Tuning pre-trained bonito model

[ktan8]

We've been applying Bonito to some of our PromethION datasets. However, we noticed that the model seems to be basecalling some repeat elements incorrectly. To address this, we figured that providing the model with more training examples of these repeat sequences might help base call these sequences correctly.

However, we understand that re-training the model from scratch might take a bit of time. Is there a way for us retrain the pre-trained Bonito models with our extra training examples?

Thanks!

[iiSeymour]

Good question @ktan8, this is possible. Just use --pretrained to specify a model to fine tune and lower the learning rate.

bonito training --epochs 1 --lr 5e-4 --pretrained dna_r9.4.1@v3.2 --directory new-training-data fine-tuned-model

[ktan8]

Thanks!