I wrapped the transform stage into the onnx model to avoid cpu stuff I guess, along with some other minor improvements this should slightly increase performance on non nvidia GPU's