app.py

import base64
import json
from datetime import datetime
import gradio as gr
import torch
import spaces
from PIL import Image, ImageDraw
from qwen_vl_utils import process_vision_info
from transformers import Qwen2VLForConditionalGeneration, AutoProcessor
import ast
import os
import numpy as np
from huggingface_hub import hf_hub_download, list_repo_files

# Define constants
DESCRIPTION = "[UGround Demo](https://osu-nlp-group.github.io/UGround/)"
_SYSTEM = "You are a very helpful assistant."
MIN_PIXELS = 256 * 28 * 28
MAX_PIXELS = 1344 * 1344

# Specify the model repository and destination folder
# https://huggingface.co/osunlp/UGround-V1-2B
model_repo = "osunlp/UGround-V1-2B"
destination_folder = "./UGround-V1-2B"

# Ensure the destination folder exists
os.makedirs(destination_folder, exist_ok=True)

# List all files in the repository
files = list_repo_files(repo_id=model_repo)

# Download each file to the destination folder
for file in files:
    file_path = hf_hub_download(repo_id=model_repo, filename=file, local_dir=destination_folder)
    print(f"Downloaded {file} to {file_path}")

model = Qwen2VLForConditionalGeneration.from_pretrained(
    destination_folder,
    torch_dtype=torch.bfloat16,
    device_map="cpu",
)

# Load the processor
processor = AutoProcessor.from_pretrained("Qwen/Qwen2-VL-2B-Instruct", min_pixels=MIN_PIXELS, max_pixels=MAX_PIXELS)

# Helper functions
def draw_point(image_input, point=None, radius=5):
    """Draw a point on the image."""
    if isinstance(image_input, str):
        image = Image.open(image_input)
    else:
        image = Image.fromarray(np.uint8(image_input))

    if point:
        x, y = round(point[0]/1000 * image.width), round(point[1]/1000 * image.height)
        ImageDraw.Draw(image).ellipse((x - radius, y - radius, x + radius, y + radius), fill='red')
    return image

def array_to_image_path(image_array, session_id):
    """Save the uploaded image and return its path."""
    if image_array is None:
        raise ValueError("No image provided. Please upload an image before submitting.")
    img = Image.fromarray(np.uint8(image_array))
    filename = f"{session_id}.png"
    img.save(filename)
    return os.path.abspath(filename)

def crop_image(image_path, click_xy, crop_factor=0.5):
    """Crop the image around the click point."""
    image = Image.open(image_path)
    width, height = image.size
    crop_width, crop_height = int(width * crop_factor), int(height * crop_factor)

    center_x, center_y = int(click_xy[0] * width), int(click_xy[1] * height)
    left = max(center_x - crop_width // 2, 0)
    upper = max(center_y - crop_height // 2, 0)
    right = min(center_x + crop_width // 2, width)
    lower = min(center_y + crop_height // 2, height)

    cropped_image = image.crop((left, upper, right, lower))
    cropped_image_path = f"cropped_{os.path.basename(image_path)}"
    cropped_image.save(cropped_image_path)

    return cropped_image_path

@spaces.GPU
def run_showui(image, query, session_id, iterations=1):
    """Main function for iterative inference."""
    image_path = array_to_image_path(image, session_id)
    
    click_xy = None
    images_during_iterations = []  # List to store images at each step

    for _ in range(iterations):
        messages = [
            {
                "role": "user",
                "content": [
                    {"type": "text", "text": "You are a very helpful assistant"},
                    {"type": "image", "image": image_path, "min_pixels": MIN_PIXELS, "max_pixels": MAX_PIXELS},
                    {"type": "text", "text": f"""Your task is to help the user identify the precise coordinates (x, y) of a specific area/element/object on the screen based on a description.

  - Your response should aim to point to the center or a representative point within the described area/element/object as accurately as possible.
  - If the description is unclear or ambiguous, infer the most relevant area or element based on its likely context or purpose.
  - Your answer should be a single string (x, y) corresponding to the point of the interest.

