app.py

import gradio as gr
import PIL.Image
import transformers
from transformers import PaliGemmaForConditionalGeneration, PaliGemmaProcessor
import torch
import os
import string
import functools
import re
import numpy as np
import spaces

# Model IDs
MODEL_IDS = {
    "paligemma-3b-ft-widgetcap-waveui-448": "agentsea/paligemma-3b-ft-widgetcap-waveui-448",
    "paligemma-3b-ft-waveui-896": "agentsea/paligemma-3b-ft-waveui-896"
}
COLORS = ['#4285f4', '#db4437', '#f4b400', '#0f9d58', '#e48ef1']
# Device configuration
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

# Load models and processors
models = {name: PaliGemmaForConditionalGeneration.from_pretrained(model_id).eval().to(device)
          for name, model_id in MODEL_IDS.items()}
processors = {name: PaliGemmaProcessor.from_pretrained(processor_id)
              for name, processor_id in MODEL_IDS.items()}

###### Transformers Inference
@spaces.GPU
def infer(
    image: PIL.Image.Image,
    text: str,
    max_new_tokens: int,
    model_choice: str
) -> str:
    model = models[model_choice]
    processor = processors[model_choice]
    inputs = processor(text=text, images=image, return_tensors="pt").to(device)
    with torch.inference_mode():
        generated_ids = model.generate(
            **inputs,
            max_new_tokens=max_new_tokens,
            do_sample=False
        )
    result = processor.batch_decode(generated_ids, skip_special_tokens=True)
    return result[0][len(text):].lstrip("\n")

def parse_segmentation(input_image, input_text, model_choice):
    out = infer(input_image, input_text, max_new_tokens=100, model_choice=model_choice)
    objs = extract_objs(out.lstrip("\n"), input_image.size[0], input_image.size[1], unique_labels=True)
    labels = set(obj.get('name') for obj in objs if obj.get('name'))
    color_map = {l: COLORS[i % len(COLORS)] for i, l in enumerate(labels)}
    highlighted_text = [(obj['content'], obj.get('name')) for obj in objs]
    annotated_img = (
        input_image,
        [
            (
                obj['mask'] if obj.get('mask') is not None else obj['xyxy'],
                obj['name'] or '',
            )
            for obj in objs
            if 'mask' in obj or 'xyxy' in obj
        ],
    )
    has_annotations = bool(annotated_img[1])
    return annotated_img

######## Demo

INTRO_TEXT = """## PaliGemma WaveUI\n\n

Two fine-tuned models on the [WaveUI dataset](https://huggingface.co/datasets/agentsea/wave-ui) from different bases:\n\n

- [paligemma-3b-ft-widgetcap-waveui-448](https://huggingface.co/agentsea/paligemma-3b-ft-widgetcap-waveui-448)
- [paligemma-3b-ft-waveui-896](https://huggingface.co/agentsea/paligemma-3b-ft-waveui-896)

Note:\n\n

- the task they were fine-tuned on was detection, so it may not generalize to other tasks.

Usage: write the task keyword "detect" before the element you want the model to detect. For example, "detect profile picture".
"""

with gr.Blocks(css="style.css") as demo:
    gr.Markdown(INTRO_TEXT)
    with gr.Tab("Detection"):
        model_choice = gr.Dropdown(label="Select Model", choices=list(MODEL_IDS.keys()))
        image = gr.Image(type="pil")
        seg_input = gr.Text(label="Detect instruction (e.g. 'detect sign in button')")
        seg_btn = gr.Button("Submit")
        annotated_image = gr.AnnotatedImage(label="Output")
        
        examples = [["./airbnb.jpg", "detect 'Amazing pools' button"]]
        gr.Examples(
            examples=examples,
            inputs=[image, seg_input],
        )

        seg_inputs = [
            image,
            seg_input,
            model_choice
        ]
        seg_outputs = [
            annotated_image
        ]
        seg_btn.click(
            fn=parse_segmentation,
            inputs=seg_inputs,
            outputs=seg_outputs,
        )

_SEGMENT_DETECT_RE = re.compile(
    r'(.*?)' +
    r'<loc(\d{4})>' * 4 + r'\s*' +
    '(?:%s)?' % (r'<seg(\d{3})>' * 16) +
    r'\s*([^;<>]+)? ?(?:; )?',
)

def extract_objs(text, width, height, unique_labels=False):
    """Returns objs for a string with "<loc>" and "<seg>" tokens."""
    objs = []
    seen = set()
    while text:
        m = _SEGMENT_DETECT_RE.match(text)
        if not m:
            break
        print("m", m)
        gs = list(m.groups())
        before = gs.pop(0)
        name = gs.pop()
        y1, x1, y2, x2 = [int(x) / 1024 for x in gs[:4]]
        
        y1, x1, y2, x2 = map(round, (y1*height, x1*width, y2*height, x2*width))
        mask = None

        content = m.group()
        if before:
            objs.append(dict(content=before))
            content = content[len(before):]
        while unique_labels and name in seen:
            name = (name or '') + "'"
        seen.add(name)
        objs.append(dict(
            content=content, xyxy=(x1, y1, x2, y2), mask=mask, name=name))
        text = text[len(before) + len(content):]

    if text:
        objs.append(dict(content=text))

    return objs

#########

if __name__ == "__main__":
    demo.queue(max_size=10).launch(debug=True)