Why model.log_prediction=false is not working in NeMo 2.1.0

[mehadi92]

I was tryting to finetune ASR model. But can't disable log_prediction

python speech_to_text_finetune.py \
    --config-path="/app/conf" \
    --config-name="fastconformer_hybrid_transducer_ctc_bpe.yaml" \
    model.train_ds.manifest_filepath="/data/manifest.json" \
    ++model.log_prediction=false \
    model.train_ds.batch_size=12 \

[AudranBert]

Hi,
I personally set it to False just before launching training if it can help:

asr_model = get_base_model(trainer, cfg)
asr_model.wer.log_prediction=False
trainer.fit(asr_model)

But for hybrid decoding "nvidia/stt_be_fastconformer_hybrid_large_pc" I did not successfully hide both predictions.

[mehadi92]

Hello @AudranBert,

Do you mean setting asr_model_.wer.log_prediction=False before this line

Thanks

[AudranBert]

Hi, yes before this line. If you find a way for hybrid decoding models to hide both predictions, I will be happy to know it !