Elektrine lite

← Feed

@naturepoker@genomic.social

Post #2749777

2026-05-19 08:58 UTC

I've been wondering about role of reference assembly contiguity when training/tuning a basecaller model. Figured out how to export ONT's pytorch weights to hdf5 for processing in #Julia with NNlib, and attention map generated from POD5 is fascinating. Whatever is happening in there, the basecaller behavior differs quite a bit. #bioinformatics

Replies (1)

  • @naturepoker@genomic.social 2026-05-19 09:20

    Main difference being that weights trained on complete genome reference keeps on looking toward other distant signals to make a call on current input, while contig level genome reference trained model constantly focuses on immediate signal neighbors more often than not.

    Open ##3588557