In Tensorflow 2.0, I want to process images before writing a TFRecord file. My images come as numpy arrays. It seems I cannot find the good combination between reader and writer