better DC2 precision & recall for the dimmest light sources

prob-ml / bliss

Bayesian Light Source Separator

MIT License

41 stars 11 forks source link

better DC2 precision & recall for the dimmest light sources #973

Closed jeff-regier closed 4 months ago

jeff-regier commented 5 months ago

Figures 2 & 3 in the DC2 manuscript show BLISS being outperformed by the LSST pipeline for the dimmest light sources. We should confirm that these numbers are correct by rerunning everything. If they are correct, we should improve BLISS's image normalization for these dim light sources.

YicunDuanUMich commented 5 months ago

roger that

jeff-regier commented 5 months ago

If we need to improve BLISS's image normalization, the arcsinh transform (torch.arcsinh) may give us better results than log, which is what we use here: https://github.com/prob-ml/bliss/blob/3b79b7e057777bcc504331b1fa84fd33ac3c6d31/bliss/encoder/image_normalizer.py#L93

Fiddling with some of the thresholds may help too. Currently the thresholds we use for log transforms are set here: https://github.com/prob-ml/bliss/blob/3b79b7e057777bcc504331b1fa84fd33ac3c6d31/bliss/conf/base_config.yaml#L86

YicunDuanUMich commented 5 months ago

we can now outperform lsst

jeff-regier commented 5 months ago

Wonderful!