[Add SOTA model in Neuralforecast] : TSMixer by Google

[VinayChaudhari1996]

Description

Hello @mergenthaler
Is it possible to add new TSMixer model in library?
Reference article and code :

Code : https://www.datasciencewithmarco.com/blog/tsmixer-the-latest-forecasting-model-by-google

Paper:
https://arxiv.org/abs/2303.06053

Thanks

Use case

No response

[cchallu]

Hi @VinayChaudhari1996 ! Thanks for the suggestion. Yes, we plan to add TSMixer to the library soon.

[candalfigomoro]

How does TSMixer compare to TiDE? (#610)

[candalfigomoro]

Unofficial pytorch implementation: https://github.com/ditschuk/pytorch-tsmixer

[candalfigomoro]

Pay attention to the fact that there are apparently 2 different models named "TSMixer":

• TSMixer (by Google): https://arxiv.org/abs/2303.06053
• TSMixer (by IBM), renamed to PatchTSMixer: https://arxiv.org/abs/2306.09364

[MLfreakPy]

This would surely be another significant contribution further underscoring the strength of this library!

[cchallu]

Hi @steffenrunge. It will be a great addition! We almost have it ready, and will include it soon :).

[candalfigomoro]

Hi @steffenrunge. It will be a great addition! We almost have it ready, and will include it soon :).

@cchallu Is your implementation based on the "ext" version (like this: https://github.com/ditschuk/pytorch-tsmixer/blob/main/torchtsmixer/tsmixer_ext.py) with static / hist / future exogenous variables support?

[candalfigomoro]

I think this issue can be closed, see:

• [FEAT] Added TSMixer model #914
• [FEAT] Added TSMixerx model #921