PPO config __init__ is bloated

[GauravVirmani]

What does this PR do?

Fixes: #233

Moving total_ppo_epochs, forward_batch_size and log_with to post init method and let the dataclass automatically assign the other member variables.

cc @edbeeching @lvwerra

[HuggingFaceDocBuilderDev]

The documentation is not available anymore as the PR was closed or merged.

[younesbelkada]

Awesome work, that looks great to me. Thanks a mile for fixing

[edbeeching]

This is great, that your for addressing this.
Would it be worthwhile removing the long args description and using fields + metadata?
For example: log_with: Optional[str] = field(default=None, metadata={"help": "use 'wandb' to log with wandb"})

[younesbelkada]

Hi @GauravVirmani ,
Can you run the styling checks?

make style && make quality

[younesbelkada]

Hi @GauravVirmani
I think there is a mismatch between our black versions, mine is:

>>> import black
>>> black.__version__
'23.1.0'

Can you please double check?

[GauravVirmani]

Hey @younesbelkada

My version is same:

Is there a way to run these checks in my local machine? Let me try running the styling checks again.

[younesbelkada]

Hmm in my experience you just need to run make style && make quality. Otherwise I can do it for you

[younesbelkada]

Thanks a lot for all your work on this!

[GauravVirmani]

Hi @younesbelkada

I have added changes to incorporate the 2 new arguments(early_stopping and target_kl) added in PR #238 to resolve merge conflict with main branch.

@edbeeching @lvwerra It would be great if you can review and get this merged.

[lvwerra]

Thanks @GauravVirmani for fixing the merged PR!