Cpp inference of YOLOV8

Support YOLO
, ,
1

hi guys!
i have C++ code to run YOLOV8 and it works well on RTX4090.
but when i run the same code with same CUDA , cudnn and OPENCV versions on Quadro K620, i get error in the following line:

net.forward(outputs, net.getUnconnectedOutLayersNames());

can someone help me with this?
is this a problem with QUADRO or something else?
and since i am running the code in release mode, i can’t see the cause of the error.

2

Hi there!

It’s great to hear that your C++ code for running YOLOv8 works well on the RTX 4090! The issue you’re encountering with the Quadro K620 could be due to several factors. Here are a few things to consider:

  1. CUDA and cuDNN Compatibility: Ensure that the versions of CUDA and cuDNN you are using are compatible with the Quadro K620. The K620 is an older GPU, and it might not support the latest versions of these libraries. You can check the compatibility matrix on the NVIDIA CUDA Toolkit Documentation.

  2. Compute Capability: The Quadro K620 has a compute capability of 5.0, which might not be sufficient for some of the operations required by YOLOv8. You can verify the compute capability requirements in the YOLOv8 documentation.

  3. OpenCV Version: Ensure that the OpenCV version you are using is compatible with both your CUDA and cuDNN versions. Sometimes, mismatches can cause issues during the net.forward call.

  4. Error Logging: Since you’re running in release mode and can’t see the error details, consider running a debug build to capture more information about the error. This can help pinpoint the exact cause.

  5. Driver Updates: Make sure that your GPU drivers are up to date. Sometimes, updating the drivers can resolve compatibility issues.

Here’s a small checklist to help you troubleshoot:

  • Verify CUDA and cuDNN versions compatibility with Quadro K620.
  • Check the compute capability requirements for YOLOv8.
  • Ensure OpenCV version compatibility.
  • Run a debug build to capture detailed error logs.
  • Update GPU drivers.

If the issue persists, you might want to consider using a different GPU that meets the requirements for YOLOv8. For additional details, you can refer to the Ultralytics YOLO FAQ.

I hope this helps! If you have any more questions, feel free to ask. :blush:

1 Like
3

thanks for your help.
my cudnn and CUDA(11.8) versions are compatible with opencv 4.7 in rtx 4090.
so i think the problem is not with opencv version.
nvidia drivers are also up to date.
but i tested it also on rtx 1050 and the same issue appeared.

4

@Nima_Chelongar a few things:

  1. You mentioned an error message but didn’t share it. This makes it much more difficult to understand the issue you’re facing. It’s important to remember that no one has as much information about the problem as you, the less you share about it, the less likely anyone will be able (or possibly willing) to help. Whenever you encounter an error, you should post the entire error message and the code that causes the error.

  2. There is no RTX 1050, you could you mean GTX 1050? The GTX 1050

  3. The K620 GPU is quite old and I suspect might not have a supported PyTorch version, or at least isn’t supported in the oldest PyTorch version that Ultralytics supports. It’s also likely to be very slow as it uses PCIe 2.0.

  • The environment (CUDA Toolkit, cuDNN, and/or PyTorch) you use for the RTX 4090 is unlikely to also work with the K620, and I would guess it might work with a GTX 1050 GPU.
  1. When it comes to C++ I’m not very familiar with the functionality, and I suspect there will be a limited number of users who actually are that use Ultralytics. All the recommendations made in the post above are good ones to follow and should be your starting point with this problem.

One final note. Without a matching system, it’s pretty close to impossible for anyone to replicate your problem, so the conclusion will have to be (until you find the solution) that it’s a hardware or environment setup related issue. Consider that every PC component, driver, and software version as a variable, plus the OS and your specific code; so there are really a lot of variables to consider in diagnosing this issue. The most obvious source would be version/hardware incompatibility since you have it working with an RTX 4090 and only see an issue with other hardware.

5

thank you very much!
yes my bad, it is GTX 1050.
since i am running release version, i don’t have any error massage to show you. :slight_smile:
but if you need anything more of my code, i would share here.

6

Nima,

It sounds like the issue may be due to hardware limitations of the Quadro K620, which is significantly less powerful than the RTX 4090, especially concerning its CUDA cores and overall computational capability. This discrepancy can affect the performance and compatibility when handling AI inference tasks like those required by YOLOv8.

Here are a few steps to troubleshoot and potentially resolve your issue:

  1. Ensure your GPU Drivers are up to date: Outdated drivers can sometimes cause compatibility issues.
  2. Check GPU capabilities: Make sure that the Quadro K620 meets the minimum requirements for the configurations used in your YOLOv8 model.
  3. Debugging: Since you mentioned running in release mode limits the visibility of the error, try running it in debug mode if possible to get more detailed error logs. This might pinpoint what’s causing the error on the Quadro K620.

If continuous problems arise, considering the hardware gap, using a more capable GPU may be necessary to run your YOLOv8 model efficiently.

Let me know if this helps or if you have any other questions!

1 Like
7

Hi,

As per my knowledge it’s due to the Quadro K620 not fully supporting the required CUDA or cuDNN features used by YOLOv8. The RTX 4090 is a much more powerful and recent GPU compared to the Quadro K620, which may not support the necessary compute capabilities or tensor operations required by the YOLOv8 model.

You must ensure that the Quadro K620 has the required compute capability for the version of YOLOv8 you are using. YOLOv8 may require a GPU with a compute capability of 6.0 or higher, while the Quadro K620 has a compute capability of 5.0, which could be a limitation.

Thanks