Docs on Model Evaluation pipeline

descriptinc / descript-audio-codec

State-of-the-art audio codec with 90x compression factor. Supports 44.1kHz, 24kHz, and 16kHz mono/stereo audio.

https://descript.notion.site/Descript-Audio-Codec-11389fce0ce2419891d6591a68f814d5

MIT License

1.23k stars 116 forks source link

Docs on Model Evaluation pipeline #72

Open Sonal-Monteiro opened 6 months ago

Sonal-Monteiro commented 6 months ago

I am trying to reproduce the DAC training and evaluation. I am getting negative values for SI-SDR loss even for the pretrained weights provided in the repo. Could you please provide documentation for evaluating the model? Thanks

yinyxl commented 3 months ago

Me too. I got negative values when calculating the SI-SDR and I think the evaluation script uses SI-SDR loss, not SI-SDR