Fix interaction between PyroParam and torch.func.grad #3328
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This was referenced Feb 15, 2024
|
I think I'm missing some edge cases. Marking this WIP while I strengthen the test and see if it still works... |
|
Addressed the rest of the edge cases, should be ready for review |
ordabayevy
approved these changes
Feb 16, 2024
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Addresses a bug described downstream in BasisResearch/chirho#393
This PR adds a fix for compatibility of
PyroModules andPyroParams withtorch.func.gradand the other functional automatic differentiation transforms intorch.func. The fix is basically to replace eachpyro.paramstatement or other interaction with the parameter store with a dummy version that does not store and retrieve a parameter tensor from a nonlocal state (which is invisible to the tracing machinery intorch.func).Without this fix, gradient computations in
torch.func.graddo not propagate to the unconstrained parameters behind constrainedPyroParams even when usingpyro.settings.set(module_local_param=True)and are always zero. After this fix, the functional AD system intorch.funcbehaves correctly withAutoGuides and otherPyroModules whenmodule_local_param=True, though it is still fundamentally incompatible with the global parameter store state whenmodule_local_param=False.Tested: