Encountered problem while using YOLOv8

[poor-soul]

I was trying to run YOLOv8 using visual code studio. Installed ultralytics and ran the yolo predict model=yolov8n.pt source='https://ultralytics.com/images/bus.jpg' command on the vs code terminal. However the output i received was 2 persons, 1 bicycle, 5 cars, 10 motorcycles, 73 boats, 3 stop signs, 1 dog, 10 horses, 10 cows, 32 bears, 1 giraffe, 63 umbrellas, 6 handbags, 9 frisbees, 15 snowboards, 5 surfboards, 12 knifes, 5 beds, 37 dining tables which are clearly not part of the picture. I will be providing a attachment of the image.

When i first installed ultralytics and tried running torch there was a missing dependency error. fbgemm.ddl was missing. Later when i installed vs_BuildTools this issue was resolved. Then i proceeded to run the code in a virtual environment where a program using torch ran without any errors. Then i proceeded to type this code snippet and encountered this problem. I also tried running using command prompt and jupyter notebook but the same issue persists. I also did check if the versions are compatible, which they are. I haven't installed cuda yet is it becuase of that or are there any other issues which i am not aware of ? Somebody please help.

[glenn-jocher]

It sounds like you're encountering an unusual issue with YOLOv8 predictions. First, ensure you're using the latest version of the Ultralytics package by running pip install --upgrade ultralytics. If the problem persists, it might be related to corrupted model weights or an environment-specific issue. Try re-downloading the model weights with yolo predict model=yolov8n.pt source='https://ultralytics.com/images/bus.jpg'. If the issue continues, consider running the same command in a clean virtual environment to rule out any conflicts. CUDA is not required for CPU inference, so it's unlikely to be the cause. If you still face problems, please provide additional details or error messages.

[poor-soul]

I did all the things you mentioned but the problem still remains unsolved.

1. First updated ultralytics. Then i ran the command but it was still not working.

2. Then i proceeded to run on a new environment but the result is the same.

I did mention before about fbgemm.ddl file missing and installing vs_BuildTools resolved that issue. Is the problem caused by something related to this ?

[glenn-jocher]

@poor-soul thank you for providing the additional details. Given that you've already updated Ultralytics and tried running the command in a new environment, let's explore a few more steps. First, ensure that your model weights are not corrupted by re-downloading them using the yolo predict command. If the issue persists, it might be related to the initial fbgemm.dll error. Although you resolved it by installing vs_BuildTools, there could still be underlying issues with your PyTorch installation. Try reinstalling PyTorch and its dependencies. If the problem continues, please share any error messages or logs you encounter. This will help us diagnose the issue more accurately.

[balthaczar23]

I am having the exact same issue on my Windows system. I encountered the same error with torch and fixed it by downloading the missing .dll file and adding it to System32. Then when trying to run prediction I got the same nonsensical output with dozens of predictions clustered at the top of the image.

Were you able to make any progress on this issue?

[glenn-jocher]

@balthaczar23 thank you for sharing your experience. To address this issue, please ensure you are using the latest versions of both Ultralytics and PyTorch. Re-download the YOLOv8 model weights to rule out any corruption. If the problem persists, it might be related to your initial fix with the .dll file. Reinstalling PyTorch and its dependencies could help resolve any underlying issues. If you continue to face problems, please provide any error messages or logs for further assistance.

[balthaczar23]

If anyone is also coming across this issue I ended up solving it by uninstalling torch, torchaudio, and torchvision then installing CUDA Toolkit and reinstalling torch as the CUDA version, found here: https://pytorch.org/get-started/locally/

This fixed the original dependency error with torch and for some reason YOLO then started giving accurate predictions.

[pderrenger]

@balthaczar23 thank you for sharing your solution. It's great to hear that reinstalling PyTorch with CUDA resolved the issue and improved YOLOv8's predictions. For others experiencing similar problems, ensuring compatibility between PyTorch, CUDA, and your hardware is crucial. If issues persist, always verify with the latest versions of the packages. If further assistance is needed, please refer to our documentation or open a new issue.