MouseLand / cellpose

a generalist algorithm for cellular segmentation with human-in-the-loop capabilities
https://www.cellpose.org/
BSD 3-Clause "New" or "Revised" License
1.24k stars 359 forks source link

[FEATURE] - better documentation of and/or integration for use with training-run-tracking tools #927

Open bethac07 opened 2 months ago

bethac07 commented 2 months ago

Is your feature request related to a problem? Please describe. We love making Cellpose models, like, a lot, but especially on the trickier ones, it can be hard to remember what the magic secret sauce was for training.

Describe the solution you'd like In other projects in the lab, we're big users of Neptune.ai , and have colleagues who are more into WeightsAndBiases. It would be nice to get a sense of the least-worst way to use Cellpose with either one of these tools - unfortunately, Cellpose's awesome encapsulation of its training at the command line I think makes it harder to easily incorporate it that way, though I suspect it's easier when training in a notebook. We will almost certainly get into doing this ourselves, and can contribute back how easy (or not) it was if/when we do, but if it's something you already are aware of people supporting or doing, we're always happy to copy off of someone else's homework.

Thanks!

carsen-stringer commented 1 month ago

With Cellpose 3, we've put our training functions in a separate module, so these sorts of logging programs may be easier to deploy, but we are not using them and don't know anyone else using them. The losses are useful, but I'm guessing you probably want to run AP metrics when the training ends anyway, so this would come after calling the train_seg function I think.