【Pytorch API笔记9】用torchnvjpg加速图像解码

骚操作之，opencv读取图像太慢了，图像的加载与解码都直接用cuda

一、安装

conda activate my_base # 先CD到自己的anconda下

git clone https://github.com/itsliupeng/torchnvjpeg.git
 
cd torchnvjpeg/
 
python3 setup.py bdist_wheel  # 编译生成whl文件
 
cd dist

python -m pip install torchnvjpeg-0.1.0-cp37-cp37m-win_amd64.whl

二、使用

注意，要先import torch

否则会报

ImportError: DLL load failed: 找不到指定的模块。

或者

ImportError: libc10.so: cannot open shared object file: No such file or directory

通过先import torch，就可以找到这个文件了

import glob, os
import time
import torch
import torchnvjpeg
import cv2

if __name__ == "__main__":
    path = "./images/"
    img_paths = glob.glob(os.path.join(path, "*.jpg"))
    img_size = len(img_paths)
    decoder = torchnvjpeg.Decoder()
    a = 0
    for img_path in img_paths:
        start = time.time()
        # img = cv2.imread(img_path)
        image_data = open(img_path, 'rb').read()
        image_tensor = decoder.decode(image_data)
        np = image_tensor.data.cpu().numpy()
        end = time.time()
        a += end - start
    print("cost time = {}".format(a / img_size))