search

tyiannak / amvoc

A Python Tool for Analysis of Mouse Vocal Communication
MIT License
13 stars 2 forks source link

Improve Unsupervised Vocalization (syllable) detection. Compare to existing outputs (Cezar to send data) #1

Closed tyiannak closed 4 years ago

tyiannak commented 4 years ago

The goal of this issue is to create a robust and more accurate syllable detector (unsupervised). Towards this end, we will experiment with feature extraction and thresholding parameters. Results will be compared wrt other similar software and/or human generated ground truth.

tyiannak commented 4 years ago

@cvargas4 what is the last column of the csv file? (e.g. number 13 in the following row) TGExampleWithTimeStamps,B148_test,685,296.0448,296.2272,13

cvargas4 commented 4 years ago

@cvargas4 what is the last column of the csv file? (e.g. number 13 in the following row) TGExampleWithTimeStamps,B148_test,685,296.0448,296.2272,13

The last column is listing which "Repertoire Unit (RU)" that USV belongs to. In this case, I asked MUPET to cluster into 40 RUs. That syllable in your example happens to be in the 13th cluster. I believe this is done by k-means.

So far we (the lab) have not used these RUs as they are not necessarily consistant across analyses (very much within batch/recording).

tyiannak commented 4 years ago

as we discussed today @cvargas4 , these are the subtasks to be completed before finalizing this issue. 1) change F1/F2 to be static (code) 2) add the spectral energy + dynamic threshold + energy + ratio (see next step) as a second plot in the evaluation.py script (code) 3) add some energy/spectral_energy ratio in the thresholding rule (it seems that true vocalizations have ratios closer to 1) (code) 4) more parameters tuned in evaluation.py (code) 5) re run for the whole recordings (report) 6) prepare report + examples (QA)

tyiannak commented 4 years ago

I've updated the results and comparison to mupet here (for the best parameters). @cvargas4 it would be great if you could randomly select some areas from the 2 spectrogram plots to estimate precision and recall of the amvoc detector.

You can also commend the respective document here.

I am also working on adding the MSA comparison and after that I think we're done with the vocalization detection task.

tyiannak commented 4 years ago

@cvargas4 document and results updated. The following is missing to close this issue: 1) @cvargas4 to proceed with Precision/Recall evaluation (see document for the link to the 2 html files to be used for the evaluation) 2) @tyiannak and @cvargas4 to check the low agreement results with MSA

tyiannak commented 4 years ago

Can be considered closed after precision/recall has been added