Is there a way to set an OpenCL device in OpenCV with Python?

[yigitayaz262]

How can I set a specific device for OpenCL to use in OpenCV in Python 3? When i run this its using Intel UHD graphics.

 

import cv2
import numpy as np
import time

video = cv2.VideoCapture(0)
boyut = 320
ClassDosyasi = 'coco.names'
ClassIsimleri = []
minKararlilik = 0.5
nmsTreshold = 0.3

with open(ClassDosyasi, 'rt') as f:
    ClassIsimleri = f.read().rstrip('\n').rsplit('\n')

modelConf = 'yolov3-320.cfg'
modelWeight = 'yolov3-320.weights'
net = cv2.dnn.readNetFromDarknet(modelConf, modelWeight)
net.setPreferableBackend(cv2.dnn.DNN_BACKEND_OPENCV)
net.setPreferableTarget(cv2.dnn.DNN_TARGET_OPENCL_FP16)


def objeBulucu(ciktilar, frame):
    yukseklik, genislik, channels = frame.shape
    bbox = []
    classIDs = []
    confs = []
    for cikti in ciktilar:
        for detection in cikti:
            scores = detection[5:]
            classID = np.argmax(scores)
            kararlilik = scores[classID]
            if kararlilik > minKararlilik:
                w, h = int(detection[2] * genislik), int(detection[3] * yukseklik)
                x, y = int(detection[0] * genislik) - w / 2, int((detection[1] * yukseklik) - h / 2)
                bbox.append([x, y, w, h])
                classIDs.append(classID)
                confs.append(float(kararlilik))
    indices = cv2.dnn.NMSBoxes(bbox, confs, minKararlilik, nmsTreshold)
    for i in indices:
        i = i[0]
        box = bbox[i]
        x, y, w, h = box[0], box[1], box[2], box[3]
        endX = int(x + w)
        endY = int(y + h)
        cv2.rectangle(frame, (int(x), int(y)), (endX, endY), (0, 255, 0), 2)
        cv2.putText(frame, f'{ClassIsimleri[classIDs[i]].upper()} {int(confs[i] * 100)}%',
                    (int(x), int(y) - 10), cv2.FONT_HERSHEY_SIMPLEX, 0.7, (0, 0, 255), 2)


while True:
    start_time = time.time()
    basari, frame = video.read()

    blob = cv2.dnn.blobFromImage(frame, 1 / 255, (boyut, boyut), [0, 0, 0], 1, crop=False)
    net.setInput(blob)
    layerIsimleri = net.getLayerNames()
    cikisIsimleri = [layerIsimleri[i[0] - 1] for i in net.getUnconnectedOutLayers()]
    ciktilar = net.forward(cikisIsimleri)
    objeBulucu(ciktilar, frame)
    cv2.putText(frame, "FPS: " + str(int(1.0 / (time.time() - start_time))), (5, 20), cv2.FONT_HERSHEY_SIMPLEX, 0.8,
                (0, 0, 255),2)
    cv2.imshow('Sonuc', frame)
    cv2.waitKey(1)

 

[Sauron]

You can set the OPENCV_OPENCL_DEVICE environment variable https://github.com/opencv/opencv/issues/5165

[yigitayaz262]

14 minutes ago, Sauron said:

You can set the OPENCV_OPENCL_DEVICE environment variable https://github.com/opencv/opencv/issues/5165

How can I set this on windows? @Sauron

 

[yigitayaz262]

@Sauronthanks made it but nothing changed it was 3 fps with Intel UHD and now 3 fps with gtx 1650. How can I speed up this?

[yigitayaz262]

 

 

Its using the only copy section. Is it the reason of 3 fps?

[jj9987]

That GPU doesn't support CUDA either, so not a lot of options there for you.

Is your OpenCV version compiled with OpenCL support?

[yigitayaz262]

Just now, jj9987 said:

That GPU doesn't support CUDA either, so not a lot of options there for you.

Is your OpenCV version compiled with OpenCL support?

yes its working with opencl and using gpu. My gpu is gtx 1650 and cuda software installed.

[Sauron]

6 minutes ago, yigitayaz262 said:

@Sauronthanks made it but nothing changed it was 3 fps with Intel UHD and now 3 fps with gtx 1650. How can I speed up this?

are you certain it's actually using the gpu? kinda seems like it's using the cpu... try using DNN_TARGET_OPENCL instead of DNN_TARGET_OPENCL_FP16, the latter may reference a different env variable.

