You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
CIP-0010: GPy2 readiness — what should be deprecated, and what should keep working?
#1297
I've been thinking about a successor to GPy (provisionally GPy2). The working assumption is that it ships under a separate import name, so GPy1 and the successor can be installed side by side and you can move one script at a time.
Before successor code lands, GPy1 should prepare the upgrade path and give visible warnings for features that may not carry forward. That work is described in CIP-0010 (Proposed). The CIP is about GPy1 readiness; successor design is a separate CIP. Framework choice (e.g. JAX) is recorded as input, not decided here.
CIP-0010 also brings together work already under way (Cython in CIP-000E, paramz in CIP-000F, parameter tying in CIP-0002, the latent-space explorer in CIP-000B, kernel coverage in CIP-000D) and adds:
A deprecation policy. Nothing that works is removed from GPy 1.x solely because it is successor-bound. Features that will not carry forward get a visible FutureWarning that names a replacement. (Today GPy silences all DeprecationWarnings on import; we plan a measured fix with narrow third-party filters so CI is not flooded.)
An upgrade contract. A recorded list of the public API (documented / __all__ / behavioural checklist: models, kernels, parameter names, optimize, predict, print(m), regex indexing, constraints, save/load) that a compatibility layer must support.
A golden-value test corpus (first pass first). Recorded GPy1 results for a small set of standard models, so "the successor gives the same answers" can be checked within tolerances rather than claimed. Expansion (SVGP, MRD, …) is backlog after the readiness bar.
Saved-model migration through to_dict / versioned JSON. We do not assume the successor can read GPy1 pickles directly.
Requirements (WHAT):REQ-0001 scripts under compatibility with at most an import change; REQ-0002 saved models load; REQ-0003 deprecations visible at least one minor release before the successor's first stable release.
The candidate deprecation table in the CIP is undecided — it is a Discussion list, not a hit list.
I'd especially like to hear from the community on these questions:
Which features do you use? In particular: MPI (SparseGP_MPI, var_dtc_parallel), the GPU psi-statistics (PyCUDA), GPy.inference.mcmc, GPy.util.datasets. These are deprecation candidates, and I don't want to drop something people depend on.
Do you maintain custom kernels or likelihoods? If so, roughly how many, and do they implement hand-written gradients (update_gradients_full, dlogpdf_dlink, …)? We are considering a GPy1 helper where you only write K/Kdiag (or logpdf) and GPy1 supplies finite-difference gradients — so custom code is already in the shape a successor would need.
How do you save models? Pickle, m.save / GPy.load, or to_dict?
Import name for the successor:gpy2, GPy2, or something else?
What would make an upgrade painless for you? For example a conversion tool, a migration guide, or running both versions side by side for a while.
Anything missing from the deprecation candidate list in the CIP, or anything on it you need kept?
Comments on the CIP itself — including the readiness bar, workstreams, open questions, and the appendix of design lessons from other GP libraries — are very welcome here too.
(If the CIP / requirements links 404 briefly: they land when the Proposed files are on main.)
reacted with thumbs up emoji reacted with thumbs down emoji reacted with laugh emoji reacted with hooray emoji reacted with confused emoji reacted with heart emoji reacted with rocket emoji reacted with eyes emoji
Uh oh!
There was an error while loading. Please reload this page.
Uh oh!
There was an error while loading. Please reload this page.
I've been thinking about a successor to GPy (provisionally GPy2). The working assumption is that it ships under a separate import name, so GPy1 and the successor can be installed side by side and you can move one script at a time.
Before successor code lands, GPy1 should prepare the upgrade path and give visible warnings for features that may not carry forward. That work is described in CIP-0010 (Proposed). The CIP is about GPy1 readiness; successor design is a separate CIP. Framework choice (e.g. JAX) is recorded as input, not decided here.
CIP-0010 also brings together work already under way (Cython in CIP-000E, paramz in CIP-000F, parameter tying in CIP-0002, the latent-space explorer in CIP-000B, kernel coverage in CIP-000D) and adds:
FutureWarningthat names a replacement. (Today GPy silences allDeprecationWarnings on import; we plan a measured fix with narrow third-party filters so CI is not flooded.)__all__/ behavioural checklist: models, kernels, parameter names,optimize,predict,print(m), regex indexing, constraints, save/load) that a compatibility layer must support.to_dict/ versioned JSON. We do not assume the successor can read GPy1 pickles directly.The candidate deprecation table in the CIP is undecided — it is a Discussion list, not a hit list.
I'd especially like to hear from the community on these questions:
SparseGP_MPI,var_dtc_parallel), the GPU psi-statistics (PyCUDA),GPy.inference.mcmc,GPy.util.datasets. These are deprecation candidates, and I don't want to drop something people depend on.update_gradients_full,dlogpdf_dlink, …)? We are considering a GPy1 helper where you only writeK/Kdiag(orlogpdf) and GPy1 supplies finite-difference gradients — so custom code is already in the shape a successor would need.m.save/GPy.load, orto_dict?gpy2,GPy2, or something else?Comments on the CIP itself — including the readiness bar, workstreams, open questions, and the appendix of design lessons from other GP libraries — are very welcome here too.
(If the CIP / requirements links 404 briefly: they land when the Proposed files are on
main.)All reactions