app.py

import gradio as gr
import torch
from transformers import AutoModelForObjectDetection, AutoImageProcessor
from PIL import Image, ImageDraw

# Definir el repositorio en Hugging Face
repo_id = "facebook/detr-resnet-101"

# Cargar el modelo en modo FP16 para mayor velocidad en GPU
device = "cuda" if torch.cuda.is_available() else "cpu"
model = AutoModelForObjectDetection.from_pretrained(repo_id).to(device).half()
image_processor = AutoImageProcessor.from_pretrained(repo_id)

# Función para la inferencia
def predict(img):
    img = img.convert("RGB")  # Asegurar formato RGB
    inputs = image_processor(images=img, return_tensors="pt", pin_memory=True).to(device)

    with torch.no_grad():
        outputs = model(**inputs)

    # Procesar los resultados
    target_sizes = torch.tensor([img.size[::-1]], device=device)
    results = image_processor.post_process_object_detection(outputs, threshold=0.5, target_sizes=target_sizes)[0]

    # Dibujar las detecciones en la imagen
    draw = ImageDraw.Draw(img)
    detecciones = []

    for score, label, box in zip(results["scores"], results["labels"], results["boxes"]):
        box = [round(i, 2) for i in box.tolist()]
        x, y, x2, y2 = box
        draw.rectangle([x, y, x2, y2], outline="red", width=3)
        class_name = f"Clase {label.item()} - Confianza: {round(score.item(), 2)}"
        draw.text((x, y), class_name, fill="red")
        detecciones.append(class_name)

    return img, "\n".join(detecciones)

# Crear la interfaz y lanzarla con Gradio
gr.Interface(
    fn=predict,
    inputs=gr.Image(type="pil"),
    outputs=[gr.Image(), gr.Text()],
    examples=['raccoon-133.jpg', 'raccoon-108.jpg'],
    concurrency_limit=2
).launch(share=False)