APT to JAABA Issue

[Hoopfer Lab]

Hello,

We created a multipoint tool through APT with 23 key points for fly grooming and are trying to integrate this into JAABA to create behavioral classifiers (see image 1). We have tracked single flies with FlyTracker-1.0.5; the perframe features from FlyTracker are attached (see image 2).

 

We were wondering if there is a matching config for these features? We compared the features we have from FlyTracker to the Eyrun_flies config. Many features we have are log transformed or normalized versions of several features from the eyrun config. We also have dist_to_wall features while the eyrun config does not. I am not sure if this is causing the problem with the errors shown below.

There is a persisting error that we run into when trying to load the experiment directory and reconstitute the files in JAABA. Could this be an issue with what the trk file should be named? Even if we regenerate the files, the perframe features remain missing. What should the experimental file directory hierarchy look like?

Thank you for your help,

David and Yeseo from the Hoopfer Lab

[Mayank Kabra]

Hi,

The reason for error is that the trk file as specified in "Trk File Name" (EH250924_HEAT_1_CH6_GroomTrx2_main_V2.trk) in APTProject doesn't exist in your experiment directory D:\Hoopferlab\APT_tracking\movies_to_track\EH250924_HEAT_1_CH6\EH250924_HEAT_1_CH6_JAABA directory. Similar to movies for JAABA, the trk files should have the same name in each experiment directory. Have a consistent naming trk file convention and use that in APTProject. For example if you name the trk file as APT_output.trk then specify that in the APTProject GUI.

As for using the feature set you get from FlyTracker, I'll ask Alice to respond.

HTH,

Mayank

--
You received this message because you are subscribed to the Google Groups "JAABA" group.
To unsubscribe from this group and stop receiving emails from it, send an email to jaaba+un...@googlegroups.com.
To view this discussion visit https://groups.google.com/d/msgid/jaaba/4e854bc1-4cdd-4b60-b96a-4d26e1ebe912n%40googlegroups.com.

[Alice Robie]

Hi

I did a deep-dive on the eyrun target type a couple months ago because I also ran into trouble with the current out of flytracker and Eyrun_flies not working. 

What I found was that the transforms of the Flytracker outputs were added later and Eyrun has posted an updated config.xml file for the Eyrun_flies that corresponded to the new outputs. https://kristinbranson.github.io/FlyTracker/FAQ.html (link is broken but I tracked down a copy).I've now uploaded this to the JAABA repo as eyrun_flies2, and this seemed to work for me. I didn't want to break the original eyrun_flies for folks that might be still using an old version of flytracker. 

Note the further complication of pairs vs groups in the FAQ that I didn't deal with. 

However, you may have noted that there are far fewer perframe features for this target type than for flies or flies_wing which are used with the output of Ctrax. Our Fly Disco Pipeline has code for converting the movie_JAABA/trx.mat into a form that can use flies_disco which has the ctrax set of features if you're interested in that. 

Since you have APT keypoints you may not need as many body and social features from body tracking, but it is something to keep in mind if you're having trouble training a classifier. You can see the classifier performance quantification I did comparing body, keypoints, and both in Fig 8 of the Disco manuscript. 

Cheers

Alice 

To view this discussion visit https://groups.google.com/d/msgid/jaaba/CAGgnnKwHmnCQ2TeF-7-kOmkxwim25mjcbY4T9qm-M%2BL7NVKnVw%40mail.gmail.com.