How to choose loss functions for a different dataset.

RetroCirce / HTS-Audio-Transformer

The official code repo of "HTS-AT: A Hierarchical Token-Semantic Audio Transformer for Sound Classification and Detection"

https://arxiv.org/abs/2202.00874

MIT License

341 stars 62 forks source link

Closed the6thsense closed 1 year ago

the6thsense commented 1 year ago

I am working on a dataset other than those used in paper. What factors should I consider to choose the loss function for it.

RetroCirce commented 1 year ago

Hi,

If the dataset is designed for multi-class classification (i.e. many targets for one sample), BCE would be a good use, Otherwise, CE is a good use.

Best, Ke