How to checkpoint Datamodule's hyperparams/parameters on Model Checkpoint?

[FeryET]

I want to create an inference model, and I have multiple datasets. For this issue, I want to save some settings about my DataModule and indicate what is the test indices, what is the dataset location and etc.

How can I add these to model_checkpoint?

[rohitgr7]

there is a on_save_checkpoint hook which receives the checkpoint (dict) before its saved. You can attach whatever you need inside your datamodule.

[FeryET]

@rohitgr7

How to load the checkpoint info later into the datamodule? So far I use LightningModule.load_from_checkpoint to load a model, but I don't know how to parse the whole dict here?

[rohitgr7]

parse the whole dict here?

can you elaborate more on this? checkpoint argument is a dict here. so you can easily access whatever you save inside on_save_checkpoint by checkpoint[some_key].

[FeryET]

parse the whole dict here?

can you elaborate more on this? checkpoint argument is a dict here. so you can easily access whatever you save inside on_save_checkpoint by checkpoint[some_key].

I have a LightningModule instance here, let's call it PLModule and a LightningDataModule let's call it dm. When I save a checkpoint via a trainer, on_save_checkpoint callbacks get called and something is saved in the modules. The problem is that when I want to load the checkpoint, PLModule.load_checkpoint clearly does not load the LightningDataModule information I want. I want to know if there is anyway to load the DataModule seperately without including it in the PLModule as a field?

[rohitgr7]

you mean PLModule.load_from_checkpoint here? if that's so then yes it won't load datamodule or callback related stuff since it just works when you resume your training. I think we should load datamodule states as well. Mind opening an issue? Let's see what others think :)
https://github.com/PyTorchLightning/pytorch-lightning/blob/dbe1662dc38d5217328ab459743b1113869a628c/pytorch_lightning/trainer/trainer.py#L1017-L1018