Skip to content

Latest commit

 

History

History
41 lines (34 loc) · 1.28 KB

notes.md

File metadata and controls

41 lines (34 loc) · 1.28 KB

espnet-based

  • evaluation

    • end2end: beam-search on scores produced by ASR, LM + whatever
    • ASR-only: greedy-search

  • pretrained espnet model

  • take pretrained librispeech model, evaluate on spanish dataset

  • no batch inference yet?

TODO

  • transformer-encoder only
    • compare vs. full-architecture
  • spanish data

deepspeech.pytorch

install apex

  • if on hpc-node: do: module load nvidia/cuda/10.1 && module load comp
  • install it: git clone https://github.com/NVIDIA/apex && cd apex && OMP_NUM_THREADS=8 pip install -v --no-cache-dir --global-option="--cpp_ext" --global-option="--cuda_ext" ./

train

  • on frontend:
OMP_NUM_THREADS=2 wandb init
  • on hpc 
    module load comp
export PYTHONPATH=$HOME/SPEECH/speech-to-text:$HOME/SPEECH/speech-recognition:$HOME/UTIL/util:$HOME/SPEECH/fairseq
WANDB_MODE=dryrun python train.py

evaluate

python evaluation.py --model libri_960_1024_32_11_04_2020/deepspeech_9.pth.tar --datasets test-clean

on hpc

module load comp export PYTHONPATH=$HOME/SPEECH/speech-to-text:$HOME/SPEECH/speech-recognition:$HOME/UTIL/util python lightning/train_lightning.py