Function build_data_episodes

eval/run_predict_odyssey.py:330–409

Transfrom the episode trajectory into Odyssey data. This function work as the data preprocess of the origin repo. Args: episode_dir (str): the directory of the episodes. episodes_files (List[str]): the files in the episodes directory. dataset_name (str): the na

(episode_dir:str,
                        episodes_files:List[str],
                        dataset_name:str,
                        his_len:int = 4)

        raise NotImplementedError

def build_data_episodes(episode_dir:str,
                        episodes_files:List[str],
                        dataset_name:str,
                        his_len:int = 4) -> List[Dict]:
    """
    Transfrom the episode trajectory into Odyssey data.

    This function work as the data preprocess of the origin repo.

    Args:
        episode_dir (str): the directory of the episodes.
        episodes_files (List[str]): the files in the episodes directory.
        dataset_name (str): the name of the dataset.
        his_len (int, optional): the max length of history included. Defaults to 4.

    Returns:
        List[Dict]: a series of data that would be processed later.
    """

    res:list = []

    for episodes_file in episodes_files:
        episodes_path = os.path.join(episode_dir, episodes_file, f"{episodes_file}.json")
        try:
            with open(episodes_path, 'r', encoding='utf-8') as f:
                episodes = json.load(f)
        except Exception as e:
            print(f"Failed to load {episodes_path}: {e}")
            continue
            # Skip this file on error

        history_action:list = []
        history_screenshot:list = []

        for episode in episodes:

            image_path = os.path.join(episode_dir, episodes_file, f"{episodes_file}_{episode['step_id']}.jpeg")
            if not os.path.exists(image_path):
                image_path = image_path.replace(".jpeg", ".png")
                if not os.path.exists(image_path):
                    image_path = episode['image_path']

            gt = transform_actions(episode)

            if dataset_name == 'android_control_high_test':
                instruction = episode["low_instruction"]
            else:
                instruction = episode["instruction"]

            question = f"Picture 1: <img>{image_path}</img>\nI'm looking for guidance on how to {instruction}"

            history_action = history_action[-his_len :]

            if IMAGE_HISTORY:
                if len(history_action) > 0:
                    his_img = f'\nPrevious screenshots: <img>image-history: {image_path}</img>'
                    his_str = '\nPrevious Actions: '
                    for idx, hi in enumerate(history_action):