PP-OCRv4_mobile_rec

[fightingshao]

使用create_pipeline中ocr识别，其中使用模型PP-OCRv4_mobile_rec、PP-OCRv4_mobile_det，在paddlex/temp/下面生成大量的文本分割图片，请问如何设置取消图片保存

[TingquanGao]

paddlex/temp/中的内容会在程序退出时自动释放，是在程序退出后没有释放吗？系统环境是什么呢？

[fightingshao]

是CentOS Linux release 7.6.1810(Core)

[fightingshao]

请问可以取消图片保存吗？我这是服务，所以需要一直后台启动。。。

[liuhongen1234567]

您好，PaddleX在识别时，当图片传入的时ndarray会默认创建临时目录保存图片，可以通过修改 https://github.com/PaddlePaddle/PaddleX/blob/22e7501aea85166cc973a193cdbe7f0a4e809459/paddlex/inference/components/transforms/image/common.py 中的 ReadImage 来取消掉创建临时目录，具体修改如下： `class ReadImage(_BaseRead): INPUT_KEYS = ["img"] OUTPUT_KEYS = ["img", "img_size", "ori_img", "ori_img_size"] DEAULT_INPUTS = {"img": "img"} DEAULT_OUTPUTS = { "img": "img", "input_path": "input_path", "img_size": "img_size", "ori_img": "ori_img", "ori_img_size": "ori_img_size", }

_FLAGS_DICT = {
    "BGR": cv2.IMREAD_COLOR,
    "RGB": cv2.IMREAD_COLOR,
    "GRAY": cv2.IMREAD_GRAYSCALE,
}

SUFFIX = ["jpg", "png", "jpeg", "JPEG", "JPG", "bmp", "PDF", "pdf"]

def __init__(self, batch_size=1, format="BGR"):
    """
    Initialize the instance.

    Args:
        format (str, optional): Target color format to convert the image to.
            Choices are 'BGR', 'RGB', and 'GRAY'. Default: 'BGR'.
    """
    super().__init__(batch_size)
    self.format = format
    flags = self._FLAGS_DICT[self.format]
    self._img_reader = ImageReader(backend="opencv", flags=flags)
    self._pdf_reader = PDFReader()
    self._writer = ImageWriter(backend="opencv")

def apply(self, img):
    """apply"""

    def process_ndarray(img):
        # with temp_file_manager.temp_file_context(suffix=".png") as temp_file:
        #     img_path = Path(temp_file.name)
        #     self._writer.write(img_path, img)
        #     print("="*100)
        #     print("img_path", img_path)
        #     if self.format == "RGB":
        #         img = img[:, :, ::-1]
        img_path = ""
        return {
            "input_path": img_path,
            "img": img,
            "img_size": [img.shape[1], img.shape[0]],
            "ori_img": deepcopy(img),
            "ori_img_size": deepcopy([img.shape[1], img.shape[0]]),
        }`

需要注意的是，这种改法目前只在OCR测试有效，并且可能导致其他产线运行失败，后续我们会考虑支持该功能。