Question about training dataset for pre-trained model

[Howard-Liang]

Hi, I'm trying to reproduce the pre-trained bonito model. However, with the same model architecture and the dataset provided by using: bonito download --training, my models seem to converge with accuracy lower approximately 0.5~0.8% than the pre-trained model. I'm wondering was the pre-trained model trained only with the provided dataset? Thank you.

[marcpaga]

According to: https://github.com/nanoporetech/bonito/issues/69 (end of the issue)

The pre-trained dataset is trained on other data.

[Howard-Liang]

Isn't the dataset now consisting of 1221470 reads? Thanks.

[marcpaga]

Ah, then they updated the dataset. My answer is invalid then.

[danwarrior]

Where is stored the reads data after running bonito download --training ? I was expecting a .fast5 as a result but I just got some .npy files

[iiSeymour]

@danwarrior see https://github.com/nanoporetech/bonito/issues/4#issuecomment-562876731

[iiSeymour]

@Howard-Liang no, the available training set does not fully represent that used for the pre-trained models.