Test ELBO higher than train ELBO

broadinstitute / CellBender

CellBender is a software package for eliminating technical artifacts from high-throughput single-cell RNA sequencing (scRNA-seq) data.

https://cellbender.rtfd.io

BSD 3-Clause "New" or "Revised" License

285 stars 52 forks source link

Test ELBO higher than train ELBO #317

Open ellimarip opened 9 months ago

ellimarip commented 9 months ago

I have used the raw feature-barcode matrix from https://support.10xgenomics.com/single-cell-gene-expression/datasets/2.1.0/pbmc4k and have ran CellBender with defaults. I am wondering why the test ELBO seems to be higher than train ELBO?

learning_curve

Ratoncito commented 6 months ago

Yes I also have experienced this on some of my runs. Not sure how to interpret this. Perhaps this means that the pretrained model was better suited for the testing set than for the training set for some reason. Not sure why this is but am somewhat confidant this isn't a big deal/ isnt necessarily a bad thing.