from unet import Unet
n = Unet.load('models/unet3-weights-improvement-04-0.68.hdf5', num_classes=7)
# n.test('/junction/dental_unet/files/train/1.bmp', 'res.bmp')
n.test('/junction/dental_unet/files/train/2.bmp', 'res.bmp')
model.save(args.checkpoint, cur_iter, loss,
optimizer.param_groups[0]['lr'])
if best_loss > loss:
best_loss = loss
last_checkpoint = model.last_checkpoint(
args.checkpoint, '.pth')[0]
shutil.copy(
last_checkpoint,
os.path.join(args.checkpoint, f'best_model.pth'))
if args.iter_lr and cur_iter % args.iter_lr == 0:
# decay lr
scheduler.step()
model = Unet(3, 64)
state_dict = model.load(args.checkpoint)
def inference(image, mask):
global model, result_dir
model.eval()
# 0 for hole 1 for valid
mask = torch.where(mask > 0, torch.tensor(0), torch.tensor(1))
mask = torch.cat([mask, mask, mask]).float()
out_img, _ = model(image.unsqueeze(0), mask.unsqueeze(0))
result = out_img.squeeze(0) * (1 - mask) + mask * image
# save result
time = datetime.today().strftime('%H_%M')
savepoint = os.path.join(result_dir, f'{time}_{image.filename}')
result_img = torchvision.transforms.ToPILImage()(result)
