Question: How to start RE-Train for new images in training-Folder?

[ozett]

i am filling up the train-images-folder with new images. How do i start re-training of all folder? or of a certain single folder, after coping images into it?

If i switch the dropdown to a certain "person/folder", the sync-button is still inactive. do i have to restart the whole containter for that?

i pulled the latest beta, but it still shows 10x. ...instead of 10.2...?

[jakowenko]

Does this issue still persist with the latest / beta build?

[ozett]

i will check with 1.0. But i have to pull it first...hold on. i come back

[ozett]

maybe related? i have images with training-timeout. but cannot move& train this single image. a Gui-button would be helpful.

if i use the dropdown to select a target-person other than the untrained image, this new selection disables the image and train therefore not possible.

v1.0.0-39cb3c3:latest

[jakowenko]

Select the image, click the trash button in the top right, then you should be able to retrain it. Clicking "sync" would also try to retrain it for that preson.

[ozett]

for clarification: i am on the the train-pages. i did like you suggested: i clicked the trash-button and the image was trashed. expectable. no more re-train. image was deleted. not really a solution... 😢

[jakowenko]

Ah I guess because it was never trained with a 200 response I have logic in there to delete it.

You could've also clicked the "sync" button which would've tried to train that image again for the user.

[ozett]

is ist possible to retrain all folder at once? ( i switched compreface to custom-image with mobilenet and i think i have to relearn all faces) now i search for this button .. 😄

[jakowenko]

Right now you can retrain all the images for a single person at a time. When you select the name from the dropdown click the "sync" button to resync/retrain all the images for that person.

[ozett]

thanks. i changed the compreface modell and wanted to compare the detection precision between the default model (facenet - i guess) and the heaviest modell (insightface, looks promising. github says it could run as tensorflow..maybe of interest for your coral experiments..) - on all my trained images.

is there an easier way to retrain all images on all faces - at once?

[jakowenko]

Currently you have to retain each person at a time. I don't see that as much work though unless you have dozens of people. If you want to put in a FR to do everything at once, I can try to get to it at some point. It probably won't be a high priority item though as there's some other things I want to work on next.

[jakowenko]

I will checkout that other model and project you talked about. Seems promising!

[ozett]

i was wondering if doubl-take could use the normal compreface API for face-rec, but with different compreface-models?

do you tested this? does compreface aggregate face-recognitions form different models== "embedding" servers":3000 like they tell and show on architeture page. https://github.com/exadel-inc/CompreFace/blob/master/docs/Architecture-and-scalability.md

would make a great distributed system...

(or i have to use two instances of compreface with different models - as different detektors for doubltake).. 💇