Updated Napari and now all my labels are in the wrong place!

[d45ha]

Is there an existing issue for this?

• I have searched the existing issues

Operating System

Windows 11 enterprise

DeepLabCut version

3.0.0rc13

What engine are you using?

pytorch

DeepLabCut mode

single animal

Device type

NVIDIA RTX 2000 Ada Generation

Bug description 🐛

Hi, I was having the same issue described here.

I am training a single-animal model which tracks the animal based on 5 body markers (nose, left hip, right hip, tail base, tail tip), plus 8 additional markers to identify corners of the arena. I found that my model was mostly predicting marker positions correctly, however with very low likelihoods (often >0.6), so I decided to extract outlier frames using the "uncertain" algorithm to try to improve certainty.

This was working fine - I had additional frames for my chosen videos with plotted predictions which I was able to adjust. However if the model didn't identify e.g. the tail tip, then I wasn't able to add this body marker back in, I could only adjust body markers the model had already predicted. Based on the advice from this post, I updated to the latest version of Napari. However, now all of my manual labels have moved to the wrong place and the predictions on the newly extracted outlier frames have completely disappeared. The GUI also is missing some of the features from before (e.g. choosing loop vs sequential plotting).

I might be missing something very obvious but please can I have some advice as I spent a lot of time manually labelling images for them to all have disappeared! :(

Steps To Reproduce

No response

Relevant log output

Anything else?

No response

Code of Conduct

• I agree to follow this project's Code of Conduct

[C-Achard]

Hello @d45ha,

Thanks for the report, and no worries, most likely your annotations are fine, it sounds like the plugin is not loading correctly, which prevents the remapping and controls from running correctly and makes the labels look incorrect in the viewer.

Just to clarify :

The GUI also is missing some of the features from before (e.g. choosing loop vs sequential plotting)

Do you mean that those controls simply cannot be opened any more (via Plugins > Keypoint controls), or are they always greyed out ?
Could you send the output of:

python -c "import sys, platform, importlib.metadata as m, napari_deeplabcut, pathlib; pkgs=['napari-deeplabcut','napari','QtPy','PySide6','PyQt6','shiboken6','numpy','pydantic']; print('## Environment'); print(f'- python: {sys.version.replace(chr(10), \" \")}'); print(f'- platform: {platform.platform()}'); [print(f'- {p}: {m.version(p)}') if True else None for p in pkgs if True]; print(f'- plugin_module_path: {pathlib.Path(napari_deeplabcut.__file__).resolve()}')"

to get your package versions ? This way I can know whether napari-DLC was also updated and the PySide6/shiboken6 deps are fine.

It'd also be great to have your config.yaml to have better context.

Best,
Cyril

[d45ha]

Hi @C-Achard,

Thank you so much for your quicky reply!

The controls are greyed out (screenshot attached).

Here is the output:
(DEEPLABCUT) C:\Windows\System32>python -c "import sys, platform, importlib.metadata as m, napari_deeplabcut, pathlib; pkgs=['napari-deeplabcut','napari','QtPy','PySide6','PyQt6','shiboken6','numpy','pydantic']; print('## Environment'); print(f'- python: {sys.version.replace(chr(10), " ")}'); print(f'- platform: {platform.platform()}'); [print(f'- {p}: {m.version(p)}') if True else None for p in pkgs if True]; print(f'- plugin_module_path: {pathlib.Path(napari_deeplabcut.file).resolve()}')"

Environment

• python: 3.12.7 | packaged by conda-forge | (main, Oct 4 2024, 15:47:54) [MSC v.1941 64 bit (AMD64)]
• platform: Windows-11-10.0.26200-SP0
• napari-deeplabcut: 0.2.1.8
• napari: 0.6.6
• QtPy: 2.4.3
• PySide6: 6.9.3
  Traceback (most recent call last):
  File "C:\Users\xw21269.conda\envs\DEEPLABCUT\Lib\importlib\metadata_init_.py", line 397, in from_name
  return next(cls.discover(name=name))
  ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  StopIteration

During handling of the above exception, another exception occurred:

Traceback (most recent call last):
File "", line 1, in
File "C:\Users\xw21269.conda\envs\DEEPLABCUT\Lib\importlib\metadata_init_.py", line 889, in version
return distribution(distribution_name).version
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "C:\Users\xw21269.conda\envs\DEEPLABCUT\Lib\importlib\metadata_init_.py", line 862, in distribution
return Distribution.from_name(distribution_name)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "C:\Users\xw21269.conda\envs\DEEPLABCUT\Lib\importlib\metadata_init_.py", line 399, in from_name
raise PackageNotFoundError(name)
importlib.metadata.PackageNotFoundError: No package metadata was found for PyQt6

Many thanks,
Dasha

[d45ha]

Config file attached here.

config.yaml

[C-Achard]

Very good, thanks!

