Nixtla's Open Source Time Series Ecosystem.

Nixtla Community

I did installation before as: pip install neuralforecast
do i still need to run below?
pip install git+<https://github.com/Nixtla/neuralforecast.git@nbeats_predict>

I think that branch was already merged, can you tried without it?

I got the same error when run:
```Y_df, _, _ = LongHorizon.load(directory='./', group='ETTm2')```


```TypeError: sort_values() got an unexpected keyword argument 'ignore_index'```


Screen Shot 2022-07-15 at 4.14.41 PM.png

I am unable to replicate the ignore_index problem

Do you have the latest neuralforecast version <@U03QF5D8APJ>?

I installed neuralforecast via pip last week: pip install neuralforecast

I was unable to replicate the 'ignore_index' bug in my runs of the <https://colab.research.google.com/drive/1WjBbQzaivQhOldGolzymOtLmo6QX4Ieg#scrollTo=HXKT2-fpUD0Z|N-HiTS Colab Example>
I don't know if Colab can vary in its execution of pandas versions

Here is the original code for the Long horizon datasets:
<https://github.com/Nixtla/neuralforecast/blob/main/nbs/data_datasets__long_horizon.ipynb>

I am checking on google if any similar issue. pandas version might be a good point as well. Thanks.

guess it is related to pandas version..
<https://discuss.streamlit.io/t/streamlit-typeerror-sort-index-got-an-unexpected-keyword-argument-ignore-index/15960>
but after i upgrade pandas to the latest.. it start giving another issue when
import neuralforecast as nf
        module 'pandas' has no attribute 'Panel'
panel might has been removed from Pandas module 0.25.0 onwards.
<https://stackoverflow.com/questions/65600306/module-pandas-has-no-attribute-panel>
what pandas version is neuralforecast using?

interestingly, the example need both 'Panel' and 'ignore_index' from pandas. while 'Panel' was removed since version 0.25.0, 'ignore_index' was added to DataFrame.sort_values() since version 1.0.0 onward. so logically these two requirements can not be fulfilled both no matter what version pandas to be installed. Not sure how the colab setup the dependency to make it work.. I was confused.

opened an issue: <https://github.com/Nixtla/neuralforecast/issues/270>

Hi, when I try on colab below statement:
model = nf.models.nbeats.nbeats.NBEATS(
    n_time_in=input_size, n_time_out=output_size,
    n_x=4, n_x_hidden=[4],
    frequency='W-TUE', seasonality=4
)
got an error:
TypeError: __init__() got an unexpected keyword argument 'seasonality'
any idea about the issue? Thanks.
<https://github.com/Nixtla/neuralforecast/blob/main/examples/getting_started.ipynb>

Tried it on colab, it is successful without issue.

I remember on my local machine, was getting the issue that pandas version could not meet both requirements, will try it again.

Hey Musa did you solve the pandas version problem?

So sorry, can you open an issue, we have been putting all our attention to the <https://github.com/Nixtla/statsforecast|StatsForecast> package lately.  We will be attending this next week.

I opened one on July 18:
<https://github.com/Nixtla/neuralforecast/issues/270>

The 'Nixtla/hierarchicalforecast' repository doesn't contain the 'examples/TourismSmall.ipynb' path in 'main'.

The example notebook is not available, pls update

Have you tried thid one: <https://github.com/Nixtla/hierarchicalforecast/blob/main/nbs/examples/TourismSmall.ipynb>

We have this work in progress paper:
<https://arxiv.org/abs/2110.13179|https://arxiv.org/abs/2110.13179>

Please let us know what your thoughts are after trying it out.

We have used &gt;=1.3.0, you are probably right, and downgrade to an older version.

Screenshot 2022-09-13 at 16.55.38.png

Great to see how much conformal is growing. We are also planning to include it at some point :slightly_smiling_face:

Hi <@U0449S695P0>! Yes, this is the default behaviour of all these models. It will train a single model for all the time-series in the dataset

Hello, <@U03619XRWKD> Thanks for the quick response. So, we can use the trained single model on multiple time series to forecast an individual time series from the data set, correct ?

Hey Kumar,

We are refactorizing NeuralForecast's code but here is the github README.md with the pip and conda installation lines:
<https://github.com/Nixtla/neuralforecast>

And here is a working GPU example that runs on Google Colab:
<https://colab.research.google.com/github/Nixtla/neuralforecast/blob/main/examples/mqnhits.ipynb>

image.png

Hi <@U0449S695P0>! Yes, we deprecated the seasonality hyperparameter.

HI <@U03619XRWKD>: Is the entire getting_started.ipynb deprecated? I am getting the below error now.  On the 1st of October, the below cell was running in the colab but now it is Not. Could you please point out the working starting notebooks that from which I start my journey?

Hi <@U0449S695P0>, yes! We are releasing a new version of NeuralForecast, with a much simpler interface, more capabilities and models! We are working on the new examples, but you can start looking to the documentation at <https://nixtla.github.io/neuralforecast/examples/getting_started_with_nbeats_and_nhits.html>

Thanks, <@U03619XRWKD> for sharing the updated example for the NBEATS and NHITS. It is running smoothly on the colab. Are there any plans to add examples for the Global time series forecasting using the NBEATS, NHITS, and other deep learning models?

Yes, we will add more examples. This pipeline is already a global model, so if you add more series in the dataframe a single model will be used for all

Ohh..Ok. Since by default the Airline passenger data is univariate time series So, I asked the question regarding it. Thanks again for the help. Please keep posting all the latest updates in the group so we test them on our datasets.

Hi <@U044J3FUXGF>, thanks for your interest on our N-HiTS model. We just released a new version of our library with a much easier interface. We will incorporate exogenous variables for the N-HiTS in the following days, so I suggest you to try our new interface :slightly_smiling_face:

Hey Crisitan, thank you coming back on that! Sweet, then I'll look for the new release and the new documentation in a few days! Thanks!!

The first version of the new release is out :slightly_smiling_face:

I started playing with the new release - do I get this right that the 'auto' models are used for like hyperopt or what is the difference between those AutoNHITS and NHITS model classes?

The 'auto' models already perform hyperparameter using Ray Tune