how do I use the coloured training icons within the classification page?

[ElectronicBattle]

Hello all

0.17 introduced classifications but I don't know what to do with the different colours of clickable icons once a load of candidate cropped images have populated.

Once I had started by "seeding" the state or object with a few example images, I had a green icon which when clicked gave a spinning icon before returning to green once again. I assumed green meant things were going well.

Then those initial images were added to by Frigate, in their dozens or hundreds, and I think the icon went blue. I could then click on it, so I did, and the spinning icon occurred until the icon finally went red. Clicking on it when red again causes the same process, which returns to red in the end. Red doesn't appear to be an indication of success, so I may have messed things up.

Examples:

Since yesterday Frigate has automatically built up 50 to 150 images per object, and 50+ images for each of the two states of a state classification.

[NickM-27]

Red means it failed to train, blue means it is ready to be trained (meaning you have added at least one image)

[ElectronicBattle]

My mistake; thanks. I see the idea now.

Should I correct the configuration and retrain (press the blue icon) or will I need to delete the images and start again? If so, it wouldn't take long to build up a new library.

[NickM-27]

it is not clear to me at this point what is not working as expected.

[ElectronicBattle]

It isn't clear to me whether it is working properly or not due to that red icon result which I mentioned earlier. I imagined that the blue training button might turn green, or would be disabled, or would go grey, when successful but I haven't yet found a description in the documentation which explains it.

what I am doing is waiting for images to arrive, and then clicking through each of them to check that the detected image matches with the assigned sub_label name (for the object which is dog in this case) out of the two I have defined; or selecting "none" if the image isn't relevant. Here are some more images which arrived just now:

I see the count in the training button increment each time I manually process an image as per the above.

Then, once they have all been processed manually, I don't know whether I am now supposed to submit them by pressing the aforementioned blue "train model" button or not.

I'm doing the same for the blue gate (and telling the training whether the gate is open or closed) and the same for cats (and selecting one of two sub_label names for them).

Here's the associated yaml:

classification:
  bird:
    enabled: true

  custom:
    blue gate:
      enabled: true
      name: blue gate
      threshold: 0.8
      state_config:
        cameras:
          drive:
            crop:
              - 0.24339351982506427
              - 0.24351424028441887
              - 0.34788469604283856
              - 0.42927633133823984
          blue_gate:
            crop:
              - 0.5241821621727178
              - 0.31128359900885627
              - 0.7417993263426347
              - 0.6981585575331527
        motion: true
    pet dogs:
      enabled: true
      name: pet dogs
      threshold: 0.8
      object_config:
        objects:
          - dog
        classification_type: sub_label
    Pet Cats:
      enabled: true
      name: Pet Cats
      threshold: 0.8
      object_config:
        objects:
          - cat
        classification_type: sub_label

edit: I ran the training process for cats (79 images added, so not too many if I have to re-do them in some sense) and initially I got a green message in the UI saying that the training had started OK

but then after 60s I got a red failure again as reported earlier

with the only associated logs showing as

[NickM-27]

yes, you train the new model and it will turn into a loading button, then it will be disabled with (0) in the button

[ElectronicBattle]

Something isn't working at all, in that case.
I see this button:

I press it and it changes to this button:

and this notification appears in the middle of that page:

which then gets replaced with this one, after about a minute or so:

and the logs it advises to check simply say:

with the preceding logs saying things like:

at which point the training icon remains red as above and clicking on it again simply repeats the process from the dimmed-blue "in progress" icon, unless I restart Frigate at which point I get the first (bright blue) icon.

[ElectronicBattle]

yes, you train the new model and it will turn into a loading button, then it will be disabled with (0) in the button

Thanks - I will need to raise a ticket at a later date because, having upgraded to 0.17.1 and rebooted and so on, this still isn't working. I have REMd out the yaml for now and I'll try again when time permits.