One thing in your config.yaml that is likely not related but could be a problem: there are a lot of spaces in the paths, which our current parsing method has trouble with. If possible, I would recommend avoiding their use altogether, e.g. :

project_path: C:\Users\xw21269\Documents\Dasha - Deep Lab Cut - NP Pilot 
  3\NP_pilot3-Dasha-2026-02-06 - Copy

This has been improved in rc14, if you would like to update.

Otherwise, outputs look fine, one big issue with the current plugin version (#160) is that drag-and-dropping interactions only partially work, and to properly enable controls one has to use :

File > Open with plugin > Load folder...

AFTER opening the Keypoint controls for the layer to be correctly registered.

Note that this may not fix the broken ordering, as there could be a separate issue in remapping.
If that's the case, could you send me a sample broken labeled-data folder ? So I can try and reproduce.

---

Side note: Many of these oddities and load/save issues are being fixed in the plugin refactor.
If you do not mind trying an experimental build, please backup your data and install via:

pip install shiboken6>=6.4.2
pip install git+https://github.com/DeepLabCut/napari-deeplabcut.git@refs/pull/183/head 

If there is a problem, you can safely revert to the latest release with:
pip install napari-deeplabcut==0.2.1.8

Best,
Cyril

[d45ha]

Hi Cyril,

Thank you for this, I have tried the method you suggested and now the blank images are interleaved with labelled ones (I think model prediction labels as a few of them a slightly off). But the predictions do actually relate to the correct image now.

I tried to attach the labelled folder as a zip file but it isn't working, I'll respond separately via email.

Thank you for sending the experimental build, I would like to see if I can get it working in the current environment first and if that doesn't work try the experimental build - hope that's ok.

[d45ha]

Hi Cyril, Please find labelled folder: ​[https://res.public.onecdn.static.microsoft/assets/fluentui-resources/1.1.0/app-min/assets/item-types/24/zip.png]NP2_1 - neutral context - 24 March 2026 - 08-53-46 AM - Trim_trim 2.zip<https://uob-my.sharepoint.com/:u:/g/personal/xw21269_bristol_ac_uk/IQDY1phi91C_ToebHp3p7qciATeXukfJMFapuonPHlnL5sY>​ Many thanks, Dasha ​​ --------------------------------------------------- Dasha Anderson Research Associate Purple Lab E15 Biomedical Sciences Building University of Bristol

…

________________________________ From: Cyril Achard ***@***.***> Sent: 24 April 2026 10:59 To: DeepLabCut/DeepLabCut ***@***.***> Cc: Dasha Anderson ***@***.***>; Mention ***@***.***> Subject: Re: [DeepLabCut/DeepLabCut] Updated Napari and now all my labels are in the wrong place! (Issue DeepLabCut/napari-deeplabcut#189) [https://avatars.githubusercontent.com/u/94955160?s=20&v=4]C-Achard left a comment (DeepLabCut/napari-deeplabcut#189)<#189 (comment)> Very good, thanks! One thing in your config.yaml that is likely not related but could be a problem: there are a lot of spaces in the paths, which our current parsing method has trouble with. If possible, I would recommend avoiding their use altogether, e.g. : project_path: C:\Users\xw21269\Documents\Dasha - Deep Lab Cut - NP Pilot 3\NP_pilot3-Dasha-2026-02-06 - Copy This has been improved in rc14, if you would like to update. Otherwise, outputs look fine, one big issue with the current plugin version (DeepLabCut/napari-deeplabcut#160<#160>) is that drag-and-dropping interactions only partially work, and to properly enable controls one has to use : File > Open with plugin > Load folder... AFTER opening the Keypoint controls for the layer to be correctly registered. image.png (view on web)<https://github.com/user-attachments/assets/796a5946-732e-447e-9f0f-7db49353d7f1> Note that this may not fix the broken ordering, as there could be a separate issue in remapping. If that's the case, could you send me a sample broken labeled-data folder ? So I can try and reproduce.

