch %d][%d/%d] lr :%f Floss: %.4f MSE: %.4f PSNR: %.4f SSIM: %.4f"%(epoch+1, i+1, \
len(dataset), learnrate, total_loss.item(), mse, psnr, ssim))
psnr_t1, ssim_t1, psnr_t2, ssim_t2 = test(args, restorer, embedder, device, epoch)
metric.append([psnr_t1, ssim_t1, psnr_t2, ssim_t2])
print("[epoch %d] Test images PSNR1: %.4f SSIM1: %.4f"%(epoch+1, psnr_t1,ssim_t1))
load_excel(metric)
save_checkpoint({'epoch': epoch + 1,'state_dict': restorer.state_dict(),'optimizer' : optimizer.state_dict()},\
args.save_model_path, epoch+1, psnr_t1,ssim_t1,psnr_t2,ssim_t2)
def test(args, restorer, embedder, device, epoch=-1):
combine_type = args.degr_type
psnr_1, psnr_2, ssim_1, ssim_2 = 0, 0, 0, 0
os.makedirs(args.output,exist_ok=True)
for i in range(len(combine_type)-1):
file_list = os.listdir(f'{args.test_input}/{combine_type[i+1]}/')
for j in range(len(file_list)):
hq = Image.open(f'{args.test_input}/{combine_type[0]}/{file_list[j]}')
lq = Image.open(f'{args.test_input}/{combine_type[i+1]}/{file_list[j]}')
restorer.eval()
with torch.no_grad():
lq_re = torch.Tensor((np.array(lq)/255).transpose(2, 0, 1)).unsqueeze(0).to("cuda" if torch.cuda.is_available() else "cpu")
lq_em = transform_resize(lq).unsqueeze(0).to("cuda" if torch.cuda.is_available() else "cpu")
hq = torch.Tensor((np.array(hq)/255).transpose(2, 0, 1)).unsqueeze(0).to("cuda" if torch.cuda.is_available() else "cpu")
starttime = time.time()
text_embedding_1,_,text_1 = embedder([combine_type[i+1]],'text_encoder')
text_embedding_2,_, text_2 = embedder(lq_em,'image_encoder')
out_1 = restorer(lq_re, text_embedding_1)
if text_1 != text_2:
print(text_1, text_2)
out_2 = restorer(lq_re, text_embedding_2)
else:
out_2 = out_1
endtime1 = time.time()
imwrite(torch.cat((lq_re, out_1, out_2, hq), dim=3), args.output \
+ file_list[j][:-4] + '_' + str(epoch) + '_' + combine_type[i+1] + '.png', range=(0, 1))
# due to the vision problem, you can replace above line by
# imwrite(torch.cat((lq_re, out_1, out_2, hq), dim=3), args.output \
# + file_list[j][:-4] + '_' + str(epoch) + '_' + combine_type[i+1] + '.png')
psnr_1 += tensor_metric(hq, out_1, 'PSNR', data_range=1)
ssim_1 += tensor_metric(hq, out_1, 'SSIM', data_range=1)
psnr_2 += tensor_metric(hq, out_2, 'PSNR', data_range=1)
ssim_2 += tensor_metric(hq, out_2, 'SSIM', data_range=1)
print('The ' + file_list[j][:-4] + ' Time:' + str(endtime1 - starttime) + 's.')
return psnr_1 / (len(file_list)*len(combine_type)), ssim_1 / (len(file_list)*len(combine_type)),\
psnr_2 / (len(file_list)*len(combine_type)), ssim_2 / (len(file_list)*len(combine_type))
os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID"
os.environ["CUDA_VISIBLE_DEVICES"] = "0"
if __name__ == '__main__':
parser = argparse.ArgumentParser(description = "OneRestore Training")
# load model
parser.add_argument("--embedder-model-path", type=str, default = "./ckpts/embedder_model.tar", help = 'embedder model path')
parser.add_argument("--restore-model-path", type=str, default = None, help = 'restore model path')
parser.add_argument("--save-model-path", type=str, default = "./ckpts/", help = 'restore model path')
parser.add_argument("--epoch", type=int, default = 300, help = 'epoch number')
parser.add_argument("--bs", type=int, default = 4, help = 'batchsize')
parser.add_argument("--lr", type=float, default = 1e-4, help = 'learning rate')
parser.add_argument("--adjust-lr", type=int, default = 30, help = 'adjust learning rate')
parser.add_argument("--num-works", type=int, default = 4, help = 'number works')
parser.add_argument("--loss-weight", type=tuple, default = (0.6,0.3,0.1), help = 'loss weights')
parser.add_argument("--degr-type", type=list, default = ['clear', 'low', 'haze', 'rain', 'snow',\
'low_haze', 'low_rain', 'low_snow', 'haze_rain', 'haze_snow', 'low_haze_rain', 'low_haze_snow'], help = 'degradation type')
parser.add_argument("--train-input", type=str, default = "./dataset.h5", help = 'train data')
parser.add_argument("--test-input", type=str, default = "./data/CDD-11_test", help = 'test path')
parser.add_argument("--output", type=str, default = "./result/", help = 'output path')
argspar = parser.parse_args()
print_args(argspar)
main(argspar)