Very interesting work! I have been trying to re-run UCF11 experiment after reading your paper. I have tried a number of formats to prepare the data to fit load_data API. I believe I am missing some steps. What's the input shape for load_data? (nb_frames, 240, 320, 3) or (nb_frames, 120, 160, 3). I noticed that you mentioned to scale it to 160:120 using ffmpeg. Although the comment in the load_data API says the original 320:240. Perhaps if you would like to share a script for the data preparation that would be great. Alternatively, if you can point to your pickle files location, that would help too. Any help would be highly appreciated. Cheers!