 

On a side note, why not use CUDA?

2 minutes ago, jj9987 said:

That GPU doesn't support CUDA either

Pretty sure all nvidia GPUs since 2008 support CUDA.

[yigitayaz262]

Just now, Sauron said:

are you certain it's actually using the gpu? kinda seems like it's using the cpu... try using DNN_TARGET_OPENCL instead of DNN_TARGET_OPENCL_FP16, the latter may reference a different env variable.

 

On a side note, why not use CUDA?

Pretty sure all nvidia GPUs since 2008 support CUDA.

When I use cuda it says

cv2.error: OpenCV(4.5.1) C:\Users\appveyor\AppData\Local\Temp\1\pip-req-build-wvn_it83\opencv\modules\dnn\src\dnn.cpp:1382: error: (-215:Assertion failed) preferableBackend != DNN_BACKEND_OPENCV || preferableTarget == DNN_TARGET_CPU || preferableTarget == DNN_TARGET_OPENCL || preferableTarget == DNN_TARGET_OPENCL_FP16 in function 'cv::dnn::dnn4_v20201117::Net::Impl::setUpNet'

[ WARN:0] global C:\Users\appveyor\AppData\Local\Temp\1\pip-req-build-wvn_it83\opencv\modules\videoio\src\cap_msmf.cpp (434) `anonymous-namespace'::SourceReaderCB::~SourceReaderCB terminating async callback

The code again:

modelConf = 'yolov3-tiny.cfg'
modelWeight = 'yolov3-tiny.weights'
net = cv2.dnn.readNetFromDarknet(modelConf, modelWeight)
net.setPreferableBackend(cv2.dnn.DNN_BACKEND_OPENCV)
net.setPreferableTarget(cv2.dnn.DNN_TARGET_CUDA)

 

[yigitayaz262]

I think i need to build it with cuda enabled but idk how to make this to a python library

[Sauron]

13 minutes ago, yigitayaz262 said:

When I use cuda it says

you're using a very recent version of opencv, cuda support isn't great for those right now... anyway did using DNN_TARGET_OPENCL make any difference?

[yigitayaz262]

48 minutes ago, Sauron said:

you're using a very recent version of opencv, cuda support isn't great for those right now... anyway did using DNN_TARGET_OPENCL make any difference?

no same fps as cpu

[yigitayaz262]

can i install older version of opencv and use it with cuda. This fps too important for me. Tried pip install opencv-python==4.0 but its not working

[Sauron]

10 minutes ago, yigitayaz262 said:

no same fps as cpu

There could be a difference in the way windows handles this compared to Linux... have you tried ubuntu? Anyway here is a full tutorial on using CUDA with opencv in python https://www.pyimagesearch.com/2020/02/03/how-to-use-opencvs-dnn-module-with-nvidia-gpus-cuda-and-cudnn/

 

Also if speed is really important for some reason you should take a look at MobileNetSSD, in my testing it's significantly faster than yolov3

[yigitayaz262]

40 minutes ago, Sauron said:

There could be a difference in the way windows handles this compared to Linux... have you tried ubuntu? Anyway here is a full tutorial on using CUDA with opencv in python https://www.pyimagesearch.com/2020/02/03/how-to-use-opencvs-dnn-module-with-nvidia-gpus-cuda-and-cudnn/

 

Also if speed is really important for some reason you should take a look at MobileNetSSD, in my testing it's significantly faster than yolov3

Thinking about windows ubuntu dualboot before this topic. External ssd i ordered will come in 3-4 days. Thanks for help. Bye

[jj9987]

4 hours ago, Sauron said:

On a side note, why not use CUDA?

Pretty sure all nvidia GPUs since 2008 support CUDA.

It's not listed in the CUDA GPU list - https://developer.nvidia.com/cuda-gpus

[Sauron]

14 minutes ago, jj9987 said:

It's not listed in the CUDA GPU list - https://developer.nvidia.com/cuda-gpus

Then that list is incomplete. Considering it runs a Turing chip (minus raytracing) perhaps it was a marketing decision to push the sale of the more expensive 20xx variants.

[yigitayaz262]

13 hours ago, jj9987 said:

It's not listed in the CUDA GPU list - https://developer.nvidia.com/cuda-gpus

But how I have cuda?