def tensor2im(image_tensor, imtype=np.uint8, normalize=True): image_numpy = image_tensor.cpu().float().detach().numpy() if normalize: image_numpy = (image_numpy+1)*255.0*0.5 else: image_numpy = (image_numpy+1)*255.0 image_numpy = np.clip(image_numpy, 0, 255) blank_image = np.zeros((image_tensor.shape[1],image_tensor.shape[2],image_tensor.shape[0]), np.uint8) if image_tensor.shape[0] == 3: blank_image[:,:,0]=image_numpy[2,:,:] blank_image[:,:,1]=image_numpy[1,:,:] blank_image[:,:,2]=image_numpy[0,:,:] else: blank_image[:,:,:]=image_numpy[:,:,:] return blank_image def im2tensor(image_numpy, normalize=True): if normalize: image_numpy = (image_numpy/255.0)*2.0-1.0 else: image_numpy = image_numpy/255.0 image_numpy = np.clip(image_numpy, -1, 1) blank_image = np.zeros((image_numpy.shape[2],image_numpy.shape[0],image_numpy.shape[1])) if image_numpy.shape[2] == 3: blank_image[2,:,:]=image_numpy[:,:,0] blank_image[1,:,:]=image_numpy[:,:,1] blank_image[0,:,:]=image_numpy[:,:,2] else: blank_image[:,:,:]=image_numpy[:,:,:] image_tensor = torch.Tensor(blank_image) return image_tensor w_size = 1024 h_size = 256 input_label = torch.zeros([input_labe.shape[0],input_labe.shape[1],h_size,w_size], dtype=torch.float32,device=input_labe.device) for i in range(input_labe.shape[0]): f_label = input_labe [i,:,:,:] f_label_img = tensor2im(f_label) f_label_img = cv2.resize(f_label_img,(w_size,h_size)) input_label[i,:,:,:] = im2tensor(f_label_img, normalize=True)