________________________________ Side note: Many of these oddities and load/save issues are being fixed in the plugin refactor. If you do not mind trying an experimental build, please backup your data and install via: pip install shiboken6>=6.4.2 pip install ***@***.***/pull/183/head If there is a problem, you can safely revert to the latest release with: pip install napari-deeplabcut==0.2.1.8 Best, Cyril — Reply to this email directly, view it on GitHub<#189 (comment)>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/AWWTLRIXX3LTKXN5CDWVVB34XM3J3AVCNFSM6AAAAACYE5HKL2VHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHM2DGMJSGMYTQOJUHE>. Triage notifications, keep track of coding agent tasks and review pull requests on the go with GitHub Mobile for iOS<https://github.com/notifications/mobile/ios/AWWTLRPRHLAIGR3V5YTJUZ34XM3J3A5CNFSNUABFM5UWIORPF5TWS5BNNB2WEL2JONZXKZKDN5WW2ZLOOQXTIMZRGIZTCOBZGQ42M4TFMFZW63VHNVSW45DJN5XKKZLWMVXHJKTGN5XXIZLSL5UW64Y> and Android<https://github.com/notifications/mobile/android/AWWTLRNGLFUKSAAZO6SFTJL4XM3J3A5CNFSNUABFM5UWIORPF5TWS5BNNB2WEL2JONZXKZKDN5WW2ZLOOQXTIMZRGIZTCOBZGQ42M4TFMFZW63VHNVSW45DJN5XKKZLWMVXHJLTGN5XXIZLSL5QW4ZDSN5UWI>. Download it today! You are receiving this because you were mentioned.Message ID: ***@***.***>

[C-Achard]

Thanks for your reply @d45ha , I'm afraid I cannot access the SharePoint link (it looks like a permissions issue), but if the problem is now mostly resolved, there’s no need to send the project.

Just to clarify, it seems the labels do relate to the correct frames now, but I'm not sure if the "interleaving" is a problem ?
Are you getting:

• Labels correctly overlaid on top of each image/frame
• Alternating panels/images, e.g. one image frame and then one labels-only frame?

If you are seeing them side-by-side it might just be grid mode, which you can disable with the lower left corner button:

Normal	Grid mode

If there is still a problem, please let me know!

Best,
Cyril

[C-Achard]

(P.S. And yes no worries about the experimental build! Definitely not a requirement, happy to help with the current release)

[d45ha]

Hi Cyril, Not sure why the folder permissions aren't working, maybe this link will allow access: ​[https://res.public.onecdn.static.microsoft/assets/fluentui-resources/1.1.0/app-min/assets/item-types/24/zip.png]NP2_1 - neutral context - 24 March 2026 - 08-53-46 AM - Trim_trim 3.zip<https://uob-my.sharepoint.com/:u:/g/personal/xw21269_bristol_ac_uk/IQDcxLyyAOqQRqsv7K0FCWFqAT8cSMZEbUYfUJjPOPki1jw>​ In terms of the interleaving images, sorry I probably wasn't very clear. When I originally extracted frames for labelling (before training the model or analysing any videos), I extracted 20 frames per video. After analysing videos and getting likelihoods, I extracted outlier frames using the uncertain algorithm. This gave me an additional 20 frames per video (now 490 total). Before updating Napari, when I loaded a folder with outlier frames it had my manually labelled frames as the first 20 and the outlier frames as the second 20. Now they appear to be all mixed up. Now when I open a folder with extracted frames, half (20/40) are labelled (and given the small inaccuracies I would assume these are the extracted outlier frames with model predictions, not my own predictions), and the other half are not labelled. Out of interest, if frames have been extracted but not labelled, could this affect model accuracy? When I first extracted frames I extracted 20 frames from all my videos but only manually labelled a subset. I assumed unlabelled images would be ignored if they have no markers but now I am doubting myself... Best wishes, Dasha --------------------------------------------------- Dasha Anderson Research Associate Purple Lab E15 Biomedical Sciences Building University of Bristol

