`.save_hyperparameters()` from a parent model doesn't save arguments of the child model

[eladsegal]

🐛 Bug

Calling .save_hyperparameters() from a parent model doesn't save arguments of the child model.

To Reproduce

This is a very simple example that has two models, each inherits from a parent.
In the first one, the parent calls .save_hyperparameters() and in the second one the child does.
You can see that difference in .hparams in each of the models.
https://colab.research.google.com/drive/1J4Z8AFanZUIPMffj0ZO0yiKtlDto1OQw?usp=sharing

Expected behavior

It shouldn't matter where .save_hyperparameters() is called in the inheritance chain, all the parameters passed to the instantiated model should be saved in .hparams.
I believe the solution is already given with collect_init_args:

        init_args = {}
        for local_args in collect_init_args(frame, []):
            init_args.update(local_args)

instead of
https://github.com/PyTorchLightning/pytorch-lightning/blob/a3def9d228b4c849e8a6222cffd22e547a17df64/pytorch_lightning/utilities/parsing.py#L220

It worked for me, so let me know if you'd like me to submit a pull request.

[awaelchli]

Hi @eladsegal
Apologies for the late response. Thanks for sending the PR and the test.
Note, calling save_hyperparameters() in a super class constructor is suboptimal. If the signature of the subclass and the super class differ, then the union of these set s of arguments will be collected. Instantiating an object either from the super class or from the subclass will be problematic (this limitation, afaik, applies regardless of your fix).
So I recommend calling save_hyperparameters() only in the most specific subclass.

[eladsegal]

Thanks @awaelchli, I understand and agree.
My use case is pretty standard where I only pass **kwargs to the parent class's __init__, so it should be safe.