app.py

import gradio as gr
import spaces
from transformers import Qwen2VLForConditionalGeneration, AutoTokenizer, AutoProcessor
from qwen_vl_utils import process_vision_info
import torch
import base64
from PIL import Image, ImageDraw
from io import BytesIO
import re


models = {
    "Qwen/Qwen2-VL-7B-Instruct": Qwen2VLForConditionalGeneration.from_pretrained("Qwen/Qwen2-VL-7B-Instruct", torch_dtype="auto", device_map="auto")
}

processors = {
    "Qwen/Qwen2-VL-7B-Instruct": AutoProcessor.from_pretrained("Qwen/Qwen2-VL-7B-Instruct")
}

DESCRIPTION = "# Qwen2-VL Object Localization Demo"


def image_to_base64(image):
    buffered = BytesIO()
    image.save(buffered, format="PNG")  # Save the image in memory as PNG
    img_str = base64.b64encode(buffered.getvalue()).decode("utf-8")  # Encode image to base64
    return img_str


def draw_bounding_boxes(image, bounding_boxes, outline_color="red", line_width=2):
    draw = ImageDraw.Draw(image)
    for box in bounding_boxes:
        xmin, xmax, ymin, ymax = box
        draw.rectangle([xmin, ymin, xmax, ymax], outline=outline_color, width=line_width)
    return image


@spaces.GPU
def run_example(image, text_input, model_id="Qwen/Qwen2-VL-7B-Instruct"):
    model = models[model_id].eval()
    processor = processors[model_id]

    messages = [
        {
            "role": "user",
            "content": [
                {"type": "image", "image": f"data:image;base64,{image_to_base64(image)}"},
                {"type": "text", "text": "You are a helpfull assistant to detect objects in images. When asked to detect elements based on a description you return bounding boxes for all elements in the form of [xmin, ymin, xmax, ymax] whith the values beeing scaled to 1000 by 1000 pixels. When there are more than one result answer with a list of bounding boxes in the form of [[xmin, ymin, xmax, ymax], [xmin, ymin, xmax, ymax], ...]."},
                {"type": "text", "text": f"detect {text_input}"},
            ],
        }
    ]

    text = processor.apply_chat_template(
        messages, tokenize=False, add_generation_prompt=True
    )
    image_inputs, video_inputs = process_vision_info(messages)
    inputs = processor(
        text=[text],
        images=image_inputs,
        videos=video_inputs,
        padding=True,
        return_tensors="pt",
    )
    inputs = inputs.to("cuda")

    generated_ids = model.generate(**inputs, max_new_tokens=128)
    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
    )

    pattern = r'\[\s*(\d+)\s*,\s*(\d+)\s*,\s*(\d+)\s*,\s*(\d+)\s*\]'
    matches = re.findall(pattern, str(output_text))
    parsed_boxes = [[int(num) for num in match] for match in matches]
    return output_text, parsed_boxes, draw_bounding_boxes(image, parsed_boxes)

css = """
  #output {
    height: 500px; 
    overflow: auto; 
    border: 1px solid #ccc; 
  }
"""

with gr.Blocks(css=css) as demo:
    gr.Markdown(DESCRIPTION)
    with gr.Tab(label="Qwen2-VL Input"):
        with gr.Row():
            with gr.Column():
                input_img = gr.Image(label="Input Picture", type="pil")
                model_selector = gr.Dropdown(choices=list(models.keys()), label="Model", value="Qwen/Qwen2-VL-7B-Instruct")
                text_input = gr.Textbox(label="Description of Localization Target")
                submit_btn = gr.Button(value="Submit")
            with gr.Column():
                model_output_text = gr.Textbox(label="Model Output Text")
                parsed_boxes = gr.Textbox(label="Parsed Boxes")
                annotated_image = gr.Image(label="Annotated Picture")

        submit_btn.click(run_example, [input_img, text_input, model_selector], [model_output_text, parsed_boxes, annotated_image])

demo.launch(debug=True)