[Breaking Change] Remove last_dim_is_batch #2544
Conversation
gpleiss
force-pushed
the
remove_last_dim_is_batch
branch
2 times, most recently
from
43f680d to
41a08c5
Compare
|
Still todo: update example notebooks, and add new documentation on additive / product kernels. |
gpleiss
force-pushed
the
remove_last_dim_is_batch
branch
from
41a08c5 to
e97d874
Compare
|
How are you thinking about warning users about breaking changes -- e.g. will there be deprecation warnings? (I do see that this is for merge into the |
|
@esantorella good question, and this is a good question for other GPyTorch 2.0 breaking changes more broadly. We could release a new minor release while we're still preparing 2.0 that adds deprecation warnings to notify users about breaking changes. @jacobrgardner @JonathanWenger thoughts? |
|
I think a minor release that has deprecations makes a lot of sense. It's not a lot of extra work to add the deprecations, and it might prevent some issues coming our way right after the planned 2.0 release. |
gpleiss
force-pushed
the
remove_last_dim_is_batch
branch
from
e97d874 to
41f6ae8
Compare
gpleiss
changed the title
gpleiss
force-pushed
the
remove_last_dim_is_batch
branch
2 times, most recently
from
bbea4f9 to
8d859b9
Compare
gpytorch/variational/grid_interpolation_variational_strategy.py
Outdated
Show resolved
Hide resolved
**Note**: This is not a breaking change; "legacy" grids were deprecated pre v1.0.
…Kernel - The functionality of both kernels has not disappeared, but both kernels cannot work without the last_dim_is_batch_option. - The examples/00_Basic_Usage/kernels_with_additive_or_product_structure.ipynb notebook describes how to replicate the functionality of both kernels without last_dim_is_batch.
- The functionality of this kernels has not disappeared, but this kernel cannot work without the last_dim_is_batch_option. - The examples/00_Basic_Usage/kernels_with_additive_or_product_structure.ipynb notebook describes how to replicate the functionality of this kernel using the gpytorch.utils.sum_interaction_terms utility.
Addresses PR review comments
gpleiss
force-pushed
the
remove_last_dim_is_batch
branch
from
8d859b9 to
e4ec21b
Compare
This PR (finally) bumps our PyTorch requirement to 2.0.
Most use cases are not affected. There are two breaking changes as part of this PR:
AdditiveStructureKernelandProductStructureKernel. The functionality for these kernels can be implemented manually, and there is a new example notebook showing how to do so.NewtonGirardAdditiveKernel. Again, the functionality of this kernel can be implemented manually, using the newsum_interaction_termsutility (which contains the the meat of the NewtonGirardAdditiveKernel anyways).