Skip to content

Latest commit

 

History

History
executable file
·
39 lines (20 loc) · 1.6 KB

File metadata and controls

executable file
·
39 lines (20 loc) · 1.6 KB

Input Parameters description


For prepare_data

Wav files and its corresponding text files should be present in the same folder with same name. eg - audio_id/audio.wav, audio_id/audio.txt

make_valid_from_train : If valid set is already present in different folder make this 0 else make it 1 to extract valid percentage from train data

train_wav_path: Directory where your train data(wav files) is present, if wav files are present in multiple folders put them under one parent directory

valid_wav_path: Used only when 'make_valid_from_train' is 0; directory where your valid data(wav files) is present, if wav files are present in multiple folders put them under one parent directory

valid_percentage: Used only when 'make_valid_from_train' is 1; percentage of data to be used for validation purpose. eg - 0.04 if 4%

prep_scripts: Path for utility scripts

For start_finetuning

config_name: This file contains configurable parameters for finetuning

gpus: Number of gpus to use

run_in_nohup: Make it 1 for running training in background

data_path: Contains files made from running prepare_data.sh

checkpoints_path Directory to save checkpoints generated after each epoch during finetuning. eg- checkpoint_best, checkpoint_last

log_path: nohup.out is saved here as <timestamp_of_running>.out

tensordboard_path: Path where tensorboard logs are to be saved

pretrained_model_path: Path of checkpoint_best.pt generate from pretraining

update_freq: To simulate n gpus by k gpus update frequency will be [n/k]

wav2vec_repo_path: Path of fairseq repository