Work around issues in load_from_checkpoint() using trainer.predict()

[TimoSommer]

Hey all,

I recently noticed that .load_from_checkpoint() has some unintuitive features, for example not setting the correct model.current_epoch (#12819) and needing users to manually save the hyperparameters.

I also noticed that if I just use the self.trainer.predict api, then self.trainer.model is automatically loaded with the correct weights and current_epoch is also correctly the epoch of the checkpoint I saved, in contrast to when using .load_from_checkpoint(), plus the original, correct hyperparamters:

y_pred = self.trainer.predict(
                                    model=self.trainer.model,
                                    dataloaders=dataloader,
                                    ckpt_path='best'
                                    )

I am not sure though, if this also correctly loads optimizer states and everything else.

I am wondering if this is a valid workaround or if some parts are not loaded this way. I guess there will be some downside to this?

Environment:

OS: Linux
pytorch_lightning version: 1.5.10
torch version: '1.11.0.post2'

[rohitgr7]

I recently noticed that .load_from_checkpoint() has some unintuitive features, for example not setting the correct model.current_epoch (#12819) and needing users to manually save the hyperparameters.

it doesn't load the current_epoch as explained in the issue but it does load the hyperparameters if they are saved using self.save_hyperparameters().

I also noticed that if I just use the self.trainer.predict api, then self.trainer.model is automatically loaded with the correct weights and current_epoch

because model doesn't store the current epoch, it refers the current_epoch from Trainer and since you are using the same trainer instance here, the state persists.