caption.py

import argparse
import glob
import json
import os

import torch
from accelerate import PartialState
from PIL import Image
from tqdm import tqdm
from transformers import AutoModel, AutoTokenizer


class caption_processor:
    def __init__(self, vlm_name, device):
        self.vlm = AutoModel.from_pretrained(
            vlm_name,
            trust_remote_code=True,
            attn_implementation="flash_attention_2",
            torch_dtype=torch.bfloat16,
        )
        self.vlm_tokenizer = AutoTokenizer.from_pretrained(
            vlm_name, trust_remote_code=True
        )

        self.vlm = self.vlm.eval().to(device)
        self.prompt = """
            1. describe the image in brief, Avoid using phrases in [In the/The image/scene shows/contains/is a] in the captions, directly describe the contents.
            2. Imagine this picture is the first frame of a 5-second video. Please describe the video and add dynamics, including the movement of objects and themes, as well as the overall camera movement.Avoid using phrases in [In the/The video/scene shows/contains/is a] in the descriptions, directly describe the contents.
            3. Please output in JSON format.{"caption": "...","video_description": "..."}
            """  # noqa: E501

    def str_2_json(self, str):
        # Find the first occurrence of '{'
        start_idx = str.find("{")
        if start_idx == -1:
            return None

        # Find the last occurrence of '}'
        end_idx = str.rfind("}")
        if end_idx == -1 or end_idx <= start_idx:
            return None

        # Extract the JSON string
        json_str = str[start_idx : end_idx + 1]

        # Load and return the JSON
        try:
            import json

            return json.loads(json_str)
        except json.JSONDecodeError:
            return None

    def process(self, image):
        msgs = [{"role": "user", "content": [image, self.prompt]}]

        answer = self.vlm.chat(
            msgs=msgs, tokenizer=self.vlm_tokenizer, enable_thinking=False, stream=False
        )

        dict_answer = self.str_2_json(answer)
        if dict_answer is None:
            return {"response": answer}

        return dict_answer


def get_images_from_path(path):
    if os.path.isdir(path):
        return glob.glob(os.path.join(path, "*.jpg")) + glob.glob(
            os.path.join(path, "*.png")
        )
    elif os.path.isfile(path) and (path.endswith(".jpg") or path.endswith(".png")):
        return [path]
    else:
        return []


def parse_args():
    parser = argparse.ArgumentParser(description="Caption processor")
    parser.add_argument("--vlm_name", type=str, required=True)
    parser.add_argument("--output_dir", type=str, required=True)
    parser.add_argument("--paths", type=str, required=True, nargs="+")
    return parser.parse_args()


if __name__ == "__main__":
    distributed_state = PartialState()
    args = parse_args()
    output_dir = args.output_dir
    os.makedirs(output_dir, exist_ok=True)

    vlm_name = args.vlm_name
    paths = args.paths
    all_paths = []
    for path in paths:
        images = get_images_from_path(path)
        all_paths.extend(images)
    print("found", len(all_paths), "images")

    processor = caption_processor(
        vlm_name,
        distributed_state.device,
    )
    with distributed_state.split_between_processes(
        all_paths, apply_padding=False
    ) as batched_paths:
        print("GPU", distributed_state.device, "found", len(batched_paths), "images")

        for path in tqdm(batched_paths, desc="Processing images"):
            try:
                json_path = os.path.join(output_dir, os.path.basename(path) + ".json")
                if os.path.exists(json_path):
                    print(f"File {json_path} already exists, skipping.")
                    continue

                image = Image.open(path)
                output = None

                for _ in range(3):
                    output = processor.process(image)
                    if output is not None:
                        break

                if output is None:
                    raise Exception("Failed to process image after 3 attempts")
                else:
                    with open(
                        json_path,
                        "w",
                        encoding="utf-8",
                    ) as f:
                        json.dump(output, f, ensure_ascii=False, indent=2)
            except Exception as e:
                print(f"Error processing {path}: {e}")