app.py

import torch
import clip
from PIL import Image
import gradio as gr

device = "cuda" if torch.cuda.is_available() else "cpu"
model, preprocess = clip.load("ViT-B/32", device=device)

def hotornot(image, gender):
    image = Image.fromarray(image.astype("uint8"), "RGB")
    image = preprocess(image).unsqueeze(0).to(device)
    positive_terms = [f'a hot {gender}', f'a beautiful {gender}', f'an attractive {gender}']
    negative_terms = [f'a gross {gender}', f'an ugly {gender}', f'a hideous {gender}']

    pairs = list(zip(positive_terms, negative_terms))

    def evaluate(terms):
        text = clip.tokenize(terms).to(device)

        with torch.no_grad():
            logits_per_image, logits_per_text = model(image, text)
            probs = logits_per_image.softmax(dim=-1).cpu().numpy()
            return probs[0]

    probs = [evaluate(pair) for pair in pairs]
    
    positive_probs = [prob[0] for prob in probs]
    negative_probs = [prob[1] for prob in probs]

    print("+:", positive_probs)
    print("-:", negative_probs)

    hot_score = sum(positive_probs)/len(positive_probs)
    ugly_score = sum(negative_probs)/len(negative_probs)
    print(hot_score, ugly_score)
    composite = ((hot_score - ugly_score)+1) * 50
    composite = round(composite, 2)
    return composite

iface = gr.Interface(
    fn=hotornot,
    inputs=[
        gr.inputs.Image(label="Image"),
        gr.inputs.Dropdown(
            [
                'person', 'man', 'woman'
            ],
            default='person',
        )
    ],
    outputs="number",
    title="Hot or Not",
    description="A simple hot or not app using OpenAI's CLIP model.",
)
iface.launch()