Convert Fairseq data2vec1 to HF

This code found in this repository was adapted from this original HuggingFace repository. This repository contains two scripts that convert a fairseq data2vec1 checkpoint to HuggingFace 🤗 Transformers.

Procedure

1. Create a HF repo :

huggingface-cli repo create <model-name> --organization <org_of_model>
git clone https://huggingface.co/<org_of_model>/<name_of_model>

2. Convert the model

./run_convert.sh \
    --hf-path </path/to/local/hf/repo> \
    --fairseq-path </path/to/fairseq/checkpoint> \
    --size {base, large} \
    [--dict </path/to/dict>] \
    [--copy-fairseq-model]

3. Verify models behave equally

./run_forward.py \
    --hf-path </path/to/local/hf/repo> \
    --fairseq-path </path/to/fairseq/checkpoint> \
    [--finetuned </path/to/dict>]

4. Push to hub

huggingface-cli upload <your-org>/<your-model> </path/to/local/hf/repo>

Changelog

convert_data2vec1_audio_original_pytorch_checkpoint_to_pytorch.py (originally from official huggingface /transformers) was modified.

0. The fairseq model is properly imported (and class name clashes are correctly handled)

1. sampling_rate and do_normalize are both extracted from the fairseq's original configuration (e.g. cfg['task']['sample_rate']) instead of being guessed.

2. Creates preprocessor_config.json which the original didn't do for pre-trained (i.e. non-finetuned) models

3. run_forward.py was adapted from the wav2vec2 conversion script