Use the Callback state_key to disambiguate callbacks provided to trainer and from LightningModule

[ananthsub]

We still use the types of the callbacks here: https://github.com/PyTorchLightning/pytorch-lightning/blob/92d9fc2280091c300fe8848284b9f6dfcbd209eb/pytorch_lightning/trainer/connectors/callback_connector.py#L272-L283

This can lead to collisions, when perhaps we have multiple callback instances of the same type with different settings that we want to use together.

For checkpointing callback states, we have shifted to using the newly introduced callback state key here: https://github.com/PyTorchLightning/pytorch-lightning/blob/92d9fc2280091c300fe8848284b9f6dfcbd209eb/pytorch_lightning/trainer/callback_hook.py#L58

We can do the same thing in this callback connector

Originally posted by @ananthsub in #11060 (comment)

cc @justusschock @awaelchli @akihironitta @rohitgr7 @ananthsub @daniellepintz

[awaelchli]

Unfortunately this solution is incomplete. Since the removal of on_init_end, callbacks like ModelCheckpoint are not fully initialized until the setup() hook runs, but the configure_callbacks hook gets called earlier than that. Using the state_key there that includes an uninitialized attribute (like _save_on_train_epoch_end) will lead to a wrong comparison in some edge cases.

[stale]

This issue has been automatically marked as stale because it hasn't had any recent activity. This issue will be closed in 7 days if no further activity occurs. Thank you for your contributions, Pytorch Lightning Team!

[stale]

This issue has been automatically marked as stale because it hasn't had any recent activity. This issue will be closed in 7 days if no further activity occurs. Thank you for your contributions, Pytorch Lightning Team!