When downloading model weights (Not using this account, if anyone from roboflow is wondering how I am downloading weights on Growth plan) the performance is much worse. I don’t have comparison pictures (Ran out of credits, which is why I am using kaggle for YOLO11 inference), but it is significantly worse. I attached one of the images I inferred using kaggle. Using Roboflow inference, It would detect all numbers besides maybe 1 or 2. Using kaggle, well you can see the difference. Does roboflow have something that makes it better at inference?
Hi there @EthanRhode - my gut is you have an error in your deployment pipeline somewhere.
We’ve spent a ton of time ironing out common errors in our inference server, which is likely why you’re seeing good results there.
1 Like
I just tested it using an RF-DETR model and the same thing happened with reduced performance on inference. I created a new notebook in kaggle and copy pasted the RF-DETR inference script and just changed the paths for the model and images. Was there something else I was supposed to do?
Sorry - want to make sure I’m following. You’re seeing better results for yolov11 on inference, but worse results for RF-DETR?
No, I am seeing worse results for both running inference on kaggle.
My advice is to just self-host our inference server – I’m not able to troubleshoot custom deployment pipelines.
Okay, Thank you.
I assumed you wouldn’t be able to but I figured I would try my luck.
This topic was automatically closed 21 days after the last reply. New replies are no longer allowed.