Pytorch-lightning: Expose lr_scheduler to pl.LightningModule as a property (self.scheduler)?

[seannz]

What is your question?

Dear all,

Is lr_scheduler (or an array of them) exposed to pl.LightningModule as self.scheduler or something similar ? I only see self.trainer but self.scheduler would really complement the trainer!

What have you tried?

I tried looking for self.scheduler (and different variants) in the source / PDF manual but found nothing

What's your environment?

• OS: CentOS 6
• Packaging: conda
• Version: 1.0.4

[github-actions[bot]]

Hi! thanks for your contribution!, great first issue!

[rohitgr7]

it's not there yet. Mind submit a PR to add that?
Something similar for reference:
https://github.com/PyTorchLightning/pytorch-lightning/blob/7b375ed1d3ce5f55c1021038d2892dce6ae8bd66/pytorch_lightning/core/lightning.py#L117-L126

[seannz]

Looks like we should be accessing it via self.trainer.schedulers (rather than directly via self.schedulers) for consistency with self.trainer.optimizers. I'll submit a PR soon.

[MohammedAljahdali]

Hey, I think you can access the scheduler by using trainer.lr_schedulers, therefore maybe you could access from your pl_module by using self.trainer.lr_schedulers

[seannz]

@MohammedAljahdali I don't see lr_schedulers anywhere in trainer.py. Would you mind pointing me to the right location where it is defined?

[rohitgr7]

@seannz it's added here:
https://github.com/PyTorchLightning/pytorch-lightning/blob/a32bffcdea3c12fb369f2cb87cbcf8b7e91396e6/pytorch_lightning/accelerators/accelerator.py#L187-L194

Hey, I think you can access the scheduler by using trainer.lr_schedulers, therefore maybe you could access from your pl_module by using self.trainer.lr_schedulers

yes you can access it via self.trainer but I think it can be exposed to LightningModule as property just like optimizers