Clarification about training requirements/optimisation

[stowerslab]

Hi all, we are very excited to try the new keypoint moseq, but just wanted to ask some questions and maybe offer some suggestions. 1) I could not find any information on training requirements/recommendations and had to dig through slack to find that the recommended training requirements for keypoint moseq was about 20 hours of video. Does the recommended video duration change if the keypoints are higher in likelihood/confidence (say if we have brightly lit headfixed animals?) 2) Do you all have a guesstimate of what would be a good overall confidence/likehood range that keypoint-moseq could work with and when we might want to say invest more time in retraining pose estimation models or drop a bodypoint to be included while modelling? 3) Do you all have any data/idea about how deviations from that total video duration requirement might affect modelling success? (For instance, if all we can collect is 5 hours of video, would that still be sufficient if not ideal?) 4) A problem with the previous version depth-moseq is that size of the animals (say males vs females or older mice vs younger mice) was a problem, is that still an issue with keypoint-moseq (in case we wanted to combine data from both males and females)? 5) Should we blank and interpolate low confidence points for the ones we need included? 6) The noise calibration is a bit confusing to use, what does it mean by : "Continue until the regression line stabilizes.". does that mean till the slope of the line stops shifting? becomes horizontal? all the points lie close to the line? 7) Other than having prior assumptions about syllable durations, how else would you recommend optimising on kappa (i know in the previous version depth moseq, there was a way to compare syllable durations to changepoint durations to optimise kappa, but that seems to have been removed, is that method no longer valid?) 8) is the recommended number of iterations for training either the initial AR-HMM or full model the same as in the tutorial or different? (i see that the bioxirv paper uses 500 iterations for the full model 9) maybe a FAQ on the website with some of these questions and answers might be helpful, sorry if i missed something somewhere!

[calebweinreb]

Hello,

These are great questions! I followed your advice and made an FAQ page. Let me know if it addresses all your questions.

[stowerslab]

Hey caleb it absolutely does! Thanks

• Shawn from Stowerslab