Generative Hybrid Representations for Activity Forecasting with No-Regret Learning