Feature request: per batch l2-normalization at embedding generation for training on large datasets about scalign HOT 3 CLOSED

Hi @hui2000ji ,

I see the issue, the problem is in saving the embeddings. I am going to move the l2-norm into the batch-wise for loop of encoderModel_calc_embedding in backend_encoder.R. This will also better agree with the batch-wise l2-norm being performed during training. I am going run a few unit tests then an update should be pushed early tomorrow.

Also, I think this should only be an issue for train_encoder_multi.R, unless you are setting: scAlignOptions(full.norm = TRUE). scAlignOptions(norm = TRUE, full.norm = FALSE) performs only per-batch l2-norm during training and is the default parameterization.

from scalign.

I think this sort of does the job: #14

from scalign.

I've pushed an update which passed the unit tests we have for the tutorial datasets. This should fix the memory issues you were having. Let us know if it doesn't!

Thanks for taking the time to modify the code, our updates were very similar! Hopefully we will have some time to port this code to TF 2.0 and remove tensorflow for R which has caused a few issues now.

from scalign.