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
@samlobel and I are trying to create a replay table with prioritized sampler and remover. We are trying to remove low priority entries by negating the priority_exponent argument to the remover's constructor (the priority_exponent converts the priority scores to probabilities by raising the score to the exponent and dividing by the sum of scores raised to exponent). However, there is an assert in the cpp constructor for the remover selector that prevents us from specifying negative priority_exponents, even though this is a totally valid mathematical operation which would continue to yield valid sampling/removing probabilities. Here is an example culprit assert of code:
Yes we can go ahead and remove this restriction. I don't even remember the reasoning behind it to begin with.
The change will be submitted today and will be part of the next Reverb release (which is coupled with the next tensorflow release). In the meantime I'd suggest you build Reverb from source with this change included.
@samlobel and I are trying to create a replay table with prioritized sampler and remover. We are trying to remove low priority entries by negating the
priority_exponent
argument to the remover's constructor (thepriority_exponent
converts the priority scores to probabilities by raising the score to the exponent and dividing by the sum of scores raised to exponent). However, there is an assert in the cpp constructor for the remover selector that prevents us from specifying negativepriority_exponent
s, even though this is a totally valid mathematical operation which would continue to yield valid sampling/removing probabilities. Here is an example culprit assert of code:reverb/reverb/cc/selectors/prioritized.cc
Line 60 in 7e9ee1a
Would it be possible to remove that assert statement? Alternatively, is there a workaround?
The text was updated successfully, but these errors were encountered: