EPM_syllables notebook

[sfmig]

A notebook demoing some simple analysis using a trained keypoint moseq model on new data.

It should be compatible with PR #16, but I rebased to main rather than to that branch for clarity (to have just one file on this PR).

Some questions:

• I cleared the outputs in the last commit, but let me know if we prefer to have them displayed and I can revert it.
• What was the target syllable duration during training? I wrote 20 frames but not 100% sure

[niksirbi]

• What was the target syllable duration during training? I wrote 20 frames but not 100% sure

It was actually 12 frames (400 ms), and I achieved that for the AR-HMM fit. During the full model fit, while is also started at ~12 frames, it drifted upwards over time and finally settled at ~16 frames. I decided to leave it at that instead of re-training.

I will rephrase this:

We can compute the median duration of all syllables, and check if it is in line with the target of 20 frames.

to:

We can compute the median duration of all syllables in the new data.

Keep in mind that the median syllable duration in the training data was ~16 frames

[niksirbi]

Nice touch with the itertools.groupby. I'd never used it before and it's very useful.

[niksirbi]

I cleared the outputs in the last commit, but let me know if we prefer to have them displayed and I can revert it.

I prefer to have them displayed, to match the other existing notebooks.

It's also nice for posterity, if someone happens to come across these notebooks on github after the course (or completely outside the course context).

I've generated the outputs on my end and pushed them.

[niksirbi]

Excellent work @sfmig!

This notebook is a great resource beyond just this course.

I especially appreciate the fact that you went at length to match the formatting I had in the EPM_analysis.ipynb notebook. This way there was no polishing / harmonisation left for me to do.

The only finishing touch I made was to show the sample frame image as a background for the centroid trajectory and heatmap plots (as we do in EPM_analysis.ipynb as well). Apart from the visual context it adds, it also eliminates the need for manually inverting the y axis and setting an equal aspect (because matpltolib automatically applies these settings when you show an image first).

This PR is ready to go as is, but I have one lingering question out of curiosity: why did you need to use 51 colors in the ethogram if the number of syllables is much less that that?

[sfmig]

why did you need to use 51 colors in the ethogram if the number of syllables is much less that that?

Yeah good catch! I suspect it is an artefact of my explorations 😅

I reduced it to 30 and added an explanation. Even though usually ~15 are found in our example, I left it as 30 just in case people want to have a go with a different video.