wav2vec2-ctc-librispeech / run_librispeech.sh
esc-bencher's picture
Add training scripts and weights
4147459
raw
history blame contribute delete
969 Bytes
#!/usr/bin/env bash
python run_flax_speech_recognition_ctc.py \
--model_name_or_path="esc-benchmark/wav2vec2-ctc-pretrained" \
--tokenizer_name="wav2vec2-ctc-librispeech-tokenizer" \
--dataset_name="esc-benchmark/esc-datasets" \
--dataset_config_name="librispeech" \
--output_dir="./" \
--wandb_project="wav2vec2-ctc" \
--wandb_name="wav2vec2-ctc-librispeech" \
--max_steps="50000" \
--save_steps="10000" \
--eval_steps="10000" \
--learning_rate="3e-4" \
--logging_steps="25" \
--warmup_steps="5000" \
--preprocessing_num_workers="1" \
--hidden_dropout="0.2" \
--activation_dropout="0.2" \
--feat_proj_dropout="0.2" \
--do_train \
--do_eval \
--do_predict \
--overwrite_output_dir \
--gradient_checkpointing \
--freeze_feature_encoder \
--push_to_hub \
--use_auth_token