Add tree label-synchronous beam-search algorithm #129
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
…lattice from beam
larissakl
requested review from
SimBe195 and
curufinwe
and removed request for
curufinwe
# Conflicts: # apptainer/2022-10-21_tensorflow-1.15_arm_v1/makefiles/Modules.make # apptainer/2022-10-21_tensorflow-1.15_v1/makefiles/Modules.make # apptainer/2023-05-08_tensorflow-2.8_v1/makefiles/Modules.make # apptainer/2023-08-09_tensorflow-2.8_onnx-1.15_v1/makefiles/Modules.make # apptainer/2023-11-08_tensorflow-2.14_v1/makefiles/Modules.make # apptainer/2025-04-23_tensorflow-2.17_onnx-1.20_v1/makefiles/Modules.make # src/Search/Makefile # src/Search/Module.cc # src/Search/Module.hh
hannah220
requested changes
Jan 7, 2026
| currentToken(extension.nextToken), | ||
| currentState(extension.state), | ||
| lmHistory(extension.lmHistory), | ||
| length(base.length + 1), |
Contributor
There was a problem hiding this comment.
As in LexiconfreeLabelsyncBeamSearch, shouldn't length be incremented depending on the transitionType?
| extension.timeframe + 1, | ||
| {extension.score - extension.lmScore, extension.lmScore}, | ||
| {})); | ||
| break; |
Contributor
There was a problem hiding this comment.
Maybe it's better without empty lin betwen case
| } | ||
| break; | ||
|
|
||
| default: |
Contributor
There was a problem hiding this comment.
What about LABEL_LOOP?
LexiconfreeLabelsyncBeamSearch also contains blank, should this also include it?
| sentenceEndLemma = lexicon_->specialLemma("sentence-boundary"); | ||
| } | ||
| sentenceEndLabelIndex_ = sentenceEndLemma->id(); | ||
| log() << "Use sentence-end index " << sentenceEndLabelIndex_ << " inferred from lexicon"; |
Contributor
There was a problem hiding this comment.
Also better to have option to read from paramSentenceEndLabelIndex
| Nn::LabelIndex tokenIdx = network_->structure.state(successorState).stateDesc.acousticModel; | ||
|
|
||
| auto transitionType = Nn::LabelScorer::TransitionType::LABEL_TO_LABEL; | ||
| if (hyp.currentToken == Core::Type<Nn::LabelIndex>::max) { |
Contributor
There was a problem hiding this comment.
Suggested change
| if (hyp.currentToken == Core::Type<Nn::LabelIndex>::max) { | |
| if (hyp.currentToken == Nn::invalidLabelIndex) { |
| } | ||
| }; | ||
|
|
||
| recombinedHypotheses_.clear(); |
Contributor
There was a problem hiding this comment.
// Reserve capacity because future reallocations would break the raw pointer we are storing later
recombinedHypotheses_.reserve(newBeam_.size());
| break; | ||
|
|
||
| default: | ||
| defect(); // Unexpected transition type which can not be produced by `inferTransitionType` |
Contributor
There was a problem hiding this comment.
Also, why isn't inferTransitionType function defined in this class?
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.
This adds a label-synchronous search algorithm on a search tree built by the AedTreeBuilder (#127).
It is derived from the lexicon-free label-synchronous beam-search algorithm (#126). Similar to the time-synchronous treesearch (#113), global or separate pruning of within-word and word-end hypotheses (within the set of active hypotheses) is possible and a language model score is added a word end.