mrVI hyperparameter tunning

Hi all,
We get great results from mrVI and we want to further make sure we use it t optimally.
We would like to do a quick HPO on mrVI and would love to get your experienced input on which parameters to focus? When running HPO on scVI, we focused on N_latent, N_hidden and learning rate. Are there additional parameters you advise to scan for mrVI? thanks!!

Hi @shani.bmo ! Thanks for your interest in using MrVI. It’s really exciting that you are seeing good results from it. We have an upcoming reimplementation of MrVI with significant modifications, which I will recommend to use once it comes out. But in the meantime I would recommend focusing on the n_latent, epochs/learning rate, kl annealing (n_epochs_kl_warmup), hvg selection (be careful what metric you use for this), choice of sample / batch covariates.

1 Like

@Justin_Hong thank you for your fast response and for your work :slight_smile: looking forward for the new version.