Currently devolearn uses cv2 to load and run inference on images/videos. There are faster and cleaner alternatives like decord and imageio which can be used instead.