[Examples] Generalise Seq2Seq ASR to handle Whisper

[sanchit-gandhi]

What does this PR do?

Generalises run_speech_recognition_seq2seq.py to handle Whisper.

To train the "tiny.en" model on LibriSpeech dummy:

Bash script

#!/usr/bin/env bash
CUDA_VISIBLE_DEVICES=0 python run_speech_recognition_seq2seq.py \
        --dataset_name="hf-internal-testing/librispeech_asr_dummy" \
        --model_name_or_path="openai/whisper-tiny.en" \
        --dataset_config_name="clean" \
        --train_split_name="validation" \
        --eval_split_name="validation" \
        --output_dir="./" \
        --preprocessing_num_workers="1" \
        --length_column_name="input_length" \
        --overwrite_output_dir \
        --num_train_epochs="1" \
        --per_device_train_batch_size="8" \
        --per_device_eval_batch_size="8" \
        --learning_rate="3e-4" \
        --warmup_steps="500" \
        --evaluation_strategy="steps" \
        --text_column_name="text" \
        --save_strategy="no" \
        --evaluation_strategy="epoch" \
        --logging_steps="10" \
        --save_total_limit="1" \
        --generation_max_length="40" \
        --generation_num_beams="1" \
        --fp16 \
        --gradient_checkpointing \
        --group_by_length \
        --predict_with_generate \
        --do_train --do_eval \
        --do_lower_case

To train the "medium.en" model on LibriSpeech 960h:

Bash script

#!/usr/bin/env bash
CUDA_VISIBLE_DEVICES=0 python run_speech_recognition_seq2seq.py \
    --model_name_or_path="openai/whisper-medium.en" \
    --dataset_name="librispeech_asr" \
    --dataset_config_name="all" \
    --train_split_name="train.clean.100+train.clean.360+train.other.500" \
    --eval_split_name="validation.clean" \
    --max_steps="5000" \
    --output_dir="./" \
    --run_name="whisper-librispeech" \
    --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="1000" \
    --save_strategy="steps" \
    --save_steps="1000" \
    --generation_max_length="224" \
    --generation_num_beams="1" \
    --length_column_name="input_length" \
    --gradient_checkpointing \
    --group_by_length \
    --freeze_encoder \
    --fp16 \
    --overwrite_output_dir \
    --do_train \
    --do_eval \
    --predict_with_generate \
    --use_auth_token

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you read the contributor guideline,
  Pull Request section?
• Was this discussed/approved via a Github issue or the forum? Please add a link
  to it if that's the case.
• Did you make sure to update the documentation with your changes? Here are the
  documentation guidelines, and
  here are tips on formatting docstrings.
• Did you write any new necessary tests?

Who can review?

Anyone in the community is free to review the PR once the tests have passed. Feel free to tag
members/contributors who may be interested in your PR.

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint.

[patrickvonplaten]

Thanks for iterating here!

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint.

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint.

[sanchit-gandhi]

@sgugger this one's ready to go! Just an FYI in-case you wanted to take a look :)

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint.

[sgugger]

Thanks for working on this!