Aggregator return error when the tensors in batch are integer tensors.

🐛Bug Aggregator return error when the tensors in batch are integer tensors when invoking get_output_tensor() in ‘average’ mode.

To reproduce

import torch
import torchio as tio
from torch.utils.data import DataLoader
# create target image
int_im = tio.ScalarImage(tensor=(torch.rand(1, 256, 256, 256) * 65535).type(torch.int32))

# create sampler
subject = tio.Subject(im=int_im)
grid_sampler = tio.GridSampler(subject, patch_overlap=20, patch_size=128)
inference_dl = DataLoader(grid_sampler, batch_size=4)
aggregator = tio.GridAggregator(grid_sampler, 'average')

# sampler and rebuild tensor
for i, mb in enumerate(inference_dl):
    aggregator.add_batch(mb['im'][tio.DATA], mb[tio.LOCATION])
o = aggregator.get_output_tensor()

Simple fix aggregator.py https://github.com/alabamagan/torchio/commit/084caaca5c9b41bcb99aa734e46a2eee43c5a9da

Expected behavior

Reconstruct the output integer data array correctly.

Actual behavior

Produced an error.

---------------------------------------------------------------------------
RuntimeError                              Traceback (most recent call last)
<ipython-input-6-90a464355da2> in <module>()
      8 for i, mb in enumerate(inference_dl):
      9     aggregator.add_batch(mb['image'][tio.DATA], mb[tio.LOCATION])
---> 10 o = aggregator.get_output_tensor()

~/Toolkits/Anaconda2/envs/py3_cuda10/lib/python3.7/site-packages/torchio/data/inference/aggregator.py in get_output_tensor(self)
    163             self._output_tensor = self._output_tensor.type(torch.int32)
    164         if self.overlap_mode == 'average':
--> 165             output = self._output_tensor / self._avgmask_tensor
    166         else:
    167             output = self._output_tensor

RuntimeError: Integer division of tensors using div or / is no longer supported, and in a future release div will perform true division as in Python 3. Use true_divide or floor_divide (// in Python) instead.

System info

Platform: Linux-4.15.0-140-generic-x86_64-with-debian-stretch-sid
TorchIO:  0.18.37
PyTorch:  1.6.0
NumPy:    1.19.1
Python:   3.7.3 (default, Mar 27 2019, 22:11:17)
[GCC 7.3.0]
**```z**

[alabamagan]

I didn’t notice at first because I was working on integer tensors, but I think changes in be4bf4d to using true_division might lead to loss in precision if the averaged images patches are meant to be float values. Might have big impact if the averaged patches are normalized to be ranged from 0-1. Might worth revisiting.

[alabamagan]

In short, I am working on index arrays and deformation tensor, so they took form of integer tensors during slicing, but I can perform manual casting as long as the pipeline is not broken. I think you can close this issue in light that newer versions of pytorch don’t have this problem.