Can Ergen comments

Results 98 comments of


                                            Can Ergen

multi-GPU vs. single GPU - scvi 1.1.x branch

Hi @martinkim0 scVI is working with multiple samples (like n_samples_per_mc_run). Those look similar like the multi-GPU structure (n_samples, n_batch, n_genes). Quite some other functions like scANVI are not handling n_samples...

Model trained on 0.14 and model load on 0.20, error shows: no "ind_x"

I'm sorry you ran into this. MultiVI was reformatted after the initial release. We will update our release notes with scvi-tools 1.2. Specifically PR #1697 added indices to setup_anndata. This...

Implement conditional batch/layer normalization

You can copy this one from the muANVI repo and I would add a second batch key in the AnnData setup, only used for conditional norm in the encoder. We...

batch and max_kl_weight parameters ignored when mapping to scvi reference

Getting into the thread here, in all scArches tutorials it's highlighted that covariates are encoded. In our hands it doesn't make a difference for most datasets. OOD things shouldn't be...

MultiVI with Poisson/NB likelihood for ATAC?

Hi @vitkl I'm linking @marianogabitto here as he developed multiVI and was also working on a different implementation of PoissonPeakVI. What's your opinion on this request?

MultiVI with Poisson/NB likelihood for ATAC?

Happy to review the PR.

MultiVI with Poisson/NB likelihood for ATAC?

I'm sorry for the confusion that was a comment to @marianogabitto to assign me the PR if he creates one.

`use_observed_lib_size=False` fails with NaN for some datasets

Can you provide what the minimum of your library size (and maximum) is and whether this behavior is dependent on this? I have observed this when not filtering for _min_counts_....

Using torch.compile in Pyro models

Hi @vitkl, What's the purpose of the request? In our hands, speed-up was not reproducibly high. Have you made other experience? We see larger speed-up using JAX, so if it's...

Using torch.compile in Pyro models

I agree that speed-up is expected to be largely model-dependent and that scVI is small and might be a bad proxy. Adam and Martin experimented with torch.compile, however, only in...