Main motivation for this was to support newer cards - we couldn't run this on our hardware with the older version of Cuda; resultantly, pytorch and etc. have also been updated. Tried to follow some guides online as to the correct interdependent versions of each.
As part to this I've made some minimal changes to the Dockerfile to use more variables and tidy; if we continue using this project, would happily provide some more as I believe it could be improved and made a bit easier and more decoupled.
We also ran into bug #157 again so I've pulled in the fix for this from the comments and seems to work for the minimal example training.
Main motivation for this was to support newer cards - we couldn't run this on our hardware with the older version of Cuda; resultantly, pytorch and etc. have also been updated. Tried to follow some guides online as to the correct interdependent versions of each.
As part to this I've made some minimal changes to the Dockerfile to use more variables and tidy; if we continue using this project, would happily provide some more as I believe it could be improved and made a bit easier and more decoupled.
We also ran into bug #157 again so I've pulled in the fix for this from the comments and seems to work for the minimal example training.