Training custom model, will deepstack work with models made using YoloX


I’ve been collecting my own images to train a custom model for my needs and was going to do everything as explained in the deepstack docs using YoloV5 but just recently found that there’s a YoloX which is supposedly a bit better than v5. My question is will deepstack work with a YoloX model or does it have to be V5?


if you are referring to ‘yolo v5x’ (a high accuracy model variant of v5), it will definitely work. I am not familiar with a yoloX, or whether other yolo versions (3, 4, X) work with deepstack, maybe someone else knows?

edit- thanks for the info, looks interesting. maybe just try an ultra simple model with minimal labeling/epochs using yoloX and see if it works? (and let us know!)

Did you test to train a model with yolox and run it in deepstack?

I have not yet trained a model. I’m still collecting images from my security cameras. Have about 2500 so far for people. Im sure that’s plenty but I’m waiting until I have 3500. 3000 for training and 500 for testing. First time training so not 100% sure on a good train to test ratio but that’s what I’m aiming for right now.

My assumption is that deepstack only works with YoloV5 for now as I saw another post where someone was asking if yoloV3 models would work and one of the devs said deepstack only runs V5 models.


Btw, i could recommend to train anyway. I keep re-training my model with pictures from my security cameras, and i use the old training model to automatically annotate the new pictures i have collected. That way the manual task is to only verify the results in labelIm and fix errors. For me 80% are good to go instantly, and 20% needs adding a missed object, removing a false positive or changing a wrong classification. My hand got real tired of drawing the boxes in 1000+ images.

Of course there are multiple ways to automate the process, but doing it manually is not fun! :slight_smile: