> This is exactly the approach we detail below and in our paper: we modify libjpeg to output DCT coefficients directly to TensorFlow Why not just implement 2D

> This is exactly the approach we detail below and in our paper: we modify libjpeg to output DCT coefficients directly to TensorFlow Why not just implement 2D DCT in TensorFlow or PyTorch or whatever you use, and keep the entire pipeline in GPU memory? A quick glance at the TensorFlow reference shows that it already has tf.signal.fft2d. I suppose you could just implement tf.signal.dct2d similarly. (Disclaimer: I haven't used TensorFlow in a while, I've been doing everything in PyTorch of late.)