Gpytorch: Multi-task classification example?

[jyhnnhyj]

Are there any examples for multi-task classification (rather than regression) similar to https://github.com/cornellius-gp/gpytorch/blob/master/examples/03_Multitask_GP_Regression/Hadamard_Multitask_GP_Regression.ipynb ?

[gpleiss]

By multi-task classification, do you mean multi-class (non-binary) classification? If so, check out the CIFAR10 example.

[jyhnnhyj]

No, I didn't mean multi class classification. What I meant was multi-task GP where e.g. you can use linear model coregionalisation (lmc) to model the relation between f's and u's (cross co-var between outputs and latent functions) for a classification task

[jyhnnhyj]

@gpleiss any suggestions?
I tried adapting 03_Multitask_GP_Regression/Hadamard_Multitask_GP_Regression.ipynb to a classification task (instead of regression), but was not successful. The issue is that I'm not sure how to pass x's and i's to the variational strategy (as in classification models, those need to be set)
really appreciate any feedback here :)

[jyhnnhyj]

also appreciate any comment from @jacobrgardner

[gpleiss]

Right now variational models are not set up to do multitask. You can train a variational model to output several independent GPs, but that is probably not what you're looking for.

@jacobrgardner has been working on some variational multitask GP models (GPRNs, etc. - see #564) which will allow you to do multitask classification. We'll get an example notebook once those models are ready.

[gpleiss]

This should now be possible with #903 in

[jyhnnhyj]

Thanks a lot @gpleiss - are there any samples? Had a quick look into examples folder and couldn't find anything there

[gpleiss]

Whoops I realized I added a duplicate issue for this (#948). I'll close this one in favor of the more recent one. Should come out soon.