Nixtla's Open Source Time Series Ecosystem.

Nixtla Community

Screenshot 2023-12-12 at 13.22.37.png

Hi <@U03L39VLJBA>! It stopped because the max steps is 200. Pytorch Lightning prints outputs based on epochs and paints the bar red when training stops in the middle of an epoch.

Yes you need to clip negative values afterwards. Alternatively, you can try training with the Poisson distribution loss or the PMM, which always restrict to positive values.

<@U03619XRWKD> thank you - is there an easy way to specify optimal training length? Should I increase max steps - in such case would it apply stopping?

Yes, increase `max_steps` to at least 1k. If you have several thousands increase it further, to 5k-10k. This in our experience is more than enough for the NBEATS/NHITS. To add validation you need to specify a `val_size` on the `fit` method. And you can control how often the validation loss is computed with parameter `val_check_steps` of the model.