  Description: {query}

  Answer:"""}
                ],
            }
        ]

        global model
        model = model.to("cuda")

        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,temperature=0)
        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
        )[0]

        click_xy = ast.literal_eval(output_text)

        # Draw point on the current image
        result_image = draw_point(image_path, click_xy, radius=10)
        images_during_iterations.append(result_image)  # Store the current image

        # Crop the image for the next iteration
        image_path = crop_image(image_path, click_xy)

    return images_during_iterations, str(click_xy)

def save_and_upload_data(image, query, session_id, is_example_image, votes=None):
    """Save the data to a JSON file and upload to S3."""
    if is_example_image == "True":
        return

    votes = votes or {"upvotes": 0, "downvotes": 0}

    # Save image locally
    image_file_name = f"{session_id}.png"
    image.save(image_file_name)

    data = {
        "image_path": image_file_name,
        "query": query,
        "votes": votes,
        "timestamp": datetime.now().isoformat()
    }
    
    local_file_name = f"{session_id}.json"
    
    with open(local_file_name, "w") as f:
        json.dump(data, f)

    return data

def update_vote(vote_type, session_id, is_example_image):
    """Update the vote count and re-upload the JSON file."""
    if is_example_image == "True":
        return "Example image."

    local_file_name = f"{session_id}.json"
    
    with open(local_file_name, "r") as f:
        data = json.load(f)
    
    if vote_type == "upvote":
        data["votes"]["upvotes"] += 1
    elif vote_type == "downvote":
        data["votes"]["downvotes"] += 1
    
    with open(local_file_name, "w") as f:
        json.dump(data, f)

    return f"Thank you for your {vote_type}!"

with open("./assets/showui.png", "rb") as image_file:
    base64_image = base64.b64encode(image_file.read()).decode("utf-8")


# [
#                     [f"{cur_dir}/amazon.jpg",f"Search bar at the top of the page"],
#                     [f"{cur_dir}/shopping.jpg", f"delete button for the second item in the cart list"],
#                     [f"{cur_dir}/ios.jpg", f"Open Maps"],
#                     [f"{cur_dir}/toggle.jpg", f"toggle button labeled by VPN"],
#                     [f"{cur_dir}/semantic.jpg", f"Home"],
#                     [f"{cur_dir}/accweather.jpg", f"Select May"],
#                     [f"{cur_dir}/arxiv.jpg", f"Home"],
#                     [f"{cur_dir}/arxiv.jpg", f"Edit the page"],
#                     [f"{cur_dir}/ios.jpg", f"icon at the top right corner"],
#                     [f"{cur_dir}/health.jpg", f"text labeled by 2023/11/26"],


examples = [
["./examples/amazon.jpg", "Search bar at the top of the page", True],
["./examples/shopping.jpg", "delete button for the second item in the cart list", True],
["./examples/ios.jpg", "Open Maps", True],
["./examples/toggle.jpg", "toggle button labeled by VPN", True],
["./examples/semantic.jpg", "Home", True],
["./examples/accweather.jpg", "Select May", True],
["./examples/arxiv.jpg", "Home", True],
["./examples/arxiv.jpg", "Edit the page", True],
["./examples/ios.jpg", "icon at the top right corner", True],
["./examples/health.jpg", "text labeled by 2023/11/26", True],
    ["./examples/app_store.png", "Download Kindle.", True],
    ["./examples/ios_setting.png", "Turn off Do not disturb.", True],
    # ["./examples/apple_music.png", "Star to favorite.", True],
    # ["./examples/map.png", "Boston.", True],
    # ["./examples/wallet.png", "Scan a QR code.", True],
    # ["./examples/word.png", "More shapes.", True],
    # ["./examples/web_shopping.png", "Proceed to checkout.", True],
    # ["./examples/web_forum.png", "Post my comment.", True],
    # ["./examples/safari_google.png", "Click on search bar.", True],
]

def build_demo(embed_mode, concurrency_count=1):
    with gr.Blocks(title="UGround Demo", theme=gr.themes.Default()) as demo:
        state_image_path = gr.State(value=None)
        state_session_id = gr.State(value=None)

        # if not embed_mode:
        #     gr.HTML(
        #         f"""
        #         <div style="text-align: center; margin-bottom: 20px;">
        #             <div style="display: flex; justify-content: center;">
        #                 <img src="https://raw.githubusercontent.com/showlab/ShowUI/refs/heads/main/assets/showui.jpg" alt="ShowUI" width="320" style="margin-bottom: 10px;"/>
        #             </div>
        #             <p>ShowUI is a lightweight vision-language-action model for GUI agents.</p>
        #             <div style="display: flex; justify-content: center; gap: 15px; font-size: 20px;">
        #                 <a href="https://huggingface.co/showlab/ShowUI-2B" target="_blank">
        #                     <img src="https://img.shields.io/badge/%F0%9F%A4%97%20Hugging%20Face-ShowUI--2B-blue" alt="model"/>
        #                 </a>
        #                 <a href="https://arxiv.org/abs/2411.17465" target="_blank">
        #                     <img src="https://img.shields.io/badge/arXiv%20paper-2411.17465-b31b1b.svg" alt="arXiv"/>
        #                 </a>
        #                 <a href="https://github.com/showlab/ShowUI" target="_blank">
        #                     <img src="https://img.shields.io/badge/GitHub-ShowUI-black" alt="GitHub"/>
        #                 </a>
        #             </div>
        #         </div>
        #         """
        #     )

