fauxneticien/w2v2-fairseq-pretrain

wav2vec 2.0 XLS-R continued pretraining using fairseq

1. Launch Docker container docker-compose run gpu

2. Fetch XLS-R checkpoint:

   wget https://dl.fbaipublicfiles.com/fairseq/wav2vec/xlsr2_300m.pt -P tmp/

3. Download folder of audio into data, e.g. data/wav

4. Create manifest of wav folder:

   python wav2vec_manifest.py /workspace/data/wav --ext wav --valid-percent 0.01 --dest data/manifest/

5. Adjust training parameters in w2v2-large-cpt.yaml (e.g. dataset.max_tokens: 150000 to whatever suits your GPU setup)

6. Run continued pretraining:

   fairseq-hydra-train task.data=/workspace/data/manifest \
       checkpoint.finetune_from_model=/workspace/tmp/xlsr2_300m.pt \
       common.log_format=simple \
       common.fp16=True \
       --config-dir /workspace \
       --config-name w2v2-large-cpt.yaml