Is there a reason that pytorch-forecasting seemingly unnecessarily saves loss and logging_metrics multiple times?

[dr-upsilon]

• PyTorch-Forecasting version: 1.1.1
• PyTorch version: 2.4.1
• Python version: 3.12.5
• Operating System: windows

C:\...miniconda3\envs\envpt\Lib\site-packages\lightning\pytorch\utilities\parsing.py:208: Attribute 'logging_metrics' is an instance of nn.Moduleand is already saved during checkpointing. It is recommended to ignore them usingself.save_hyperparameters(ignore=['logging_metrics']).

This is caused by self.save_hyperparameters() in init method of TemporalFusionTransformer, because save_hyperparameters() uses inspect and frame to identify all the hyperparameters,

What's the reason to keep it or shall we add handling in init?

[fkiraly]

Maybe a question for @jdb78.

@dr-upsilon, is this happening in a way that is duplicative with lightning core?

[dr-upsilon]

Maybe a question for @jdb78.

@dr-upsilon, is this happening in a way that is duplicative with lightning core?

Yes I think so. BaseModel inherits from HyperparametersMixin from lighting.pytorch.core.mixins.hparams_mixin.

[fkiraly]

Looked through this multiple times and I cannot come up with a good reason (without hearing @jdb78 original rationale).

I suppose it is something that changed between different versions of lightning, so it made sense once but is now duplicative.

Would you like to contribute a PR?