Recognition Cannot Get Past Image

[mwhouser]

Which version of recognize are you using?

10.0.6

Enabled Modes

Object recognition, Face recognition

TensorFlow mode

WASM mode

Downstream App

Photos App

Which Nextcloud version do you have installed?

32.0.1

Which Operating system do you have installed?

Ubuntu 24.04

Which database are you running Nextcloud on?

Postgres?

Which Docker container are you using to run Nextcloud? (if applicable)

AIO

How much RAM does your server have?

4 GB

What processor Architecture does your CPU have?

x86

Describe the Bug

Recognize errors on one particular image, and cannot move past it. Face and Object recognition halts.

From the Nextcloud Logs:

Allowed memory size of 536870912 bytes exhausted (tried to allocate 65536 bytes) at /var/www/html/custom_apps/recognize/lib/Classifiers/Classifier.php#389
Failed to generate preview of 18666 with dimension 1024 with nextcloud preview manager: No provider successfully handled the preview generation

The above just repeats every 6 hours. Same number 18666.

Expected Behavior

Fail x number of times, put it aside, and then move on.

To Reproduce

Not sure what to tell you. The image 18666 (if I'm identifying it correctly), is quite large: a 23 MB JPG. 15,600x15,600 pixels. It's not under "Photos", is a digital drawing, and doesn't have any faces or objects to recognize.

Debug log

No response

[marcelklehr]

Hey @mwhouser
Thank you for taking the time to give feedback. This is a problem with nextcloud's image preview service. You can probably get around it by putting an empty .nomedia file in the same folder as the problematic image. Let me know how it goes!

[mwhouser]

@marcelklehr That seems to have allowed it to move past the problem file.

Thanks for the suggestion.

[marcelklehr]

Woop woop. I'm glad it works now :)

[mwhouser]

@marcelklehr Strictly speaking, the issue was not fixed. Recognize should still move past the problem file. I just forced the matter manually. If I didn't know how to debug the issue with some URL massaging, it would be stuck there forever.

[marcelklehr]

It is stuck because an unexpected error occurs in the preview service of your nextcloud. It is not expected that the preview service cannot generate a preview for an image in my opinion.

[mwhouser]

I understand that.

I'm not familiar with the internals of the Recognize app, but when I go to the Admin settings for it, it shows me the queue length of items to be processed, and that's stayed at the same number for a week (since I created this issue). I assume that Recognize is managing the queue, and telling the preview service to generate the previews.

Recognize should be able to capture the fact that the preview service failed and stop re-sending the same item to the preview service over and over again.

If this is not how it's currently structured, please let me know. But something must be taking items from the queue and processing them. Whatever that is should be able to understand "hey, I've failed processing this item 20 times already, maybe I'll stop retrying."

[mwhouser]

To add to my comment above, it was the Recognize app that logged this error:

Failed to generate preview of 18666 with dimension 1024 with nextcloud preview manager: No provider successfully handled the preview generation

So the recognize app successfully understood that the preview failed. Why is it not within the Recognize app's capability to move past the failed item?

[marcelklehr]

Recognize actually falls back to gdlib for generating a preview if nextcloud's preview service fails and then falls back to using the image without resizing it. So there is logic in place to prevent this from happening. Sadly Allowed memory size of 536870912 bytes exhausted is a terminal error, so there is no way to gracefully recover from this. Counting the times a file has been processed would require large changes in the database architecture. We could implement a max file size setting that prevents large images from ever being processed and thus clogging up the queue. IMO as it is it's good enough as the admin does get an error in the log and in the recognize admin settings and in normal circumstances will investigate.