Skip to content

nemo_asr

Directory actions

More options

Directory actions

More options

Latest commit

 

History

History

nemo_asr

readme.md

NeMo

setup

docker build -t s2t -f Dockerfile_nemo .
# see modern_asr.md in my notes
apt-get install libsox-fmt-mp3
cd NeMo && pip install -r requirements/requirements_asr.txt
cd NeMo &&
  • tensorboard
# tensorboard
pip uninstall -y tensorboard
pip install nvidia-pyindex
pip install nvidia-tensorboard-plugin-dlprof


conda install -c conda-forge tensorboard -y
tensorboard --bind_all --logdir nemo_experiments/

-> TODO: issue remains! still not working inside of container!

reproduce Sahu's results

export repo_path="/code/spanish_nemo_asr_sahu/nemo_asr_app"
python tools/NeMo/convert_old_jasper.py --config_path="${repo_path}/tools/NeMo/example_configs/config_es.yaml" --encoder_ckpt="${repo_path}/models/es_5d_mcv_finetuned/JasperEncoder-STEP-386304.pt" --decoder_ckpt="${repo_path}/models/es_5d_mcv_finetuned/JasperDecoderForCTC-STEP-386304.pt" --output_path="/data/es_finetuned.nemo"
python tools/NeMo/convert_old_jasper.py --config_path="${repo_path}/tools/NeMo/example_configs/quartznet15x5-es.yaml" --encoder_ckpt="${repo_path}/models/es_5d_mcv_finetuned/JasperEncoder-STEP-386304.pt" --decoder_ckpt="${repo_path}/models/es_5d_mcv_finetuned/JasperDecoderForCTC-STEP-386304.pt" --output_path="/data/es_finetuned.nemo"
  • not working!

fine-tuning

python nemo_asr/speech_to_text_finetune.py model.train_ds.manifest_filepath

TODO

  • unicode for manifests: json.dump(metadata, f, ensure_ascii=False)

Info

rclone sync -P --exclude ".git/**" --exclude ".idea/**" --exclude "build/**" --exclude "*.pyc" --max-size 100k $HOME/code/SPEECH/NeMo dertilo-googledrive:NeMo

Questions

  • why is preprocessor and spec_augmentation done within forward? why not in dataloader?
  • why not black formatted?
  • no sortish sampler or bucketing? "simply" take huggingface's DistributedSortishSampler
  • why soundfile which is unable to read mp3 ?
  • what about environment.yml ?
  • what is strict=False in from_pretrained good for?

why not mp3 ?

569M	/content/LibriSpeech/dev-other-processed_wav
60M	/content/LibriSpeech/dev-other-processed_mp3
78M	/content/LibriSpeech/dev-other-processed_mp3_32
149M	/content/LibriSpeech/dev-other-processed_mp3_64

wav_vs_mp3

Sagemaker

build container
  • --force-reinstall (see Dockerfile) makes Dockerfile practically non-appendable (cause it reinstalls big+fat torch!!), thus produces a base-image from which one can inherit
docker build -f Dockerfile_base -t 706022464121.dkr.ecr.eu-central-1.amazonaws.com/pytorch-nemo:1.6.0-cpu-py3-base .
aws ecr get-login-password --region eu-central-1 | docker login --username AWS --password-stdin 706022464121.dkr.ecr.eu-central-1.amazonaws.com/pytorch-nemo
docker push 706022464121.dkr.ecr.eu-central-1.amazonaws.com/pytorch-nemo:1.6.0-cpu-py3-base
  1. data *