Argument Resolution Refactoring

[justusschock]

We should definately check where to give what argument (and maybe expose more keyword arguments) in experiment (Possible ways are to pass arguments via keyword or parameter object in __init__.py or run/test/resume/kfold, so there are 4 different ways to pass arguments)

We should probably do this, after merging #166

[justusschock]

Any opinions, on what to pass which way @delira-dev/delira ?

I would prefer to expose as many options as possible and giving them good defaults, but things like loss functions, metrics, model parameters should be passed by the parameter object IMO.

[mibaumgartner]

Maybe we should create a google docs/sheets document where we collect all parameters and decide where the respective argument should be passed?

[justusschock]

You got the job :D

[mibaumgartner]

Just created the table :)
https://docs.google.com/spreadsheets/d/1c0NKu0P_STGpUyjDsxQhxH7mUojgXkTQUWXq3PFzTJc/edit?usp=sharing

[justusschock]

So to summarize this:

Init:

• model_cls
• name
• save_path
• logging_type
• logging_kwargs
• convert_to_npy
• key_mapping
• verbose

train/resume/kfold:

• params
• val_score_key
• optim_builder
• checkpoint_freq
• trainer_cls
• gpu_ids
• reduce_mode
• datamgr

test

• predictor_cls
• model
• prepare_batch
• kwargs
• datamgr
• callbacks

params

• n_epochs
• metrics
• metric_keys
• losses
• optim_cls
• optim_params