Function do_induct

utils/src/experiment/preprocess.py:212–226

(llm: list[llms.LLM], ppt_folder: str, rank: int)

    ]

    def do_induct(llm: list[llms.LLM], ppt_folder: str, rank: int):
        if not older_than(pjoin(ppt_folder, "source.pptx"), wait=wait):
            return
        llms.language_model = llm[0]
        llms.vision_model = llm[1]
        config = Config(rundir=ppt_folder)
        ppt_image_folder = pjoin(ppt_folder, "source_slides")
        template_image_folder = pjoin(ppt_folder, "template_images")
        image_model = get_image_model(f"cuda:{rank % device_count}")
        presentation = Presentation.from_file(pjoin(ppt_folder, "source.pptx"), config)
        ImageLabler(presentation, config).caption_images()
        slide_inducter = SlideInducter(
            presentation, ppt_image_folder, template_image_folder, config, image_model
        )
        slide_inducter.content_induct()

    for folder in tqdm(sorted(glob.glob("data/*/pptx/*")), desc="prepare induction"):
        do_induct(induct_llms[induct_id], folder, 0)