# dstroot for saving models.
# logroot for writting some log(s), if is needed.
dstroot = './trainedmodels/' + data_name + '/' + tag + '/'
logroot = './logs/' + data_name + '/' + tag + '/'
subprocess.check_output(['mkdir', '-p', dstroot])
subprocess.check_output(['mkdir', '-p', logroot])
# Transform
transform = transforms.Compose(
[transforms.RandomCrop((crop_size, crop_size)),
transforms.ToTensor()])
# Dataloader
convertor = data_convertors.ConvertImageSet(data_root,
imlist_pth,
data_name,
transform=transform,
is_train=True)
dataloader = DataLoader(convertor,
batch_size=bch_size,
shuffle=False,
num_workers=2)
# Make network
cleaner = cleaner().cuda()
cleaner.train()
# Optimizer and Loss
optimizer = optim.Adam(cleaner.parameters(), lr=base_lr)
L2_loss = nn.MSELoss()
if data_name == 'RESIDE':
testroot = "../data/"+data_name+"/sots_indoor_test/"
test_list_pth = "../lists/RESIDE_indoor/sots_test_list.txt"
elif data_name == 'DCPDNData':
testroot = "../data/"+data_name+"/TestA/"
test_list_pth = '../lists/'+data_name+'/testA_list.txt'
else:
print('Unknown dataset name.')
Pretrained = '../trainedmodels/'+data_name+'/'+tag+'_model.pt'
show_dst = '../cleaned_images/'+data_name+'/'+tag+'/'
#subprocess.check_output(['mkdir', '-p', show_dst])
# Set transformer, convertor, and data_loader
transform = transforms.ToTensor()
convertor = data_convertors.ConvertImageSet(testroot, test_list_pth, data_name,
transform=transform)
dataloader = DataLoader(convertor, batch_size=1, shuffle=False, num_workers=1)
# Make the network
cleaner = cleaner().cuda()
cleaner.load_state_dict(torch.load(Pretrained))
cleaner.eval()
I_HAZE = "D:\Image_dataset\# I-HAZY NTIRE 2018\hazy1\\"
I_HAZE_GT = "D:\Image_dataset\# I-HAZY NTIRE 2018\GT1\\"
O_HAZE="D:\Image_dataset\# O-HAZY NTIRE 2018\hazy1\\"
O_HAZE_GT="D:\Image_dataset\# O-HAZY NTIRE 2018\GT1\\"
SOTSI = "C:\\Users\FQL\Desktop\RESIDE-standard\SOTS\indoor\hazy\\"
SOTSI_GT = "C:\\Users\FQL\Desktop\RESIDE-standard\SOTS\indoor\gt1\\"
SOTSO = "F:\SOTS\hazy\\"
SOTSO_GT = "F:\SOTS\clear\\"
# Set pathes
data_root = '../data/' +data_name+'/train/train/'
imlist_pth = '../lists/'+data_name+'/train_list.txt'
# dstroot for saving models.
# logroot for writting some log(s), if is needed.
dstroot = './trainedmodels/'+data_name+'/'+tag+'/'
logroot = './logs/'+data_name+'/'+tag+'/'
subprocess.check_output(['mkdir', '-p', dstroot])
subprocess.check_output(['mkdir', '-p', logroot])
# Transform
transform = transforms.ToTensor()
# Dataloader
convertor = data_convertors.ConvertImageSet(data_root, imlist_pth, data_name,
transform=transform, is_train=True,
with_aug=with_data_aug, crop_size=crop_size)
dataloader = DataLoader(convertor, batch_size=bch_size, shuffle=False, num_workers=4)
# Make network
cleaner = cleaner().cuda()
cleaner.train()
# Optimizer and Loss
optimizer = optim.Adam(cleaner.parameters(), lr=base_lr)
L1_loss = nn.L1Loss()
# Start training
print('Start training...')
for epoch in range(epoch_size):
for iteration, data in enumerate(dataloader):