1.25.4

• Addition of automatic INT8 calibration process for RGBA 4-channel images.
• It should be noted that the MS-COCO image set does not include an alpha channel, so this auto-calibration does not allow for decent quantization.
• There is no correlation between the channel being 4 and the input data being an image. Therefore, if the model to be converted is data other than a 4-channel image, automatic calibration should not be used.
  • The following means and standard deviations are used as fixed values.

    mean = np.asarray([[[[0.485, 0.456, 0.406, 0.000]]]], dtype=np.float32)
    std = np.asarray([[[[0.229, 0.224, 0.225, 1.000]]]], dtype=np.float32)

  • Also, use a fixed value of 0.5 for the alpha channel of the calibration data.

    new_element_array = np.full((*calib_data.shape[:-1], 1), 0.500, dtype=np.float32)

• magic_touch.onnx.zip
  
• [TODO] Add 4 channels of image data to the sample data for quantization #411

What's Changed

• Addition of automatic INT8 calibration process for RGBA 4-channel images by @PINTO0309 in #670

Full Changelog: 1.25.3...1.25.4