Flashlight and Pyctcdecode decoders

[karpnv]

Preserve Flashlight and Pyctcdecode beamsearch with Ngram LM

Support Flashlight and Pyctcdecode decoding with pure KenLM and NeMo KenLM
Standardize API of CLI inference scripts

Collection: ASR

Changelog

• Fix install script install_beamsearch_decoders.sh
• Create flashlight_lexicon file during scripts/asr_language_modeling/ngram_lm/train_kenlm.py
• Unify parameters for eval_beamsearch_ngram_ctc.py, speech_to_text_eval.py and training
  -- Get logprobs from Hypothesis
  -- Use "pyctcdecode" strategy as default beamsearch algorithm denoted as "beam"
  -- Remove default seq2seq strategy
  -- Remove kenlm_path and add word_kenlm_path (using lmplz), nemo_kenlm_path (train_kenlm.py)
  -- Check decoding_type and search_type combinations
  -- Support empty string in nemo_kenlm_path and word_kenlm_path for beamsearch without LM (ZeroLM)
• Fix bug with EncDecHybridRNNTCTCModel in examples/asr/transcribe_speech.py
• Support AggregateTokenizer in scripts/asr_language_modeling/ngram_lm/create_lexicon_from_arpa.py

python3 scripts/asr_language_modeling/ngram_lm/eval_beamsearch_ngram_ctc.py \
model_path=am_model.nemo  \
dataset_manifest=manifest.json  \
preds_output_folder=/tmp   \
ctc_decoding.strategy=flashlight \
ctc_decoding.beam.nemo_kenlm_path=am_model.kenlm \
ctc_decoding.beam.beam_size=[4]   \
ctc_decoding.beam.beam_alpha=[0.5]   \
ctc_decoding.beam.beam_beta=[0.5] \
ctc_decoding.beam.flashlight_cfg.lexicon_path=am_model.flashlight_lexicon \ # DEFAULT_TOKEN_OFFSET
batch_size=32  \
beam_batch_size=1 \
cuda=1

python3 examples/asr/speech_to_text_eval.py  \
model_path=am_model.nemo \ 
dataset_manifest=manifest.json \
decoder_type=ctc  
ctc_decoding.strategy=flashlight \  
ctc_decoding.beam.nemo_kenlm_path=am_model.kenlm \
ctc_decoding.beam.beam_size=4   \
ctc_decoding.beam.beam_alpha=0.5   \
ctc_decoding.beam.beam_beta=0.5 \
ctc_decoding.beam.flashlight_cfg.lexicon_path=am_model.flashlight_lexicon \
ctc_decoding.beam.return_best_hypothesis=true \
batch_size=32  \
output_filename=/tmp/manifest_out.json 
cuda=1

PR Type:

• [ V] New Feature
• Bugfix
• Documentation

If you haven't finished some of the above items you can still open "Draft" PR.

Who can review?

Additional Information

• Related to "FileNotFoundError: KenLM binary file not found at : None" thrown when decoding without N-gram LM #9067

[github-actions]

This PR is stale because it has been open for 14 days with no activity. Remove stale label or comment or update or this will be closed in 7 days.

[github-actions]

This PR was closed because it has been inactive for 7 days since being marked as stale.

[michiahe]

If you install the pypi package for Flashlight decoders, the maximum supported ngram length is 6. If you try to use a longer Ngram you'll see an error even if you built kenlm with support for longer Ngrams.

This model has order 10 but KenLM was compiled to support up to 6.  If your build system supports changing KENLM_MAX_ORDER, change it there and recompile.  With cmake:
 cmake -DKENLM_MAX_ORDER=10 ..
With Moses:
 bjam --max-kenlm-order=10 -a
Otherwise, edit lm/max_order.hh.

[karpnv]

Yes, this is what I fixed for other decoders except Flashlight. Let's fix it next time

[nithinraok]

Note: eval_beamsearch_ngram_ctc.py and transducer.py I guess also requires changes for hypothesis that was updated recently for log probs.