Dockerfile not compatible with Tesla V100

[burnzzz]

The Dockerfile you provide is based on cuda 8.0 and cannot be used with current Tesla Cards because they need at least cuda 9.0. If you need the 32GB V100 to run the training effectively at full resolution, then there is a mismatch.

Will the code be updated to pytorch 1.0 in the future? pytorch 0.4.0 has problems with current cuda releases.

[dustinfreeman]

Hey @burnzzz , I constructed the docker file empirically & independently (I don't work at NVIDIA), and was experimenting with P100s on Google Cloud machines. I've been considering experimenting with V100s (and CUDA 9.0), but haven't yet. If you can construct a V100 dockerfile, let me know!

[dustinfreeman]

It "just worked" for me on V100 when I updated the Dockerfile like so: #67

[rmitsch]

Didn't work for me unfortunately, I had to go with P100s instead - see #73.