struct_caption.py

import json
import torch

from transformers import Qwen2_5_VLForConditionalGeneration, AutoProcessor
from qwen_vl_utils import process_vision_info

SYSTEM_PROMPT = "I need you to generate a structured and detailed caption for the provided video. The structured output and the requirements for each field are as shown in the following JSON content: {\"subjects\": [{\"appearance\": \"Main subject appearance description\", \"action\": \"Main subject action\", \"expression\": \"Main subject expression  (Only for human/animal categories, empty otherwise)\", \"position\": \"Subject position in the video (Can be relative position to other objects or spatial description)\", \"TYPES\": {\"type\": \"Main category (e.g., Human)\", \"sub_type\": \"Sub-category (e.g., Man)\"}, \"is_main_subject\": true}, {\"appearance\": \"Non-main subject appearance description\", \"action\": \"Non-main subject action\", \"expression\": \"Non-main subject expression (Only for human/animal categories, empty otherwise)\", \"position\": \"Position of non-main subject 1\", \"TYPES\": {\"type\": \"Main category (e.g., Vehicles)\", \"sub_type\": \"Sub-category (e.g., Ship)\"}, \"is_main_subject\": false}], \"shot_type\": \"Shot type(Options: long_shot/full_shot/medium_shot/close_up/extreme_close_up/other)\", \"shot_angle\": \"Camera angle(Options: eye_level/high_angle/low_angle/other)\", \"shot_position\": \"Camera position(Options: front_view/back_view/side_view/over_the_shoulder/overhead_view/point_of_view/aerial_view/overlooking_view/other)\", \"camera_motion\": \"Camera movement description\", \"environment\": \"Video background/environment description\", \"lighting\": \"Lighting information in the video\"}"

class StructCaptioner:
    def __init__(self, model_path):
        self.model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
            model_path,
            torch_dtype=torch.bfloat16,
            # attn_implementation="flash_attention_2",
            device_map="cuda",
        )
        self.processor = AutoProcessor.from_pretrained(model_path)

    def __call__(self, video_path):
        messages = [
            {
                "role": "user",
                "content": [
                    {
                        "type": "video",
                        "video": video_path,
                        "max_pixels": 360 * 420,
                        "fps": 2.0,
                    },
                    {"type": "text", "text": SYSTEM_PROMPT},
                ],
            }
        ]

        text = self.processor.apply_chat_template(
            messages, tokenize=False, add_generation_prompt=True
        )
        image_inputs, video_inputs, video_kwargs = process_vision_info(messages, return_video_kwargs=True)

        inputs = self.processor(
            text=[text],
            images=image_inputs,
            videos=video_inputs,
            padding=True,
            return_tensors="pt",
            **video_kwargs,
        )

        inputs = inputs.to("cuda")

        generated_ids = self.model.generate(**inputs, max_new_tokens=2048, temperature=0.05)
        generated_ids_trimmed = [
            out_ids[len(in_ids) :] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)
        ]
        output_texts = self.processor.batch_decode(
            generated_ids_trimmed, skip_special_tokens=True, clean_up_tokenization_spaces=False
        )
        caption = json.loads(output_texts[0])
        caption = json.dumps(caption, indent=4, ensure_ascii=False)
        return caption