Open mjsteinbaugh opened 3 years ago
HI @mjsteinbaugh , i would be happy if you submit the PR
OK great I'll work on that when I have some free time
@Koncopd can we make Bioconda release?
I'll try to work on making this happen
that would be great since we are releasing a new version next week
OK @M0hammadL I'll work on a draft recipe over the weekend
@mjsteinbaugh is there a reason for bioconda as opposed to conda-forge? Most scanpy-related stuff is on conda-forge and I'm trying to move scvi-tools there right now.
Either should be fine although most bioinformatics tools, including single-cell genomics utilities, are currently on bioconda
I always found it harder to get anything updated quickly with bioconda unfortunately. I like that conda-forge seems to have better infrastructure for managing things yourself.
Hi everyone - @mjsteinbaugh - did scarches ever end up on conda at all? Would love it if so, please let me know anything I can do to help.
I should add that scvi-tools is on conda and contains the functionality to use scarches with scVI, scANVI, and totalVI.
Oh that's helpful. Can you point to docs on how to do this?
To be clear, any scvi-tools model that is in this repo is directly using the scvi-tools implementation, instead of a reimplementation.
@dannykwells-sab @Koncopd we will have the conda release nextweek containing 11 reference models
@dannykwells-sab Haven't had free time to work on this, but may have some bandwidth this weekend.
I have another issue, maybe for @adamgayoso
When we're using scarches-scvi with GPU, we are getting this message
sca.models.SCVI.setup_anndata(source_adata, batch_key=condition_key) WARNING:jax._src.lib.xla_bridge:No GPU/TPU found, falling back to CPU. (Set TF_CPP_MIN_LOG_LEVEL=0 and rerun for more info.)
Even given this, the GPU acceleration is clearly working. when we do
vae = sca.models.SCVI()
vae.train()
We can confirm that pytorch can see cuda.
Any thought on what we can do to address this, or what it's doing?
This is nothing to worry about, models are trained using pytorch. What you're seeing there is a warning derived from a jax-based function we use to check if the user supplied data is count data.
If the train command says that it's using the gpu then it's using the gpu.
Would it be possible to publish this on Bioconda in addition to PyPi? I'm happy to help write and submit a recipe.