        with gr.Row():
            with gr.Column(scale=3):
                imagebox = gr.Image(type="numpy", label="Input Screenshot", placeholder="""#Try UGround with screenshots!

                
                Windows:  [Win + Shift + S]
                macOS:  [Command + Shift + 3]
                
                Then upload/paste from clipboard 🤗
                """)
                
                # Add a slider for iteration count
                iteration_slider = gr.Slider(minimum=1, maximum=3, step=1, value=1, label="Refinement Steps")

                textbox = gr.Textbox(
                    show_label=True,
                    placeholder="Enter a query (e.g., 'Click Nahant')",
                    label="Query",
                )
                submit_btn = gr.Button(value="Submit", variant="primary")

                # Examples component
                gr.Examples(
                    examples=[[e[0], e[1]] for e in examples],
                    inputs=[imagebox, textbox],
                    outputs=[textbox],  # Only update the query textbox
                    examples_per_page=3,
                )

                # Add a hidden dropdown to pass the `is_example` flag
                is_example_dropdown = gr.Dropdown(
                    choices=["True", "False"],
                    value="False",
                    visible=False,
                    label="Is Example Image",
                )

                def set_is_example(query):
                    # Find the example and return its `is_example` flag
                    for _, example_query, is_example in examples:
                        if query.strip() == example_query.strip():
                            return str(is_example)  # Return as string for Dropdown compatibility
                    return "False"

                textbox.change(
                    set_is_example,
                    inputs=[textbox],
                    outputs=[is_example_dropdown],
                )

            with gr.Column(scale=8):
                output_gallery = gr.Gallery(label="Iterative Refinement", object_fit="contain", preview=True)
                # output_gallery = gr.Gallery(label="Iterative Refinement")
                gr.HTML(
                    """
                    <p><strong>Note:</strong> The <span style="color: red;">red point</span> on the output image represents the predicted clickable coordinates.</p>
                    """
                )
                output_coords = gr.Textbox(label="Final Clickable Coordinates")

                gr.HTML(
                    """
                    <p><strong>🤔 Good or bad? Rate your experience to help us improve! ⬇️</strong></p>
                    """
                )
                with gr.Row(elem_id="action-buttons", equal_height=True):
                    upvote_btn = gr.Button(value="👍 Looks good!", variant="secondary")
                    downvote_btn = gr.Button(value="👎 Too bad!", variant="secondary")
                    clear_btn = gr.Button(value="🗑️ Clear", interactive=True)

            def on_submit(image, query, iterations, is_example_image):
                if image is None:
                    raise ValueError("No image provided. Please upload an image before submitting.")
                
                session_id = datetime.now().strftime("%Y%m%d_%H%M%S")
                
                images_during_iterations, click_coords = run_showui(image, query, session_id, iterations)
                
                save_and_upload_data(images_during_iterations[0], query, session_id, is_example_image)
                
                return images_during_iterations, click_coords, session_id

            submit_btn.click(
                on_submit,
                [imagebox, textbox, iteration_slider, is_example_dropdown],
                [output_gallery, output_coords, state_session_id],
            )

            clear_btn.click(
                lambda: (None, None, None, None),
                inputs=None,
                outputs=[imagebox, textbox, output_gallery, output_coords, state_session_id],
                queue=False
            )

            upvote_btn.click(
                lambda session_id, is_example_image: update_vote("upvote", session_id, is_example_image),
                inputs=[state_session_id, is_example_dropdown],
                outputs=[],
                queue=False
            )

            downvote_btn.click(
                lambda session_id, is_example_image: update_vote("downvote", session_id, is_example_image),
                inputs=[state_session_id, is_example_dropdown],
                outputs=[],
                queue=False
            )

    return demo

if __name__ == "__main__":
    demo = build_demo(embed_mode=False)
    demo.queue(api_open=False).launch(
        server_name="0.0.0.0",
        server_port=7860,
        ssr_mode=False,
        debug=True,
    )