dev = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')
nepochs = 50
log_step = 100
# Define network
net = ESPCN(8)
net.to(dev)
print(net)
# Define loss
criterion = nn.MSELoss()
# Define optim
optimizer = optim.Adam(net.parameters(), lr=0.0001)
scheduler = optim.lr_scheduler.ReduceLROnPlateau(optimizer, patience=5, verbose=True)
pretrained = torch.load('models/unet_bn_20190912_040318.pth', map_location='cuda:0')
net.load_state_dict(pretrained['model_state_dict'])
optimizer.load_state_dict(pretrained['optimizer_state_dict'])
net.eval()
with torch.no_grad():
val_loss = 0.0
start = time.time()
for i, (inps, lbls) in enumerate(dataloaders['val']):
inps = inps.to(dev)
lbls = lbls.to(dev)
outs = net(inps)
train_set = get_training_set(args.upscale_factor)
test_set = get_test_set(args.upscale_factor)
train_data_loader = DataLoader(dataset=train_set, num_workers=args.threads, \
batch_size=args.train_batch_size, shuffle=True)
test_data_loader = DataLoader(dataset=test_set, num_workers=args.threads, \
batch_size=args.test_batch_size, shuffle=False)
net = ESPCN(upscale_factor=args.upscale_factor).to(device)
# Uncomment below to load trained weights
# weights = torch.load('data/weights/weights_epoch_30.pth')
# net.load_state_dict(weights)
criterion = nn.MSELoss()
optim = torch.optim.Adam(net.parameters(), lr=args.lr)
def train(epoch):
epoch_loss = 0
for iteration, batch in enumerate(train_data_loader, 1):
img_in, target = batch[0].to(device).type(dtype), \
batch[1].to(device).type(dtype)
optim.zero_grad()
loss = criterion(net(img_in), target)
epoch_loss += loss.item()
loss.backward()
optim.step()
num_workers=args.num_workers)
val_dataset = Dataset(path=args.path,
mode='val',
upscale_factor=args.upscale_factor)
val_dataloader = DataLoader(val_dataset,
batch_size=args.batch_size,
shuffle=False,
num_workers=args.num_workers)
print("==>Loading model")
model = ESPCN(upscale_factor=args.upscale_factor).to(device)
#data parallel
if len(args.gpu_ids) > 1:
model = nn.DataParallel(model, args.gpu_ids)
print(model)
optimizer = optim.Adam(model.parameters(), lr=args.lr)
criterion = nn.MSELoss().to(device)
psnr_best = 0
for epoch in range(args.epochs):
train(train_dataloader, model, epoch, criterion, optimizer, args)
psnr = evaluate(val_dataloader, model, epoch, criterion, args)
if psnr > psnr_best:
psnr_best = psnr
save_path_name = os.path.join(
args.save_path,
'x{}_best_model.pth'.format(args.upscale_factor))
print('model is saved at ', save_path_name)
torch.save(model, save_path_name)