Add sampling strategies to beam search (#4768)

* add node and beam samplers * refactor * get stochastic beam search working * Add MultiomialSampler, TopPSampler, TopKSampler to beam_search.py, and tests for those samplers and stochastic_beam_search * Update changelog and finalize documentation * set default to without replacement * Updated TopPSampler to remove loop, with testing and bugfix. Cleaned up documentation and sampeler code. * added p sampler test * Better error messages * Update allennlp/nn/beam_search.py * lint * default to top-k if insufficient examples when top-p sampling * formatting * minor clean up * fix CHANGELOG Co-authored-by: Jackson Stokes <[email protected]>
allenai · Nov 11, 2020 · 9f7cc24 · 9f7cc24
1 parent f6fe8c6
commit 9f7cc24
Show file tree

Hide file tree

Showing 3 changed files with 703 additions and 75 deletions.
diff --git a/CHANGELOG.md b/CHANGELOG.md
@@ -12,6 +12,9 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 - Added an optional `seed` parameter to `ModelTestCase.set_up_model` which sets the random
   seed for `random`, `numpy`, and `torch`.
+- Added sampler class and parameter in beam search for non-deterministic search, with several
+  implementations, including `MultinomialSampler`, `TopKSampler`, `TopPSampler`, and
+  `GumbelMaxSampler`. Utilizing `GumbelMaxSampler` will give [Stochastic Beam Search](https://api.semanticscholar.org/CorpusID:76662039).
 
 ### Changed