Model checkpoints/files required for predictions


I’m training my NMT systems with google colab/drive, so I have limited storage. As training multiple models can take up a lot of storage, which of the ‘outputs’ from the model are the most important to keep if I want to be able to make predictions with the model but not waste lots of storage space? I.e. is it enough to just keep the most recent checkpoint or is there anything else important which gets outputted and is worth keeping?




You only need the latest checkpoint. Both OpenNMT-py and OpenNMT-tf have an option to only keep the latest checkpoint on disk.

For OpenNMT-py:

keep_checkpoint: 1

For OpenNMT-tf:

  keep_checkpoint_max: 1

ok, thanks. I may keep a few more checkpoints on the disk just so I can average them at the end, but I’ll delete the earlier checkpoints after I’ve averaged them.