Fix confidence loss to scale up correctly #33
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Outdated suggestions cannot be applied.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.
Previously, weighting of confidence loss would jump from
step_frac
just beforestep_frac > warmup_frac
to1.0
whenstep_frac > warmup_frac
.For example, if
warmup_frac = 0.1
, then whenstep_frac = 0.1
,coef = step_frac = 0.1
, and then at the next stepstep_frac > 0.1 = warmup_frac
, socoef = 1.0
.The paper describes this weighting as increasing smoothly. The change makes it increase linearly from 0 to 1 during the first
warmup_frac
steps.We change
>
to>=
to allowwarmup_frac = 0.0
, then on the first stepstep_frac = 0.0
andcoef = 1.0
, otherwise we'd get a divide by 0 error on the first step.