[Question] How could I hide the "Finding best initial lr" message from pytorch_lightning when using Darts' Torch Forecasting Models?

[fmerinocasallo]

I am unable to hide the Finding best initial lr message when calling the lr_find method associated with Darts' Torch Forecasting Models, such as BlockRNNModel:

Based on my understanding, this message is generated by pytorch-lightning. In particular, by the on_train_batch_start method from the _LRCallback class. At least in this specific case 🤔

I have tried the following:

1. Including verbose=False when calling the lr_find method.
2. Passing a TQDMProgressBar(refresh_rate=0) instance through the callbacks list in the pl_trainer_kwargs dict passed to the BlockRNNModel constructor.
3. Including an "enable_progress_bar": False in the pl_trainer_kwargs dict passed to the BlockRNNModel constructor.

So far, no luck 😞

I don't know if I have misunderstood something or I am missing some critical bit of information 😬

Could you help me solve this issue? 🙏

Any feedback would be much appreciated ☺️

---

I am using:

• python 3.11.0
• darts 0.28.0
• pytorch_lightning 1.9.5

Let me know if there are any missing but relevant and clarifying details I should mention.

[dennisbader]

Hi @fmerinocasallo. Unfortunately, I think there is currently no built-in/easy way to disable the lr_find progress bar.
As you say, the progress bar comes from pytorch_lightning, spefically from pytorch_lightning.tuner.tuning.Tuner.lr_find(). It seems like lr_find() currently doesn't offer a way to disable the progress bar.

If this was supported by lightning, we can add support for it to Darts as well.