LAION-AI / audio-dataset

Audio Dataset for training CLAP and other models
632 stars 53 forks source link

decoding speed / benchmark #12

Open faroit opened 2 years ago

faroit commented 2 years ago

This repo is great. I always wanted to benchmark webdataset for audio. A couple of questions:

  1. did you find flac to be a good trade-off between decoding performance and file-size? have you tried mp3 instead?
  2. did you benchmark the pipeline against plain torch.data with torchaudio or the new torch data pipes? Maybe adding the benchmark to https://github.com/faroit/python_audio_loading_benchmark/ to give this a go?
  3. How is partial decoding seeking be typically done with webdatasets, when storing long audio but at decoding stage, only random chunks are being read. Is seeking supported? If yes, does this slow down the i/o pipeline?