Does YOLOv5 6.2 work with CUDA 12.x?

[rkn]

I've added a 3060 12GB to my BI PC and installed CUDA 12.6.2 and ran the install script for cuDNN. Immediately I could see the existing YOLOv5 .NET using DirectML was indeed using the 3080 GPU instead of the integrated one. Simlarly if I switch to YOLOv8 it also starts and uses the GPU using CUDA, but this module has limited models at present so wanted to try YOLOv5.

However YOLOv5 6.2 starts and just uses the CPU even though its info says to use CUDA. I see a couple of posts from last year on various forums saying it doesn't work with CUDA v12 so is that still the case or have I missed something?

Info for reference (using CPAI 2.6.5):

Module 'Object Detection (YOLOv5 6.2)' 1.9.2 (ID: ObjectDetectionYOLOv5-6.2)
Valid:            True
Module Path:      <root>\modules\ObjectDetectionYOLOv5-6.2
Module Location:  Internal
AutoStart:        True
Queue:            objectdetection_queue
Runtime:          python3.7
Runtime Location: Shared
FilePath:         detect_adapter.py
Start pause:      1 sec
Parallelism:      0
LogVerbosity:    
Platforms:        all,!raspberrypi,!jetson
GPU Libraries:    installed if available
GPU:              use if supported
Accelerator:    
Half Precision:   enable
Environment Variables
   APPDIR             = <root>\modules\ObjectDetectionYOLOv5-6.2
   CUSTOM_MODELS_DIR  = <root>\modules\ObjectDetectionYOLOv5-6.2\custom-models
   MODELS_DIR         = <root>\modules\ObjectDetectionYOLOv5-6.2\assets
   MODEL_SIZE         = Medium
   USE_CUDA           = True
   YOLOv5_AUTOINSTALL = false
   YOLOv5_VERBOSE     = false
Status Data:  {
  "inferenceDevice": "CPU",
  "inferenceLibrary": "",
  "canUseGPU": "false",
  "successfulInferences": 13,
  "failedInferences": 0,
  "numInferences": 13,
  "averageInferenceMs": 131.3846153846154
}

 

[mvoss]

I believe you have to install Cuda before you install Code Project.

Michael

 

[MikeLud1]

I've added a 3060 12GB to my BI PC and installed CUDA 12.6.2 and ran the install script for cuDNN. Immediately I could see the existing YOLOv5 .NET using DirectML was indeed using the 3080 GPU instead of the integrated one. Simlarly if I switch to YOLOv8 it also starts and uses the GPU using CUDA, but this module has limited models at present so wanted to try YOLOv5.

However YOLOv5 6.2 starts and just uses the CPU even though its info says to use CUDA. I see a couple of posts from last year on various forums saying it doesn't work with CUDA v12 so is that still the case or have I missed something?

Info for reference (using CPAI 2.6.5):

Module 'Object Detection (YOLOv5 6.2)' 1.9.2 (ID: ObjectDetectionYOLOv5-6.2)
Valid:            True
Module Path:      <root>\modules\ObjectDetectionYOLOv5-6.2
Module Location:  Internal
AutoStart:        True
Queue:            objectdetection_queue
Runtime:          python3.7
Runtime Location: Shared
FilePath:         detect_adapter.py
Start pause:      1 sec
Parallelism:      0
LogVerbosity:   
Platforms:        all,!raspberrypi,!jetson
GPU Libraries:    installed if available
GPU:              use if supported
Accelerator:   
Half Precision:   enable
Environment Variables
   APPDIR             = <root>\modules\ObjectDetectionYOLOv5-6.2
   CUSTOM_MODELS_DIR  = <root>\modules\ObjectDetectionYOLOv5-6.2\custom-models
   MODELS_DIR         = <root>\modules\ObjectDetectionYOLOv5-6.2\assets
   MODEL_SIZE         = Medium
   USE_CUDA           = True
   YOLOv5_AUTOINSTALL = false
   YOLOv5_VERBOSE     = false
Status Data:  {
  "inferenceDevice": "CPU",
  "inferenceLibrary": "",
  "canUseGPU": "false",
  "successfulInferences": 13,
  "failedInferences": 0,
  "numInferences": 13,
  "averageInferenceMs": 131.3846153846154
}

Since the RTX 3060 was not installed when the YOLOv5 6.2 module was installed it is configured for CPU.
To have it configured for the RTX 3060 you need to uninstall the YOLOv5 6.2 module then reinstall.
I have the same RTX 3060 GPU in my main BI system and the YOLOv5 .NET using DirectML is faster then the YOLOv5 6.2 module using CUDA.

 

[rkn]

Since the RTX 3060 was not installed when the YOLOv5 6.2 module was installed it is configured for CPU.
To have it configured for the RTX 3060 you need to uninstall the YOLOv5 6.2 module then reinstall.
I have the same RTX 3060 GPU in my main BI system and the YOLOv5 .NET using DirectML is faster then the YOLOv5 6.2 module using CUDA.

Thanks Mike, but I tried this and no change as it still only offers and uses CPU for YOLOv5.

If its not going to be faster then its not worth me wasting much time on this but could the fact I am still running CPAI 2.6.5 rather than a later release have any impact on YOLOv5 not using the GPU?