search

haotianteng / Chiron

A basecaller for Oxford Nanopore Technologies' sequencers
Other
122 stars 53 forks source link

How to train a similar model #110

Closed mrubio-chavarria closed 3 years ago

mrubio-chavarria commented 3 years ago

Hello,

I am trying to train a basecaller from scratch. With that aim in mind, I am studying the structure behind Chiron. The problem is that I fall all the time in a local minimum in which all the predicted character is the blank character, I have not seen that described in the paper.

Did you see this problem while training Chiron? I have read that is quite a common phenomenon. If that is the case, how did you solve it?

Thanks,

haotianteng commented 3 years ago

Hi, what data are you working with? For RNA data the model is more likely to fail into the local minimum, my suggestion is to use SGD with momentum instead of Adam when a train on RAN data. And for the DNA dataset, it's less likely to fail into this local minimum, but it still can, the solution is to run several initializations.