Error when training with torch 1.5.1

[mkmohangb]

There are some additional checks added in the pytorch framework which is causing the error when running the training script with torch 1.5.1. With torch version 1.2.0 training proceeds without error.

pytorch/pytorch@3e8d813

python train.py --gpu 0

Traceback (most recent call last):
File "train.py", line 231, in
main()
File "train.py", line 178, in main
preds = model(images)
File "/home/xxx/miniconda3/lib/python3.7/site-packages/torch/nn/modules/module.py", line 550, in call
result = self.forward(*input, **kwargs)
File "/home/xxx/miniconda3/lib/python3.7/site-packages/torch/nn/parallel/data_parallel.py", line 153, in forward
return self.module(*inputs[0], **kwargs[0])
File "/home/xxx/miniconda3/lib/python3.7/site-packages/torch/nn/modules/module.py", line 550, in call
result = self.forward(*input, **kwargs)
File "/home/xxx/Self-Correction-Human-Parsing/networks/AugmentCE2P.py", line 298, in forward
x = self.relu1(self.bn1(self.conv1(x)))
File "/home/xxx/miniconda3/lib/python3.7/site-packages/torch/nn/modules/module.py", line 550, in call
result = self.forward(*input, **kwargs)
File "/home/xxx/Self-Correction-Human-Parsing/modules/bn.py", line 119, in forward
self.training, self.momentum, self.eps, self.activation, self.slope)
RuntimeError: Some elements marked as dirty during the forward method were not returned as output. The inputs that are modified inplace must all be outputs of the Function.

[GoGoDuck912]

I suppose this error is accompanied by different PyTorch versions and caused by the implementation of InplaceABN.

In order to upgrade to pytorch>1.0, we adopt [InplaceABN v0.1](the https://github.com/mapillary/inplace_abn/tree/v0.1), which empirically works well on most PyTorch versions.

[mkmohangb]

Created Pull request: #22, please review.