List of Callbacks Raises ValueError("Expected a parent")

[amirkhas]

Bug description

If the callbacks list, passing to the trainer, has more than one element, it raises the error ValueError("Expected a parent").
If you pass any of those elements alone the code runs fine.

p.s. I'm already importing like this: import lightning.pytorch as pl which is version >=2. specific.

What version are you seeing the problem on?

2.0+

How to reproduce the bug

No response

Error messages and logs

# Error messages and logs here please

ValueError: Expected a parent

Environment

Current environment

#- Lightning Component (e.g. Trainer, LightningModule, LightningApp, LightningWork, LightningFlow):
#- PyTorch Lightning Version (e.g., 1.5.0):
#- Lightning App Version (e.g., 0.5.2):
#- PyTorch Version (e.g., 2.0):
#- Python version (e.g., 3.9):
#- OS (e.g., Linux):
#- CUDA/cuDNN version:
#- GPU models and configuration:
#- How you installed Lightning(`conda`, `pip`, source):
#- Running environment of LightningApp (e.g. local, cloud):

More info

No response

cc @awaelchli

[carmocca]

Can you share your code?

I've seen this happen when lightning.pytorch and pytorch_lightning imports are mixed together. You need to use only one of them. This includes the callback parent

import lightning.pytorch as lp
import pytorch_lightning as pl

class MyCallback(pl.Callback):
    ...

trainer = lp.Trainer(callbacks=MyCallback())  # bad!

[turian]

@carmocca I had this bug and your suggestion fixed it

[awaelchli]

@carmocca Do you think we could inspect sys.modules and detect whether there are both pytorch_lightning and lightning loaded, and warn the user? Not sure whether that would help.

[turian]

@awaelchli I mean, better warnings are always great.

[mimbres]

#17485 (comment)

Can you share your code?

I've seen this happen when lightning.pytorch and pytorch_lightning imports are mixed together. You need to use only one of them. This includes the callback parent

import lightning.pytorch as lp
import pytorch_lightning as pl

class MyCallback(pl.Callback):
    ...

trainer = lp.Trainer(callbacks=MyCallback())  # bad!

this also solved a similar bug for learning rate monitor callbacks

[carmocca]

@carmocca Do you think we could inspect sys.modules and detect whether there are both pytorch_lightning and lightning loaded, and warn the user? Not sure whether that would help.

I feel like that's too strong of a check. Imagine you are in a team of data scientist who share utility python packages and two packages use lightning, each with a type of import. This check would break their code just on import even if they don't mix the classes. Fixing it might also be out of the control of the person who just tried to import it.

I think it would be simpler that we add a specific check for this in is_overridden and _maybe_unwrap_optimized

[amirkhas]

Can you share your code?

I've seen this happen when lightning.pytorch and pytorch_lightning imports are mixed together. You need to use only one of them. This includes the callback parent

import lightning.pytorch as lp
import pytorch_lightning as pl

class MyCallback(pl.Callback):
    ...

trainer = lp.Trainer(callbacks=MyCallback())  # bad!

That solved my error. Thanks!

[davidgilbertson]

I'm seeing this error with the Optuna PyTorchLightningPruningCallback. They extend pytorch_lightning.callbacks.callback.Callback and I can't control that, but I'm using the new lightning package and don't want to change that.

My rather hacky workaround is this:

import lightning
import optuna

class PatchedCallback(
    optuna.integration.PyTorchLightningPruningCallback, lightning.Callback
):
    pass

I think since there's going to be two lightnings floating around for years, instance checks should be updated to be if isinstance(obj, (old_pl.Callback, new_pl.Callback)), before eventually being reverted after a suitable deprecation period.

Also, will the fix #17548 just mean that I get an error if I use a package that is still on old Lightning? Or is it assuming direct use of the class and it doesn't handle the inheritance case?

[lez-s]

could you please fix the error msg for this? The error message is useless. in my case, it raised TypeError: __init__() got an unexpected keyword argument ‘gpus’ after the ValueError("Expected a parent"), but I didn't even pass gpus args. And it took so long to find this thread.

[nathanpainchaud]

I'm posting this here for visibility, since this seems to be the issue that gets picked up by search engines when looking for ValueError("Expected a parent") errors with Lightning.

The check for mixed imports implemented by @carmocca fails to detect and warn users in at least some configurations of old/new imports. For a minimal example that still produces the ValueError("Expected a parent"), see this issue I've raised with PyTorch Geometric.

To fix the issue on their end I've proposed that they update their way of importing Lightning to use the new style. However, the uninformative error message made it difficult to troubleshoot the root cause of the error. I thought it was relevant to bring this up to the attention of the Lightning devs, to see if you could maybe improve the detection of mixed import styles.