(windows10 / python3.6 / conda:latest / all-other-packages:latest)

There is a memory leak somewhere in the multiprocessing section:

Running 8 workers on a system with 16GB of ram will lead to a crash within ~1-2mins and a nasty pagefile increase to 40GB! I re-wrote the script using the older Process() api and no shared variables between instances which fixed the issue so its probably something to do with data passing.

All of my tests were done on the full celebahq dataset.