Fixing a bug with Touch CUDA version

[garyfeng]

If you run the original notebook on CoLab, you likely will encounter the following bug

RuntimeError: cuda runtime error (11) : invalid argument at /pytorch/aten/src/THC/generic/THCTensorMathPairwise.cu:225

There are likely two sources of errors, both resulting from the fact that Google CoLab is giving me a Tesla T4 GPU instance with CUDA 10.1.

• The default torch, installed via requirements.txt, seems to be for a different cuda version (9?). Need to follow the instruction at https://pytorch.org/get-started/previous-versions/#linux-and-windows-15 to reinstall the current version for cuda 10.x
• The 3 custom models that requires recompiling custom cuda didn't specify the cuda version that support Tesla T4. Need to follow the instruction at https://medium.com/@patrickorcl/compile-with-nvcc-3566fbdfdbf.

Specifically, find the setup.py files under each model/networkd/* folder, and add the following:

  #'-gencode', 'arch=compute_70,code=compute_70',
    '-gencode', 'arch=compute_75,code=sm_75',
    '-gencode', 'arch=compute_75,code=compute_75'

This will add the support for Tesla T4 while also supporting previous versions.

[garyfeng]

!wget https://download.pytorch.org/whl/cu100/torch-1.0.0-cp36-cp36m-linux_x86_64.whl
!pip uninstall torch
!pip install torch-1.0.0-cp36-cp36m-linux_x86_64.whl