Micael Carvalho
Micael Carvalho
Work in progress
Cluster management tools might send SIGKILL or other signals to our job. In this case, we want to checkpoint before killing the job.
Current code crashes if `last` is provided for the first job. We want this behavior to be standard and not resume if the log dir does not exist.
This PR has a dependency on https://github.com/Cadene/bootstrap.pytorch/pull/2
With these changes I was able to run build_graph for the replica dataset. They bypass some assumptions about floors and stairs