I have a numpy image i tried converting to a tensor
tensor_img = torch.from_numpy(numpy_img).permute(2, 1, 0).unsqueeze(0)
I got a rotated and flipped image. What might the problem
So I figured it out in case someone falls into same problem. I had earlier splitted the original image into blocks where each block is stored in a list. This list is then converted into numpy and reshaped accordingly, resulting into the first image above.
import torchvision.transforms as T
transform = T.ToTensor()
tensor_img = transform(numpy_img).unsqueeze(0)
solved the problem
While ToTensor should also work, I think the original issue is caused by a wrong permutation as most likely your numpy image has the shape [H, W, C] which you would like to transform to a tensor in [C, H, W]. To do so you would need to call .permute(2, 0, 1).
2 Likes