Bug: manual labels of sorting analyzer are not reset in the Spikeinterface-GUI

[chiyu1203]

Dear Spikeinterface community,
I have been comparing the performance of several sorters and manually curating and labeling them on spikeinterface-gui .
My pipeline is created on jupyter notebook and involves saving the sorting analyser of different sorters into the disk and then load them one by one when doing the curation.
Here is the code.

sorter_list=["spykingcircus2","tridesclous2","mountainsort5","herdingspikes"]
this_sorter=sorter_list[3]
#oe_folder=r'Y:\GN25060\251130\coherence\session1\2025-11-30_14-25-01'
sorter_suffix = generate_sorter_suffix(this_sorter)
phy_folder_name = "phy" + sorter_suffix
analyser_folder_name = "analyser" + sorter_suffix + ".zarr"
sorting_folder_name = "sorting" + sorter_suffix
if analysis_methods.get("load_analyser_from_disc")==True:
    sorting_analyzer = si.load_sorting_analyzer(oe_folder / analyser_folder_name)
run_mainwindow(sorting_analyzer,curation=True)

Then I do the work on the GUI and click the buttons to save the labels (good, mua or noise) in the analyser and export them as a json file. While I switch to the next sorter, I close the GUI, which essentially finish executing that cell. Then I changed the index number of the sorter list to move on to the next sorter. With this pipeline, I noticed that labels made for previous sorters on the spikeinterface-gui would remain there (e.g. unit ID 10 is noise) for the next sorter. This is not a big issue but will leads to a bug when I did not remove those legacy labels but save/export them altogether with the new one. For example, Mountainsort5 identified only 29 units. However, because I first did manual curation on tridesclous2, which identified more than 80 units, and those labels seemed to carry over to the Mountainsort one. The resulting json file will have both information, attached file as an example.
manual_curation_1st.json

Is this a fixable bug? Should I move away from using this pipeline on jupyter notebook?
below is the software/hardware version.
spikeinterface 0.103.2
spikeinterface-gui 0.12.0
python 3.10.14
Windows 11

[samuelgarcia]

Hi.
This is indeed a strange behavior.
Are you sure of this line if analysis_methods.get("load_analyser_from_disc")==True:
That would explain that the analyzer is not reloaded and so you always inspect the same analyzer.
Otherwise the curation should be reset at each "new and fresh" created analyzer.

[chiyu1203]

Hi! I meant when I moved to a new sorter, I would manually change the index number of sorter list in the beginning.
this_sorter=sorter_list[3] #when using herdingspikes

this_sorter=sorter_list[2] #when using mountainsort5

And those analyser folders were saved seperately so I am sure that at every executeable cell, I only load one analyser folder.
With that pipeline, I only see one sorter's result at a time on the GUI. Only the labels was carried over to the next round.

[alejoe91]

@chiyu1203 does this happen as well if you restart the kernel? Where are you downloading the curation.json file?

[alejoe91]

This is the piece of code that loads a previous curation: https://github.com/SpikeInterface/spikeinterface-gui/blob/main/spikeinterface_gui/controller.py#L312-L362

As you can see, previous curation data is only loaded if you pass a curation_dict to the run_mainwindow function or if you saved the curation data to the analyzer folder.

Not sure if somewhow empty_curation_data is updated (but it shouldn't!): https://github.com/SpikeInterface/spikeinterface-gui/blob/main/spikeinterface_gui/controller.py#L360
We might want to force a deepcopy there.

[samuelgarcia]

Oh yes this is it we need a deepcopy of course becaue this is a dict of dict!!
I do a PR

[samuelgarcia]

Done in #225.
Thanks a lot @chiyu1203 for find this bug.

[chiyu1203]

@alejoe91 @samuelgarcia thanks for your reply! I have ran several tests and on another computer to reproduce the bug.
Firstly, restarting the kernel can prevent this bug. Also, I did not load labels from an json file. The abovementioned codes are the only thing I use.

In addition, this bug happens only if I click save analyser (so has nothing to do with the json file)
Correct: This bug happens as long as something is saved on the GUI (both export to json and savd analyser will affect the next one)

Interesting, I have saved labels in one analyser (Analyser A) and moved on to another analyser (Analyser B) to add some labels there. When I have done with Analyser B and moved back to Analyser A again, what is saved in the Analyser B does not affect the existing labels in the Analyser A. So I reckon this bug happens only when "labels" is first initialised in an analyser.

[samuelgarcia]

Did you test with this fix #225 ?

[chiyu1203]

Did you test with this fix #225 ?

not yet. I will give it a try later

[chiyu1203]

@samuelgarcia Hi! I can confirm that update fixed the bug. Thank you!!!