Restore the best model

[mpariente]

What would be the most lightning way to restore the best model? Either directly after training (in the same script) or for later use (in another script)?

Thanks in advance !

[neggert]

You can use the checkpont callback to save the only the best model as described here: https://williamfalcon.github.io/pytorch-lightning/Trainer/Checkpointing/ (note that that doc needs to be updated. Use save_top_k=1 instead of save_best_only)

You can then use the load_from_checkpoint method to restore your checkpoint: https://williamfalcon.github.io/pytorch-lightning/LightningModule/methods/#load_from_metrics

[Borda]

in the last version, there is also Trainer initialization parameter resume_from_checkpoint

[mpariente]

You can use the checkpont callback to save the only the best model as described here: https://williamfalcon.github.io/pytorch-lightning/Trainer/Checkpointing/ (note that that doc needs to be updated. Use save_top_k=1 instead of save_best_only)

I didn't pay attention, I was using save_best_only, thanks !

Thank you both for you answers, but what if I want to keep my checkpoints, and still restore the best one automatically, how would I do?

[neggert]

Right now, I don't think the score is saved with the checkpoint. You'd need to take a look at your logs and manually select the best one.