Nixtla's Open Source Time Series Ecosystem.

Nixtla Community

image.png

Hello <@U05EK0TTXM1>! You can find the tune.Tuner method in the neuralforecast/common/_base_auto.py file, in line 180.

There is no way currently to change the RunConfig when instantiating the auto model. You will need to clone the repo and change it directly. Can you let us know if it works? If so, we will add a parameter to change the `storage_path` from the auto model.

I think this PR is trying to address that: <https://github.com/Nixtla/neuralforecast/pull/735>

<@U03619XRWKD> thanks for this tips, I'll check this out

<@U03UX9KR75W> <@U03619XRWKD> can confirm that the addition of line doesn't help. This path sync issue was also discussed here <https://github.com/ray-project/ray/issues/36448>

while Ray fixes the issue I recommend you to try the NHITS directly. We know that Tune has a lot of issues, we are considering implementing alternatives when Tune is not an option.

<@U03619XRWKD> yup, we are currently using NHITS, quite successfully, but just that it feels we're missing out for not leveraging the AutoNHITS given the automation

<@U03619XRWKD> if you have some recommendations on which other libraries for `hyperparameter tuning` that works well with `neuralforecast`, please let me know. At the moment, I'm thinking of going with `optuna`  <https://forecastegy.com/posts/multiple-time-series-forecasting-nbeats-python/>

This certainly looks promising! Let us know how it works.

<@U03619XRWKD> I tested `optuna` with `NHITS`, and able to run the hyperparameter tuning, and more importantly I can make use of the GPU on my windows server