metadata
language:
- en
tags:
- esb
datasets:
- esb/datasets
- edinburghcstr/ami
To reproduce this run, first install Whisper from the Transformers compatible repo patrickvonplaten/whisper:
pip install git+https://github.com/openai/whisper.git
Then execute the command:
#!/usr/bin/env bash
CUDA_VISIBLE_DEVICES=0 python run_speech_recognition_whisper.py \
--model_name_or_path="medium.en" \
--dataset_name="esb/datasets" \
--dataset_config_name="ami" \
--max_steps="2500" \
--output_dir="./" \
--run_name="whisper-ami" \
--dropout_rate="0.1" \
--wandb_project="whisper" \
--per_device_train_batch_size="64" \
--per_device_eval_batch_size="16" \
--logging_steps="25" \
--learning_rate="1e-4" \
--warmup_steps="500" \
--report_to="wandb" \
--preprocessing_num_workers="16" \
--evaluation_strategy="steps" \
--eval_steps="500" \
--save_strategy="steps" \
--save_steps="500" \
--generation_max_length="224" \
--length_column_name="input_lengths" \
--gradient_checkpointing \
--group_by_length \
--freeze_encoder \
--fp16 \
--overwrite_output_dir \
--do_train \
--do_eval \
--do_predict \
--predict_with_generate \
--use_auth_token