Update Nn::LabelScorer: allow enabling of specific transition types #148
Conversation
curufinwe
changed the title
SimBe195
changed the title
| INITIAL_LABEL, | ||
| INITIAL_BLANK, | ||
| }; | ||
| break; |
There was a problem hiding this comment.
It's related to this PR #152, but TRANSDUCER and LM need additonally SENTENCE_END transition types.
| case TransitionPresetType::LM: | ||
| enabledTransitionTypes_ = { | ||
| LABEL_TO_LABEL, | ||
| INITIAL_LABEL, |
There was a problem hiding this comment.
The AedTreeBuilder and labelsync searches are still WIP and not merged yet, but at some point we will also need a preset for AED and I guess this will be the same as this one. So will we then just add a preset which is exactly the same, just with a different name? And if yes, should LM or AED be the default preset of the StatefulOnnxLabelScorer? I mean in the end it's the same, but it might become confusing because of the naming.
src/Nn/LabelScorer/LabelScorer.cc
Outdated
| : Core::Component(config) {} | ||
| : Core::Component(config), | ||
| enabledTransitionTypes_() { | ||
| enableTransitionTypes(config); |
There was a problem hiding this comment.
Sorry, I have to withdraw may approve. I just found a bug.
Here, you are calling enableTransitionTypes(config) and in enableTransitionTypes() the function defaultPreset() is called. As this is done in the constructor of LabelScorer, the defaultPreset() function of LabelScorer will always be called instead of the implementation of the derived classes one is using.
As a fix, I suggest to put enableTransitionTypes(config) to the constructor of every derived class and remove it from the base class constructor here.
There was a problem hiding this comment.
If fixed it now by removing the defaultPreset function and instead passing the default preset to the LabelScorer parent via constructor parameter.
curufinwe
changed the title
curufinwe
changed the title
This adds a parameterto the LabelScorer base class which allows users to specify a list of transition types which will always get assigned score 0 and not affect the ScoringContext. This can be especially helpful when building combined label scorers with multiple sub-scorers of different label topology such as for example CTC + LSTM LM where the LM needs to ignore all blank transitions as well as label loops and CTC on the other hand needs to ignore the sentence-end.ignored-transition-typesDo achieve this,
extendedScoringContext,computeScoreWithTimeandcomputeScoresWithTimesare implemented in the baseLabelScorerand call protected functions{extendedScoringContext|computeScoreWithTime|computeScoresWithTimes}Internalwhich are overridden in child classes.Depends on modifications to the transition type handling from #138.
Edit:
This logic was changed so that instead of a blacklist, the LabelScorer has a parameter
transition-presetwhich enables a preconfigured list of used transition types for specific models and any additional ones can be enabled via another parameterextra-transition-types. The available presets for now are"default","none","ctc","transducer"and"lm". When thedefaultpreset is used, it takes a preset that is specified by each LabelScorer subclass individually and should be the most common use-case for that class. For example forCombineLabelScorerthe default will use preset "all", forNoContextOnnxLabelScorerthe default will use preset "ctc" and so on.