The Biggest Hurdle I face while I try to transition to PyTorch from Tensorflow is Windows Compatibility, Tensorflow already has Support for Windows with CUDA 9
Here’s a lengthy thread on that matter:
One of the latest comments:
The current master is compilable on Windows. In fact, we have a Windows CI that runs on every PR to make sure that nothing breaks it. I’m not familiar with the process, but maybe @yf225 can share more on the steps to compile on Windows.
Also, there are a few channels which contain Windows64 builds on Anaconda Cloud, like this one:
From comment 304439658
# for CPU only package (also for AMD and Intel Graphics since these GPUs are unsupported) conda install -c peterjc123 pytorch # for Windows 10 and Windows Server 2016, CUDA 8 conda install -c peterjc123 pytorch cuda80 # for Windows 10 and Windows Server 2016, CUDA 9 conda install -c peterjc123 pytorch cuda90 # for Windows 7/8/8.1 and Windows Server 2008/2012, CUDA 8 conda install -c peterjc123 pytorch_legacy cuda80
I used the CI whl as per these instructions:
Works great for with winpython (ie no anaconda, just pip), but I only tried cpu.
Update: Pytorch is supported on windows now. The peterjc123 binaries are obsolete.
Although I was not able to get the conda to install. I had to do the following with pip (as per https://pytorch.org/, windows, pip, 3.6, cuda 9)
pip3 install http://download.pytorch.org/whl/cu90/torch-0.4.1-cp36-cp36m-win_amd64.whl
pip3 install torchvision
This worked on a plain python 3.6 install