[Support]: Issue with state classification

[kanemari]

Describe the problem you are having

I have created a state classification for a simple gate - 3 states - open, closed, unknown. The unknown may not be needed but I used it to classify images with artefacts due to camera pixellation or ffmpeg drop-outs.

I trained 2-3 images of the gate open, around 10 of the gate closed, 1 unknown. It was working fine for the first 2 days identifying state changes and allowing me to go back into the UI and classify any new pictures that have lower state certainty.

However, now the UI seems to time out and doesnt identify any of the previous states, nor allow new images to be assigned to an existing state for training. It still technically works as far as identifying the state, I just cant update anything in the UI anymore. There arent any errors that I can see in the logs, new classification images keep coming in, I just cant do anything with them.

Version

0.17.0-f0d69f7

What browser(s) are you using?

Chrome 145.0.7632.117

Frigate config file

only including the relevant part

classification:
  bird:
    enabled: false
  custom:
    Gate:
      enabled: true
      name: Gate
      threshold: 0.8
      state_config:
        cameras:
          driveway:
            crop:
              - 0.4636023186436883
              - 0
              - 0.6992622349446398
              - 0.41895096231280265
        motion: true
        interval: 60

Relevant Frigate log output

none

Relevant go2rtc log output

none

FFprobe output from your camera

irrelevant

Frigate stats

irrelevant

Install method

Docker Compose

docker-compose file or Docker CLI command

irrelevant

Object Detector

Coral

Network connection

Wired

Camera make and model

hikvision, tapo, ring

Screenshots of the Frigate UI's System metrics pages

Any other information that may be helpful

No response

[inkeep-ai-assistant[bot]]

Your state classification model appears to be configured correctly, and the fact that it's still technically identifying states suggests the model itself is functioning.(1)

For state classification models, the Recent Classifications tab has specific behavior regarding which images are saved. Images are only added under these circumstances:(1)

• The first classification attempt for a camera is always saved(1)
• Images are saved when the detected state differs from the current verified state(1)
• Images are saved when there's a pending state change being verified (requires 3 consecutive identical states)(1)
• Images with scores below 100% are saved even if the state matches the current state(1)

Images are not saved when the state is stable (detected state matches current state) and the score is 100%.(1)

The UI timeout behavior you're experiencing isn't directly addressed in the available documentation. Since the classification is still working but the UI interaction is failing, this could be a browser-specific issue or a session timeout problem.

Try these steps:

1. Clear your browser cache and cookies for the Frigate UI
2. Try accessing the Classification page in an incognito/private browser window
3. Check if there are any browser console errors (F12 developer tools)
4. Verify your authentication session hasn't expired

Your configuration with save_attempts set to the default of 100 for state classification models should retain classification attempts in the Recent Classifications tab.(1)(2)

If the issue persists after trying these steps, visit GitHub to report this UI-specific behavior.

📚 Sources:

1. State Classification
2. Full Reference Config

---

Have another question? Just tag @inkeep.

[kanemari]

in the UI, looking at the requests its sending out, it seems to keep repeating the same request.

Request URL
http://192.168.xx.xx:5000/4FFA347B-FAC6-48A6-9FFB-F8FF95F68291/4F88AC04-3AB1-474D-B2F5-A41AA7D6D648/from?get&nocache=1090d
Request Method
GET
Status Code
200 OK
Remote Address
192.168.xx.xx:5000
Referrer Policy
strict-origin-when-cross-origin

the thumbnail of the state which previously displayed a screenshot of the gate open now just shows the typical grey box fading in and out - which normally indicates an image is loading but it never does in this case - and keeps repeating the above request

[hawkeye217]

That request URL is not a pattern native to Frigate. Are you running behind a proxy or do you have a browser extension installed that could be intercepting requests?

[kanemari]

no - PC is on the same subnet as Frigate - connecting directly to Frigate on unauth port 5000.

GET /4FFA347B-FAC6-48A6-9FFB-F8FF95F68291/2B6BAFD2-CB93-42AB-B44C-3C68CFC3746F/from?get&nocache=17d64 HTTP/1.1
Accept: */*
Accept-Encoding: gzip, deflate
Accept-Language: en-AU,en;q=0.9,ko-KR;q=0.8,ko;q=0.7,en-GB;q=0.6,en-US;q=0.5
Connection: keep-alive
Host: 192.168.xx.xx:5000
Referer: http://192.168.xx.xx:5000/classification
User-Agent: Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/145.0.0.0 Safari/537.36
X-Requested-With: XMLHttpRequest

[kanemari]

actually it may be irrelevant - the constant polling - it does the same on other pages that dont have issues.

[kanemari]

this can be closed. it was caused by image files that were already in the dataset file structure - or supposed to be in the dataset file structure - missing from /clips/state name/state directory. I had this situation arise because of a failed network share mount, and once I resynched the directories the problem went away.

It would be good if some sort of logic to detect / handle missing files could be put in place, as I was unable to do any new image training while those files were missing. the logs didnt have any errors about the files missing either. but state identification still worked - assumingly because the trained dataset file lives somewhere else? - which was confusing.