How can I speed up inference?

When using this network for 2D image inference, it takes several seconds per image. Besides setting disable_tta to True and increasing tile_step_size, what other optimizations can be made?