zero-shot-object-tracking
zero-shot-object-tracking copied to clipboard
roboflow-ai
Doesn't work with models trained with latest version of YoloV5
Models created/trained with latest version of YoloV5 don't load with this repo code. The YoloV5 code you have needs to be updated.
@rajeshpt that is right - we will provide an update for yolov5v6.0, and in the meantime provide a hash of the version we have pinned here cc @maxhs2014
@Jacobsolawetz thanks for your quick response. Just as a suggestion, would it make sense not to keep a copy of the yolo source code and instead just use the torch.hub.load to load the model. That way this repo would work irrespective of yolo version changes.
@rajeshpt interesting... can you use torch.hub.load with models that you have trained on your custom dataset?
@rajeshpt interesting... can you use torch.hub.load with models that you have trained on your custom dataset?
yes, you can. I am currently doing that and that is the recommended approach to avoid tight dependency with the yolo source and probably also from licensing perspective if one is developing a commercial product.
@rajeshpt I'm trying to implement your fix, but I'm struggling to get it to run. Would you mind please uploading an example?
Loading your custom model - model = torch.hub.load("ultralytics/yolov5", "custom", MODEL_PATH)
for Inference - results = model(frame)
You can look at this for more details - https://github.com/ultralytics/yolov5/issues/36
@JNaranjo-Alcazar it is still not supported, @maxhs2014 and I will take a look when we get a moment, and PRs always welcome :D
Jacobsolawetz is any repo for zero-shot-object tracking compatible with YOLOv5 v6.0 (the latest version)???????
There is now a YOLOv5 v7.0, I'm guessing it is still not up to date to work with this?
