Use approximate log_beta() in .fit(), .predict() #2502
Merged
Conversation
This file contains 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
fritzo
commented
May 24, 2020
Comment on lines
+48
to
+49
| n = total_count.clamp(min=0) | ||
| k = value.masked_fill(invalid, 0) |
There was a problem hiding this comment.
this is the crux of the NAN gradient fix
martinjankowiak
previously approved these changes
May 24, 2020
| @@ -23,7 +23,7 @@ class ExtendedBinomial(Binomial): | |||
|
|
|||
| def log_prob(self, value): | |||
| result = super().log_prob(value) | |||
| invalid = ~super().support.check(value) | |||
| invalid = (value < 0) | (value > self.total_count) | |||
There was a problem hiding this comment.
curous: why this change?
There was a problem hiding this comment.
The new version is a little cheaper. The old integer_interval version additionally checks value % 1 == 0, but that is already checked in validation by the above line, and it doesn't affect numerical stability.
|
Thanks for reviewing! |
martinjankowiak
approved these changes
May 25, 2020
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Outdated suggestions cannot be applied.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 #2426
Blocking #2498 (from which this simpler PR was extracted)
This PR:
ExtendedBetaBinomialrelated to nan propagates through backward pass even when not accessed pytorch/pytorch#15506. This may have caused spurious rejections during heuristic initialization.log_beta()inCompartmentalModel.fit()and.predict()now that gradients are safe. This reduces likelihood compute cost by about 14% inSuperspredingSEIRModel, and will reduce Add overdispersed models to contrib.epidemiology #2498 more.safe_log()topyro.ops.specialwhich seems like a better home. (this change is non-breaking sincesafe_log()has not yet been released).Tested
ExtendedBinomialandExtendedBetaBinomial