MARE's Computer Vision Study.

1/25/2023

RuntimeError: CUDA error: CUBLAS_STATUS_INVALID_VALUE when calling `cublasSgemm( handle, opa, opb, m, n, k, &alpha, a, lda, b, ldb, &beta, c, ldc)`

I was suffering above (title) error during few hours.

The reason is wrong cuda version installed with pytorch.

My cuda version is 11.6, but install version is 11.7.

So I print it -> "torch.__version__"

It returend -> "1.13.1+cu117"

You can check your cuda version using this command

> nvidia-smi

so, remove all torch, torchvision packaged

> pip uninstall torch torchvision

and install again

ex)

pip3 install torch torchvision torchaudio --extra-index-url \https://download.pytorch.org/whl/cu116

Then it works well.

Thank you.

🙇🏻‍♂️

1/23/2023

PIL image to tensor

refer to code

key is "transforms.PILToTensor"

import torch
from PIL import Image
import torchvision.transforms as transforms
  
pil_image = Image.open('input.jpg')  
transform = transforms.Compose([
    transforms.PILToTensor()
])
tensor_img = transform(pil_image)
print(tensor_img)

Thx. 🙇🏻‍♂️

1/19/2023

opencv c++ gpumat mat(cpu) memory share

refer to code:

size_t frameByteSize = initMat.step[0] * initMat.rows;

#ifndef USE_UNIFIED_MEM
    /* Pinned memory. No cache */
    std::cout << "Using pinned memory" << std::endl;
    void* device_ptr, * host_ptr;
    cudaSetDeviceFlags(cudaDeviceMapHost);
    cudaHostAlloc((void**)&host_ptr, frameByteSize, cudaHostAllocMapped);
    cudaHostGetDevicePointer((void**)&device_ptr, (void*)host_ptr, 0);
    cv::Mat frame_out(height, width, CV_8UC3, host_ptr);
    cv::cuda::GpuMat d_frame_out(height, width, CV_8UC3, device_ptr);
#else
    /* Unified memory */
    std::cout << "Using unified memory" << std::endl;
    void* unified_ptr;
    cudaMallocManaged(&unified_ptr, frameByteSize);
    cv::Mat frame_out(height, width, CV_8UC3, unified_ptr);
    cv::cuda::GpuMat d_frame_out(height, width, CV_8UC3, unified_ptr);
#endif

another example

void* m_device_ptr = NULL;
void* m_host_ptr = NULL;

//memory share frame -> cuda_frame
cudaSetDeviceFlags(cudaDeviceMapHost);
size_t frameByteSize = initMat.step[0] * initMat.rows;
cudaHostAlloc((void**)&m_host_ptr, frameByteSize, cudaHostAllocMapped);
cudaHostGetDevicePointer((void**)&m_device_ptr, (void*)m_host_ptr, 0);
cv::Mat m_frame = cv::Mat(initMat.rows, initMat.cols, CV_8UC3, m_host_ptr);
cv::cuda::GpuMat m_cuda_frame = cv::cuda::GpuMat(initMat.rows, initMat.cols, CV_8UC3, m_device_ptr);

//initiation
iframes.copyTo(m_frame);

1/12/2023

Opencv Mat <-> UMat , convert Mat to UMat, UMat to Mat

Mat to UMat

c++

UMat umat; 

mat.copyTo(umat);

Python

UMat umat = mat.getUMat( flag );

UMat to Mat

C++

Mat mat = umat.getMat( flag );

Python

mat = cv2.UMat.get(umat)

1/11/2023

How to call module written with argparse in iPython notebook

Add this two line above "argparse.ArgumentParser()" line

ex)

##add sys code ##
import sys
sys.argv = ['']

##origin code##
import argparse
parser = argparse.ArgumentParser()
args = parser.parse_args()

Thx. 🙇🏻‍♂️

www.marearts.com

1/10/2023

IProgress not found. Please update jupyter and ipywidgets.

pip install ipywidgets
jupyter nbextension enable --py widgetsnbextension

thx

idx2label, label2idx python code