📣 Weights & Biases - Feedback

Hey folks!

I’m Morgan from Weights & Biases and we’re looking for feedback from the :hugs: community on our current documentation for our Hugging Face W&B integration. Any and all feedback is a real gift :gift_heart: (especially if there are things that are unclear or haven’t worked for you in the past)

Docs to Check Out :eyes:

Suggested Questions

  • What would you change or add for a Hugging Face user using W&B for the first-time?
  • What would you change or add in the docs for a Hugging Face user already using W&B?
  • Is there anything in the docs you feel is missing or could be better explained?
  • Are there any example descriptions or example code that would better explain the W&B integration?

Since I’m here, happy to hear about any suggestions or issues with the Hugging Face W&B integration here or in @boris 's thread here :slightly_smiling_face:

Myself and @boris are looking forward to hearing from you all!


Hi @morgan,

Is there any documentation on how to integrate HuggingFace Accelerate (multi-gpu on a single machine setting) with weights and biases sweep?

hey @berkin , as long as your script can accept changes to the hyperparameters being set by wandb then it should work as normal. Using Accelerate you could just call the wandb tracker and get your configs from there, e.g.

wandb_tracker = accelerator.get_tracker("wandb")
learning_rate = wandb_tracker.config["learning_rate"]

You could also call wandb directly, assuming wandb is running on that GPU

learning_rate = wandb.config["learning_rate"]