Turning convolution to matrix product requires remapping the image values into a different matrix. Presumably, this is done once as preprocessing, so it doesn't incur a cost during the actual runtime (though this probably doesn't matter too much since DNN's are not trained in a real-time setting usually).
Turning convolution to matrix product requires remapping the image values into a different matrix. Presumably, this is done once as preprocessing, so it doesn't incur a cost during the actual runtime (though this probably doesn't matter too much since DNN's are not trained in a real-time setting usually).