Upgrade to PyTorch 1.0

[meetps]

Planned updates

• Upgrade to pytorch-1.0
• Allow support for models to have custom loss functions.
• Pretrained models on S3 for common models-dataset pairs
• Compatibility matrix for dataset with models
• RefineNet and E-Net implementations
• MS-COCO, Mapillary Vistas and BDD-100k datasets
• Interface to combine datasets with a user-given mapping dictionary
• Improved metric logging

[kcyu2014]

Hi,
I have recently tested the current build on pytorch 0.4.1, and it encounter CUDA out of memory error.
There is no problem in 0.4.0 for the current time being.
Hope you could also look into this direction during the upgrade :)

kc

[Spritea]

Hi, guys:

I found a simple way to use this code with Pytorch 1.0 or 0.4.1. This error is caused by functools package used in fcn.py. So just comment related code below and it works.

#self.loss = functools.partial(cross_entropy2d, size_average=False)

Besides, for other models without functools, like segnet, it works correctly in Pytorch 1.0/0.4.1, without any modification.

See related #173