app.py

import subprocess  # 🥲

subprocess.run(
    "pip install flash-attn --no-build-isolation",
    env={"FLASH_ATTENTION_SKIP_CUDA_BUILD": "TRUE"},
    shell=True,
)
import spaces
import gradio as gr

from transformers import Qwen2VLForConditionalGeneration, AutoProcessor
from qwen_vl_utils import process_vision_info
import torch
import os
import json
from pydantic import BaseModel
from typing import Tuple

os.environ["HF_HUB_ENABLE_HF_TRANSFER"] = "1"


model = Qwen2VLForConditionalGeneration.from_pretrained(
    "Qwen/Qwen2-VL-7B-Instruct",
    torch_dtype=torch.bfloat16,
    attn_implementation="flash_attention_2",
    device_map="auto",
)
processor = AutoProcessor.from_pretrained("Qwen/Qwen2-VL-7B-Instruct")


class GeneralRetrievalQuery(BaseModel):
    broad_topical_query: str
    broad_topical_explanation: str
    specific_detail_query: str
    specific_detail_explanation: str
    visual_element_query: str
    visual_element_explanation: str


def get_retrieval_prompt(prompt_name: str) -> Tuple[str, GeneralRetrievalQuery]:
    if prompt_name != "general":
        raise ValueError("Only 'general' prompt is available in this version")

    prompt = """You are an AI assistant specialized in document retrieval tasks. Given an image of a document page, your task is to generate retrieval queries that someone might use to find this document in a large corpus.

Please generate 3 different types of retrieval queries:

1. A broad topical query: This should cover the main subject of the document.
2. A specific detail query: This should focus on a particular fact, figure, or point made in the document.
3. A visual element query: This should reference a chart, graph, image, or other visual component in the document, if present. Don't just reference the name of the visual element but generate a query which this illustration may help answer or be related to.

Important guidelines:
- Ensure the queries are relevant for retrieval tasks, not just describing the page content.
- Frame the queries as if someone is searching for this document, not asking questions about its content.
- Make the queries diverse and representative of different search strategies.

For each query, also provide a brief explanation of why this query would be effective in retrieving this document.

Format your response as a JSON object with the following structure:

{
  "broad_topical_query": "Your query here",
  "broad_topical_explanation": "Brief explanation",
  "specific_detail_query": "Your query here",
  "specific_detail_explanation": "Brief explanation",
  "visual_element_query": "Your query here",
  "visual_element_explanation": "Brief explanation"
}

If there are no relevant visual elements, replace the third query with another specific detail query.

Here is the document image to analyze:
<image>

Generate the queries based on this image and provide the response in the specified JSON format."""

    return prompt, GeneralRetrievalQuery


# defined like this so we can later add more prompting options
prompt, pydantic_model = get_retrieval_prompt("general")


def _prep_data_for_input(image):
    messages = [
        {
            "role": "user",
            "content": [
                {
                    "type": "image",
                    "image": image,
                },
                {"type": "text", "text": prompt},
            ],
        }
    ]

    text = processor.apply_chat_template(
        messages, tokenize=False, add_generation_prompt=True
    )

    image_inputs, video_inputs = process_vision_info(messages)

    return processor(
        text=[text],
        images=image_inputs,
        videos=video_inputs,
        padding=True,
        return_tensors="pt",
    )


@spaces.GPU
def generate_response(image):
    inputs = _prep_data_for_input(image)
    inputs = inputs.to("cuda")

    generated_ids = model.generate(**inputs, max_new_tokens=200)
    generated_ids_trimmed = [
        out_ids[len(in_ids) :]
        for in_ids, out_ids in zip(inputs.input_ids, generated_ids)
    ]

    output_text = processor.batch_decode(
        generated_ids_trimmed,
        skip_special_tokens=True,
        clean_up_tokenization_spaces=False,
    )
    try:
        return json.loads(output_text[0])
    except Exception:
        gr.Warning("Failed to parse JSON from output")
        return {}


title = "ColPali fine-tuning Query Generator"
description = """[ColPali](https://huggingface.co/papers/2407.01449) is a very exciting new approach to multimodal document retrieval which aims to replace existing document retrievers which often rely on an OCR step with an end-to-end multimodal approach. 

To train or fine-tune a ColPali model, we need a dataset of image-text pairs which represent the document images and the relevant text queries which those documents should match. 
To make the ColPali models work even better we might want a dataset of query/image document pairs related to our domain or task. 

One way in which we might go about generating such a dataset is to use an VLM to generate synthetic queries for us. 
This space uses the [Qwen/Qwen2-VL-7B-Instruct](https://huggingface.co/Qwen/Qwen2-VL-7B-Instruct) VLM model to generate queries for a document, based on an input document image. 

**Note** there is a lot of scope for improving to prompts and the quality of the generated queries! If you have any suggestions for improvements please open a Discussion!

This [blog post](https://danielvanstrien.xyz/posts/post-with-code/colpali/2024-09-23-generate_colpali_dataset.html) gives an overview of how you can use this kind of approach to generate a full dataset for fine-tuning ColPali models. 

If you want to convert a PDF(s) to a dataset of page images you can try out the [ PDFs to Page Images Converter](https://huggingface.co/spaces/Dataset-Creation-Tools/pdf-to-page-images-dataset) Space.

"""

examples = [
    "examples/Approche_no_13_1977.pdf_page_22.jpg",
    "examples/SRCCL_Technical-Summary.pdf_page_7.jpg",
]

demo = gr.Interface(
    fn=generate_response,
    inputs=gr.Image(type="pil"),
    outputs=gr.Json(),
    title=title,
    description=description,
    examples=examples,
)
demo.launch()