import spaces
import torch
import re
import os
import gradio as gr
from threading import Thread
from transformers import (
    TextIteratorStreamer,
    AutoTokenizer,
    AutoModelForCausalLM,
    StaticCache,
)
from PIL import ImageDraw
from torchvision.transforms.v2 import Resize

import subprocess

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

auth_token = os.environ.get("TOKEN_FROM_SECRET") or True
tokenizer = AutoTokenizer.from_pretrained("vikhyatk/moondream2")
moondream = AutoModelForCausalLM.from_pretrained(
    "vikhyatk/moondream-next",
    trust_remote_code=True,
    torch_dtype=torch.float16,
    device_map={"": "cuda"},
    attn_implementation="flash_attention_2",
    token=auth_token,
)
moondream.eval()


@spaces.GPU(duration=10)
def answer_question(img, prompt):
    if img is None:
        yield ""

    image_embeds = moondream.encode_image(img)
    streamer = TextIteratorStreamer(tokenizer, skip_special_tokens=True)
    thread = Thread(
        target=moondream.answer_question,
        kwargs={
            "image_embeds": image_embeds,
            "question": prompt,
            "tokenizer": tokenizer,
            "streamer": streamer,
        },
    )
    thread.start()

    buffer = ""
    for new_text in streamer:
        buffer += new_text
        yield buffer.strip()


@spaces.GPU(duration=10)
def caption(img, mode):
    if img is None:
        yield ""

    streamer = TextIteratorStreamer(tokenizer, skip_special_tokens=True)
    thread = Thread(
        target=moondream.caption,
        kwargs={
            "images": [img],
            "length": "short" if mode == "Short" else None,
            "tokenizer": tokenizer,
            "streamer": streamer,
        },
    )
    thread.start()

    buffer = ""
    for new_text in streamer:
        buffer += new_text
        yield buffer.strip()


def extract_floats(text):
    # Regular expression to match an array of four floating point numbers
    pattern = r"\[\s*(-?\d+\.\d+)\s*,\s*(-?\d+\.\d+)\s*,\s*(-?\d+\.\d+)\s*,\s*(-?\d+\.\d+)\s*\]"
    match = re.search(pattern, text)
    if match:
        # Extract the numbers and convert them to floats
        return [float(num) for num in match.groups()]
    return None  # Return None if no match is found


def extract_bbox(text):
    bbox = None
    if extract_floats(text) is not None:
        x1, y1, x2, y2 = extract_floats(text)
        bbox = (x1, y1, x2, y2)
    return bbox


def process_answer(img, answer):
    if extract_bbox(answer) is not None:
        x1, y1, x2, y2 = extract_bbox(answer)
        draw_image = Resize(768)(img)
        width, height = draw_image.size
        x1, x2 = int(x1 * width), int(x2 * width)
        y1, y2 = int(y1 * height), int(y2 * height)
        bbox = (x1, y1, x2, y2)
        ImageDraw.Draw(draw_image).rectangle(bbox, outline="red", width=3)
        return gr.update(visible=True, value=draw_image)

    return gr.update(visible=False, value=None)


with gr.Blocks(title="moondream vl (new)") as demo:
    gr.HTML(
        """
        <style type="text/css">
            .output-text span p { font-size: 1.4rem !important; }
        </style>
        """
    )
    gr.Markdown(
        """
        # 🌔 moondream vl (new)
        A tiny vision language model. [GitHub](https://github.com/vikhyat/moondream)
        """
    )
    with gr.Row():
        with gr.Column():
            mode_radio = gr.Radio(
                ["Caption", "Query", "Detect"],
                show_label=False,
                value=lambda: "Caption",
            )

            @gr.render(inputs=[mode_radio])
            def show_inputs(mode):
                if mode == "Query":
                    with gr.Group():
                        with gr.Row():
                            prompt = gr.Textbox(
                                label="Input",
                                value="How many people are in this image?",
                                scale=4,
                            )
                            submit = gr.Button("Submit")
                        img = gr.Image(type="pil", label="Upload an Image")
                    submit.click(answer_question, [img, prompt], output)
                    prompt.submit(answer_question, [img, prompt], output)
                    img.change(answer_question, [img, prompt], output)
                elif mode == "Caption":
                    with gr.Group():
                        caption_mode = gr.Radio(
                            ["Short", "Normal"],
                            show_label=False,
                            value=lambda: "Normal",
                        )
                        img = gr.Image(type="pil", label="Upload an Image")
                    caption_mode.change(caption, [img, caption_mode], output)
                    img.change(caption, [img, caption_mode], output)
                else:
                    gr.Markdown("Coming soon!")

        with gr.Column():
            output = gr.Markdown(label="Response", elem_classes=["output-text"])
            ann = gr.Image(visible=False, label="Annotated Image")


demo.queue().launch()