Justin Hong
Justin Hong
The documentation differentiating batch_key and categorical_covariate_keys currently does not clarify the difference between them well. From chat with Allon: 1. is there a difference if I add a discrete batch...
As mentioned in https://github.com/scverse/scvi-tools/issues/1564, if we find a way to default `training` without explicitly passing it in, this should be enough to move the necessary logic from `JaxModuleWrapper` to `JaxBaseModuleClass`...
Support for NumPyro as detailed in https://github.com/scverse/scvi-tools/issues/1366. https://github.com/scverse/scvi-tools/pull/1495 can be used as a starting point.
The jax backend currently has a very simple device management option. We would like to improve it such that users can specify specific devices to run on like with our...
Addressing comment here https://github.com/scverse/scvi-tools/blob/master/scvi/module/base/_jax_module_wrapper.py#L23
First reported in https://discourse.scverse.org/t/totalvi-nan-loss-with-few-proteins/45/3, this bug was recently reproduced in a dataset with 300 proteins. The cause of the issue is unknown, but setting `empirical_protein_background_prior=False` results in proper training. Associated...
Reimplementation of PeakVI in Jax. This is meant to both see if we can get a similar speed up in PeakVI, as well as understanding the limitations of the `JaxModuleWrapper`...
WIth the addition of AmortizedLDA, we would like to make incremental updates to improve the model. This issue will serve as the parent issue for a bunch of small features...
Is it possible this package could be added to conda forge? Although it is already on pip it would be great if conda users could also easily install pyformance!
### Question I noticed when training from scratch, the initialization of the parameters for the weight matrices in `AbstractAttention` can be particularly bad since they just use torch.empty. This ends...