You seem to be using the pipelines sequentially on gpu. in order to maximize efficiency please use a dataset