Model: DiagVI (new)#3575
Conversation
Update feature branch to match upstream/main
…pdates Re-enable some workflows
…test-linux Chore/ci disable schedule test linux
…com/quadbio/scvi-tools into 2-basic-batch-encoding-like-in-glue
…-in-glue 2 basic batch encoding like in glue
…rs-with-guidance-graph-merge
…nce-graph-merge merge to fix divergent branches
Update test_linux_custom_dataloader.yml
…el feature imputation
…cell-type label transfer
ori-kron-wis
left a comment
ori-kron-wis
left a comment
There was a problem hiding this comment.
Impressive work, everyone involved. See my several comments for review.
There was a problem hiding this comment.
is there a test to validate the functionality of: predict_celltype, posterior_predictive_sample scarches, semi-supervised training, minification, DE & DA (new in latest version) in diagvi? those will probably be used/tried by users
There was a problem hiding this comment.
There are tests covering predict_celltype, posterior_predictive_sample, and semi-supervised training. DiagVI currently does not support scArches or minification, but we plan to add that support in the future (which requires some adaptation due to DiagVI’s multi-adata / multi-modality structure). The same applies to DE and DA.
There was a problem hiding this comment.
OK. It's fine like this; we will add any capability down the road when needed.
| return self.__class__.__name__ + "(" + args_string + ")" | ||
|
|
||
|
|
||
| class Log1pNormal(Distribution): |
There was a problem hiding this comment.
Won't just using the LogNormal above with log1p function (in the generative part) be enough to save more code? Alternatively, add a "p" parameter with a default 0 (potentially 1) and use one of the classes only.
There was a problem hiding this comment.
We now removed the LogNormal class and let ZeroInflatedLogNormal inherit from torch.distributions.LogNormal. We would therefore keep Log1pNormal as it is. Do you agree with that, or would you rather implement a more general LogNormal class with a "p" parameter?
There was a problem hiding this comment.
It's fine, probably enough for most cases in sc.
5 participants
No description provided.