Now I have an action recognition dataset and I want to combine temporal and object-related information for better recognition performance. I wonder how can I train the provided spatio-temporal action detection network in mmaction2 on the action recognition dataset and use it to perform action recognition?