Where are model predictions saved by LightningCLI?

[cgoliver]

I am trying to understand how the predict subcommand works with the LightningCLI.

As far as I can tell, it would be calling the lightning module's predict_step method. However, I don't know what happens after this. I see there is a return_predictions flag but not sure where or whether the output is stored somewhere. I am not finding it in the checkpoint directory.

Thanks in advance!

[tshu-w]

Hi, the return of subcommand is not saved in LightningCLI.
https://github.com/Lightning-AI/lightning/blob/0ca3b5aa1b16667cc2d006c3833f4953b5706e72/src/pytorch_lightning/cli.py#L623-L630

You can instance CLI with run=False or Hack the LightningCLI directly to make sure it return results.

cli = LightningCLI(run=False)
cli.trainer.predict(...)

[mauvilsa]

For more context about this topic, read #12438