Refacto py4cast start

[clemoule]

implementation of :

• cli.py for executing the lightning cli

• config_cli.yaml for parsing cli arguments

• a lightningmodule.py which propose a standard structure for lightningmodule

• a datamodule.py which propose a standard structure for datamodule

• a utils.py in which are created utils class that are specialized in one task each. The idea is to create function and override methods methodically during the execution of lightningmodules. (ex: class dummylightningmodule(utils1, utils2, pl.LightningModule) (--> see method resolution order for more details)

• refacto of the old version of lightning.py for autoregressive lightning module/config/cli

• same for a masked auto encoder version (without autoregressive steps)

A faire:

(certain)

• DISCUTER DE L'ARBORESCENCE
• enlever dataset_info
• enlever torchdataloadersettings -> batch_size etc...
• rajouter une fonction de save/load checkpoints/weights
• harmoniser les méthodes (named tensor/torch tensor)
• épurer le code
• commenter le code
• on_save/load_checkpoint comme dans lightning.py ?

(peut etre)

• refaire la meme chose pour les num_inter_step
• centraliser les méthodes si possible dans utils.py

pas encore de autoencoderDataModule donc pour l'instant comportement encoder-decoder (x!=y) et pas autoencoder(x=y).

[cbovalo]

@clemoule Is this PR ready for review or is it a draft ? BTW, after a quick review, I find the code would be more readable.

[clemoule]

@clemoule Is this PR ready for review or is it a draft ? BTW, after a quick review, I find the code would be more readable

The code can be executed but i need to verify the plot/log method. We will discuss of the way to redesign the trainers, parsers, lightningmodule and datamodule this afternoon. So this is more of a draft even though it works. We will be set this afternoon.

[clemoule]

storing functions in classes (ex: PlotLightningModule) is a bit of over engineering but it helps to group functions and methods by purpose. It makes the code easier to map and understand => quicker to pick up the code => quicker to implement additions to the code and quicker to tinker

[cbovalo]

I think it's also a good thing to decouple computing metrics and making plots. I'll make some comments on the code if improvement is needed.