File size: 3,145 Bytes
fda8dae
5781b89
 
 
 
 
07d11bb
 
5781b89
1322687
 
5781b89
 
6d0cb8a
12e7969
5781b89
60e7a28
1322687
 
bac7d5d
5781b89
 
 
fda8dae
5781b89
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
6b26249
 
5781b89
07d11bb
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5781b89
 
 
 
 
 
 
 
 
1322687
5781b89
 
 
07d11bb
 
 
 
5781b89
 
07d11bb
5781b89
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
import spaces
import torch
import re
import gradio as gr
from threading import Thread
from transformers import TextIteratorStreamer, AutoTokenizer, AutoModelForCausalLM
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)

model_id = "vikhyatk/moondream2"
revision = "2024-05-20"
tokenizer = AutoTokenizer.from_pretrained(model_id, revision=revision)
moondream = AutoModelForCausalLM.from_pretrained(
    model_id, trust_remote_code=True, revision=revision,
    torch_dtype=torch.bfloat16, device_map={"": "cuda"},
    attn_implementation="flash_attention_2"
)
moondream.eval()


@spaces.GPU(duration=10)
def answer_question(img, prompt):
    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()

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() as demo:
    gr.Markdown(
        """
        # 🌔 moondream2
        A tiny vision language model. [GitHub](https://github.com/vikhyat/moondream)
        """
    )
    with gr.Row():
        prompt = gr.Textbox(label="Input", value="Describe this image.", scale=4)
        submit = gr.Button("Submit")
    with gr.Row():
        img = gr.Image(type="pil", label="Upload an Image")
        with gr.Column():
            output = gr.Markdown(label="Response")
            ann = gr.Image(visible=False, label="Annotated Image")

    submit.click(answer_question, [img, prompt], output)
    prompt.submit(answer_question, [img, prompt], output)
    output.change(process_answer, [img, output], ann, show_progress=False)

demo.queue().launch()