search

weberlab-hhu / Helixer

Using Deep Learning to predict gene annotations
GNU General Public License v3.0
164 stars 27 forks source link

phases (the three of them + None) do not always add to 1 #66

Closed alisandra closed 3 years ago

alisandra commented 3 years ago

specifically we get a pretty scattered mix of 1, and 0.5

potential link to chunk size not being divisible by pool size is under investigation

thanks @bubu227 for both catching and investigating!

alisandra commented 3 years ago

This occurred during write to h5 where values shown as 1.000000 (presumably within float precision of 1), ended up being 0.5 when written into a float16 h5 dataset. Pre-conversion to float16 in numpy dodges the issue. Fixed at 1f90da4e804256ed37bf152761c1ebe52b02253e

soi commented 3 years ago

Thanks a lot for catching and fixing!