…

________________________________ From: Cyril Achard ***@***.***> Sent: 24 April 2026 12:41 To: DeepLabCut/DeepLabCut ***@***.***> Cc: Dasha Anderson ***@***.***>; Mention ***@***.***> Subject: Re: [DeepLabCut/DeepLabCut] Updated Napari and now all my labels are in the wrong place! (Issue DeepLabCut/napari-deeplabcut#189) [https://avatars.githubusercontent.com/u/94955160?s=20&v=4]C-Achard left a comment (DeepLabCut/napari-deeplabcut#189)<#189 (comment)> (P.S. And yes no worries about the experimental build! Definitely not a requirement, happy to help with the current release) — Reply to this email directly, view it on GitHub<#189 (comment)>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/AWWTLRLFUUE2UZFZKKRZMWT4XNHGJAVCNFSM6AAAAACYE5HKL2VHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHM2DGMJSHA2TQNRVHA>. Triage notifications, keep track of coding agent tasks and review pull requests on the go with GitHub Mobile for iOS<https://github.com/notifications/mobile/ios/AWWTLRMJW6MARN2JFFTDLFL4XNHGJA5CNFSNUABFM5UWIORPF5TWS5BNNB2WEL2JONZXKZKDN5WW2ZLOOQXTIMZRGI4DKOBWGU4KM4TFMFZW63VHNVSW45DJN5XKKZLWMVXHJKTGN5XXIZLSL5UW64Y> and Android<https://github.com/notifications/mobile/android/AWWTLRNN2HL6WS7WLMYJ2Q34XNHGJA5CNFSNUABFM5UWIORPF5TWS5BNNB2WEL2JONZXKZKDN5WW2ZLOOQXTIMZRGI4DKOBWGU4KM4TFMFZW63VHNVSW45DJN5XKKZLWMVXHJLTGN5XXIZLSL5QW4ZDSN5UWI>. Download it today! You are receiving this because you were mentioned.Message ID: ***@***.***>

[C-Achard]

@d45ha I'm sorry but both in email and browser the link you sent looks broken:
[https://res.public.onecdn.static.microsoft/assets/fluentui-resources/1.1.0/app-min/assets/item-types/24/zip.png]NP2_1 - neutral context - 24 March 2026 - 08-53-46 AM - Trim_trim 3.zip<[https://uob-my.sharepoint.com/:u:/g/personal/xw21269_bristol_ac_uk/IQDcxLyyAOqQRqsv7K0FCWFqAT8cSMZEbUYfUJjPOPki1jw>​](https://uob-my.sharepoint.com/:u:/g/personal/xw21269_bristol_ac_uk/IQDcxLyyAOqQRqsv7K0FCWFqAT8cSMZEbUYfUJjPOPki1jw%3E%E2%80%8B)

Clicking the actual sharepoint link only shows:

Sorry, something went wrong
An unexpected error has occurred.
[Technical Details](https://uob-my.sharepoint.com/:u:/g/personal/xw21269_bristol_ac_uk/IQDcxLyyAOqQRqsv7K0FCWFqAT8cSMZEbUYfUJjPOPki1jw%3E%E2%80%8B#)

I do apologize but since it sounds like frames are still getting mixed up, the easiest would be for me to be able to inspect a minimal set of problematic annotations. If the issue is the zip size, no need to include everything, a 40 frames example as you mentioned is enough. That should upload fine on GitHub.

Maybe the problem comes from the CollectedData layer vs machinelabels layer. The plugin will load each h5 as a separate Points layer, which you can show/hide individually, but it sounds like the outlier frames are being added on top and may not be correctly aligned.

Regarding unlabeled frames, we do support occlusions (e.g. a specific keypoint being hidden in a given frame), but I can check in the h5 itself, this is the GT source during training so empty frames should not appear.
I know we also explicitly encode not visible vs not annotated, so overall it should be fine.

Hope this helps, let's try and find a way to share the data so I can more easily determine the exact issue. Perhaps a drive (Google or OneDrive) link would be better ? Or a smaller zip directly as GitHub attachment ?

[d45ha]

@C-Achard thank you again - and apologies for the link issue. I tried uploading here as a zip but it wasn't working either.

Yes I think you're right. I just realised the folder of frames I was trying to open was one that didn't have manual labels, in Napari only the machinelabels layer was showing up (which makes sense). However, I tried the folder for one of the videos which I did manually label and now in Napari only the CollectedData layer is coming up (across all frames). I've double checked both the CollectedData file and machine labels file are definitely in the folder.

Here is a Google drive for a folder with no CollectedData: https://drive.google.com/drive/folders/1wK5D21y1Suek_EsU_JqLmK6UgDAMTAvg?usp=sharing

And with CollectedData: https://drive.google.com/drive/folders/1HRMUtir7XojEvnJtxkXEKraBy5xBT3eu?usp=sharing

Hopefully you can access these ok.

Great, if you could confirm r.e. unlabelled frames, that would be great.