destVI bugs

[romain-lopez]

We have identified several issues with the current version of the DestVI code:

1. (critical) the current code doesn't train at all on master. Can has flagged an issue with the nobs that needs to be resolved
2. (important, minor) there is an undesirable setup on the neural net architecture of the amortization network for the proportions (not for the gamma used throughout the paper)
3. (important, minor) there is some discrepancy in the treatment of the library size that must be figured out in between CondSCVI and DestVI

[canergen]

1 is fixed by branch fix_elbo from Adam this morning https://github.com/YosefLab/scvi-tools/tree/fix_elbo (still awaiting code review)
2 I will try with simulated data and open a branch afterwards
3 Currently running some tests on using not-log library size or log library size for both models

[canergen]

3 Using the lymph node dataset from the preprint and theory, we decided to change the library size in condVI to not using logarithm.

[romain-lopez]

Now that some of these items are (almost) decided, let's plan on starting a PR after our last chat. Additionally, let's plan to add the sparsity regularizer of cell type proportion into this PR. Making progress!

[canergen]

It's now addressed in #1457.

[giovp]

not sure if related but trying destVI on a new dataset following the tutorial, I get kernel error at this stage:

sc_model.train(
    max_epochs=250,
)

Other times instead, can't reproduce, I get nan in the encoder output (again trying to reproduce but can't rn). Few months ago it was working fine though. Will try if #1457 fix it.

[giovp]

ok, having the same issue with Stereoscope implementation, I'm afraid is independent. Am currently on a mac. Apologies for hijacking the issue.

[canergen]

The reason for the PR is not numerical instability but instability for estimated cell type proportions with two bug . If it is also an issue with Stereoscope, I would suggest to check the input data first (single dataset where this is happening) and you might want to reduce learning rate in CondSCVI

[adamgayoso]

This should have been closed by #1457