FEAT: Support exog features in TimeMixer

[marcopeix]

Fix bug that allows TimeMixer to take futr_exog now.

[review-notebook-app]

Check out this pull request on 

See visual diffs & provide feedback on Jupyter Notebooks.

---

Powered by ReviewNB

[JQGoh]

@marcopeix any chance this could be part of 1.7.4 release (understood that 1.7.4 has just been released). Sounds like this is a major issue that prevents a key part of TimeMixer working properly.

[elephaint]

@JQGoh 1.7.4 was already released today, we took out the futr_exogenous functionality of TimeMixer for now, but we can make it part of the next release.

[JQGoh]

@JQGoh 1.7.4 was already released today, we took out the futr_exogenous functionality of TimeMixer for now, but we can make it part of the next release.

Thanks for the info. I was not aware that the futr_exogenous support was removed.

[elephaint]

@JQGoh 1.7.4 was already released today, we took out the futr_exogenous functionality of TimeMixer for now, but we can make it part of the next release.

Thanks for the info. I was not aware that the futr_exogenous support was removed.

Yeah @marcopeix and I couldn't make it work easily and we didn't understand it sufficiently, and we wanted to move forward with the release...

[peacocktrain]

@marcopeix Hello, I would like to use the futr_exog feature of TimeMixer, and I have pulled down this branch to test it. However, the content provided in the ipynb file does not execute correctly, and the following error is reported:

File "/workspace/neuralforecast/common/_modules.py", line 388, in forward x = self.tokenConv(x.permute(0, 2, 1)).transpose(1, 2) RuntimeError: permute(sparse_coo): number of dimensions in the tensor input does not match the length of the desired ordering of dimensions i.e. input.dim() = 4 is not equal to len(dims) = 3

Could you please help me check if the modifications in the ipynb file can be executed correctly? Thank you.

[elephaint]

@peacocktrain This is a draft PR, which means it doesn't necessarily work. Please don't use draft PRs to work with, as functionality likely doesn't work properly.

For now, I'd advise to use another model that supports futr_exog while we work on enabling this feature for TimeMixer.

[marcopeix]

Supporting exogenous features in TimeMixer is a feature that is on our radar. Note that the model as presented in the original paper does not use exogenous features. So, in the current state, the model in neuralforecast can be used to reproduce the results of the paper.

[elephaint]

The original paper doesn't seem to support exogenous features. Upon learning this I don't see a reason to support this.

Note that it's not easy to include exogenous - it is insanely hard to do this right, so it seems a limitation of the architecture that we are not placed to solve.

I'd vote for closing unless I'm misunderstanding here.