Parameter autotuning

First, thank you for the great software!

I saw that the autotuning feature

seems to have dissappeared after version 0.8.X.
At least, that’s the last version that I found a tutorial for
https://docs.scvi-tools.org/en/0.8.1/user_guide/autotune.html
I also saw that an alternative is in active development

which makes me think that the older version is gone.

I tried to run the tutorial above using 0.10.X and 0.7.X of scVI, but get
"
ModuleNotFoundError: No module named ‘scvi.inference’
"
on import statements at the beginning which makes me think things have moved around
since the tutorial was written.

Is there currently a way to run autotuning?

Unfortunately there is not a currently supported way to run autotune. In the next few weeks, the version in development will be more stable. This version uses Ray tune and will be more straightforward to use.

Thank you for the quick reply!

I’ll roll my own then.
Looking forward to the new autotuner!

A big thank you for scvi-tools!
I’m curious what the current status is on Ray tune - is it still in development?
Thanks!

Slowly still in development. You can see the PR we have up. It’s mostly wrapping Ray tune.

1 Like

Hi @adamgayoso
I would like to ask about the progress of hyperparameters tuning for scvi-tools’s models. The version that wraps Ray Tune around scvi-tools’s classes, such as SCVI, SCANVI, etc. seems not work for large datasets (as mentioned elsewhere: [tune] TypeError: __init__() got multiple values for argument 'arch' when using 'with_parameters' · Issue #11047 · ray-project/ray · GitHub).
Thank you very much for such a great library. Looking forward to hearing from you.
Kind regards,

Hi @tu.le, we are finishing up development of the autotune module for scvi-tools. We are hoping to wrap this up by the end of this week as Ray 2.1 has just been released, allowing us to address some of the bugs that were present before with PyTorch Lighting 1.7, which is our minimum version. Hopefully this will fix your issue with large datasets. Stay tuned for the next version release!

1 Like

Hi Martin @martinkim0 (hope I spell your name correctly),
Thanks for your feedback! It’s great to hear that autotune function for scvi-tools models has almost been completed. I am looking forward to further information related to its release in the near future.
Best regards,