Slicing a MetaTensor: TypeError: only integer tensors of a single element can be converted to an index #5844
Description
The goal is to slice the tensor in half. For some reason that I cannot comprehend, it doesn't work on a particular tensor:
>>> type(x)
<class 'monai.data.meta_tensor.MetaTensor'>
>>> x.shape
torch.Size([128, 1024, 8, 64])
>>> x[slice(0, 64, None), ...]
*** TypeError: only integer tensors of a single element can be converted to an index
# OK with [0:16]:
>>> x[slice(0, 16, None), ...].shape
torch.Size([16, 1024, 8, 64])
# NOT OK with [0:32]:
>>> x[slice(0, 32, None), ...]
*** TypeError: only integer tensors of a single element can be converted to an index
>>> x[0:16].shape
torch.Size([16, 1024, 8, 64])
# ARGGGGGHHHH
>>> x[0:32].shape
*** TypeError: only integer tensors of a single element can be converted to an index
>>> x[0:17].shape
*** TypeError: only integer tensors of a single element can be converted to an index
# What if we convert to Numpy? (result: it works fine)
>>> x.detach().numpy()[0:64].shape
(64, 1024, 8, 64)What could be going on here?
If it is of any use, the error occurs at this specific line: https://github.com/pashtari/factorizer/blob/fa7f03dfe8c390f0993cbc0d161922d2de819a1d/factorizer/factorization/operations.py#L331
To Reproduce
# This works fine, I cannot reproduce in this way
>>> from monai.data import MetaTensor
>>> x = MetaTensor(np.random.rand(128, 1024, 8, 64))
>>> x[slice(0, 64, None), ...].shape
torch.Size([128, 1024, 8, 64])How is it possible that a specific tensor would cause such error?
Environment
Ubuntu 20.04
python 3.10
monai 1.1.0 pyhd8ed1ab_0 conda-forge
pytorch 1.13.1 py3.10_cuda11.7_cudnn8.5.0_0 pytorch
pytorch-cuda 11.7 h67b0de4_1 pytorch
pytorch-lightning 1.8.1 pyhd8ed1ab_0 conda-forge
pytorch-mutex 1.0 cuda pytorch
torchmetrics 0.11.0 pyhd8ed1ab_0 conda-forge
torchvision 0.14.1 py310_cu117 pytorch