What is hparams exactly?

[keunwoochoi]

Hi, thanks for the nice product again.

From #525 and #599, I could guess that hparams is required to load a saved model (which I think should be mentioned somewhere in the doc btw). And from the examples, seems like hparams may be argparse.Namespace. Unfortunately though, it was not so easy to understand the concept.

What is hparams exactly? What kind of information it should/can/should not include to work properly? Is it recommended to use hyperparameter argument parser? Say, if I'm not into hyperparameter search at the moment and just want to be able to load the checkpoint model, what is the requirement on the hparams?

[s-rog]

hparams is just a dict that contains your hyperparameters for easy logging and experiments. It is optional, but does make loading models more tedious.

[neggert]

It should be an argparse.Namespace. You can get this from argparse, testtube's HyperOptArgumentParser, or create it manually from a dict like so: argparse.Namespace(**my_dict).

[keunwoochoi]

@s-rog Thanks, but it's not a dict :-)
@neggert Thanks. Yes those are the info I could assume by reading code etc. But I think those and probably more details should be documented somewhere. I'd like to contribute, but I can't document something I don't know clearly.
Also, in my opinion, it should be very strongly recommended to use it so that they can load the model later, assuming the use cases of pytorch-lightning would be more serious than just training something on MNIST. Because, for those toy example, why bother :)

[s-rog]

oh yeah I keep forgetting, you're right!

[neggert]

Yes, there's a complete overhaul of the docs in progress. We'll make sure this gets in there.

[williamFalcon]

thanks for the input!
Once we have the new docs, please let us know of any problems! :)