search

flatironinstitute / catvae

Categorical Variational Autoencoders
BSD 3-Clause "New" or "Revised" License
22 stars 3 forks source link

Batch priors #35

Closed mortonjt closed 3 years ago

mortonjt commented 3 years ago

It looks like the batch effect correction is improving. Its not great, but... it is in the ballpark.

Two things that could likely lead to improvement

  1. Optimization strategy - I've noticed that we just run it longer, we can have better estimates of batches
  2. It is very possible that our encoder architecture just sucks. I'm tempted to add arcsine transform and have a rigorous evaluation of activation functions.
mortonjt commented 3 years ago

Ok, we got something working with just plain pseudocounts. Merging.