[FEA] Gaussian Process Regressor needed

[NV-Reynaldog]

Is your feature request related to a problem? Please describe. I need the sklearn GaussianProgressRegressor available as a cuML library.

https://scikit-learn.org/stable/modules/generated/sklearn.gaussian_process.GaussianProcessRegressor.html

Describe the solution you'd like Drop-and-replace RAPIDS version of sklearn.gaussian_process.GaussianProcessRegressor

Describe alternatives you've considered CPU-only version of sklearn implementation, but it quickly becomes intractable without GPU acceleration.

Additional context . Sklearn’s implementation uses scipy underneath and the following functions are missing from cupy:

from scipy.linalg import cholesky, cho_solve

HOWEVER, NVIDIA does have a cho_solve available in cuSolver. Hopefully it's not too much work to connect cuSolver to cupy.

[github-actions[bot]]

This issue has been marked rotten due to no recent activity in the past 90d. Please close this issue if no further response or action is needed. Otherwise, please respond with a comment indicating any updates or changes to the original issue and/or confirm this issue still needs to be addressed.

[github-actions[bot]]

This issue has been marked stale due to no recent activity in the past 30d. Please close this issue if no further response or action is needed. Otherwise, please respond with a comment indicating any updates or changes to the original issue and/or confirm this issue still needs to be addressed. This issue will be marked rotten if there is no activity in the next 60d.

[berthih]

Hi, any update on this subject ?

[ianhill60]

would love this capability

[authman]

This would be a great feature to add. Let me share what my workflow currently looks like in Pyro (GPRegression, SparseGPRegression, VariationalSparseGP) vs how I'd envision a RAPIDS/cuML adoption:

• con: installation sucks, README.md and installation docs have varying instructions (README.md is accurate), and the [extra] package is necessary to use GP. VS RAPIDS conda environment install
• con: manually convert dataset from base data type -> torch tensors, and back upon completion of model inference. VS RAPIDS BYOD (bring your own datatype)
• con: manually bucket into batches for inference, to avoid gpu OoM errors VS RAPIDS "lemmie take care of that for you internally"
• con: encounter frequent torch.linalg.cholesky: U(6,6) is zero, singular U errors; occasionally at model init, and randomly during training VS RAPIDS "hey it just works"
• pro: can use apex to train on gpu by wrapping the loss function in apex
• pro: extensible - the train iteration loop is exposed, so one can add whatever additional whistles they want to the loss landscape

[teju85]

Dear @authman , Seems like gpytorch can be a good fit based on your description above. What do you think?

[teju85]

Ping @authman , did you get a chance to look at gpytorch? Does it satisfy your needs?

[authman]

@teju85 this worked =]