How to prepare the dataset?

[AI-Kyo-er]

Hello author @scofield7419, you mentioned in your paper that the AG datasets with STSG which is warpped-up with Visual-instruction-tuning method is used for training. And your code load the dataset like below:

# Action Genome dataset configuration
        self.ag = {
            "target": "motionepic.dataset.base_dataset.LazySupervisedDataset",
            "params": dict(
                data_path="./data/AG/ag_dataset.json",
                video_folder="./data/AG/",
            ),
        }

However, the original downloaded version of https://github.com/JingweiJ/ActionGenome does not have the "./data/AG/ag_dataset.json" you need. I guess some preparing method has been unfortunately missed in your repo, becaused I find the file data/AG/prepare.md is empty.

Would you like to provide any help to solve this? I will be very thankful for your attention and time, and fully appreciate your impressive work of this paper.