The in-place multiplication may cause error in the attribution by excitation-backprop

[shinkyo0513]

Hi, I found an error when I use the excitation-backprop method to attribute my Resnet.

The program raises an error when backprop by the excitation-backprop rule when it comes to the pooling layer.
The error is 'diff_view_meta->creation_meta == CreationMeta::DEFAULT INTERNAL ASSERT FAILED at /opt/conda/conda-bld/pytorch_1591914838379/work/torch/csrc/autograd/variable.cpp:82, please report a bug to PyTorch.'

I found this is because the _patch_pool's backward function contains the in-place multiplication, which has been suggested by PyTorch not to use with autograd.

TorchRay/torchray/attribution/excitation_backprop.py

Line 401 in 6a198ee

grad_input_ *= input_

I changed the in-place operation to out-of-place, and the error disappeared.