app.py

import numpy as np
from PIL import Image
import gradio as gr
from deepface import DeepFace
from datasets import load_dataset
import os
import pickle
from io import BytesIO
from huggingface_hub import upload_file, hf_hub_download, list_repo_files
from pathlib import Path
import gc
import requests
import time
import shutil

# 📁 Parámetros
DATASET_ID = "Segizu/facial-recognition"
EMBEDDINGS_SUBFOLDER = "embeddings"
LOCAL_EMB_DIR = Path("embeddings")
LOCAL_EMB_DIR.mkdir(exist_ok=True)
HF_TOKEN = os.getenv("HF_TOKEN")
headers = {"Authorization": f"Bearer {HF_TOKEN}"} if HF_TOKEN else {}

# 💾 Configuración de control de almacenamiento
MAX_TEMP_STORAGE_GB = 40
UPLOAD_EVERY = 50
embeddings_to_upload = []

def get_folder_size(path):
    total = 0
    for dirpath, _, filenames in os.walk(path):
        for f in filenames:
            fp = os.path.join(dirpath, f)
            total += os.path.getsize(fp)
    return total / (1024 ** 3)  # En GB

def flush_embeddings():
    global embeddings_to_upload
    print("🚀 Subiendo lote de embeddings a Hugging Face...")

    for emb_file in embeddings_to_upload:
        try:
            filename = emb_file.name
            upload_file(
                path_or_fileobj=str(emb_file),
                path_in_repo=f"{EMBEDDINGS_SUBFOLDER}/{filename}",
                repo_id=DATASET_ID,
                repo_type="dataset",
                token=HF_TOKEN
            )
            os.remove(emb_file)
            print(f"✅ Subido y eliminado: {filename}")
            time.sleep(1.2)  # Evita 429
        except Exception as e:
            print(f"❌ Error subiendo {filename}: {e}")
            continue

    embeddings_to_upload = []

# ✅ Cargar CSV desde el dataset
dataset = load_dataset(
    "csv",
    data_files="metadata.csv",
    split="train",
    column_names=["image"],
    header=0
)

print("✅ Validación post-carga")
print(dataset[0])
print("Columnas:", dataset.column_names)

# 🔄 Preprocesamiento
def preprocess_image(img: Image.Image) -> np.ndarray:
    img_rgb = img.convert("RGB")
    img_resized = img_rgb.resize((160, 160), Image.Resampling.LANCZOS)
    return np.array(img_resized)

# 📦 Generar y subir embeddings
def build_database():
    print("🔄 Generando embeddings...")
    batch_size = 10

    for i in range(0, len(dataset), batch_size):
        batch = dataset[i:i + batch_size]
        print(f"📦 Lote {i // batch_size + 1}/{(len(dataset) + batch_size - 1) // batch_size}")

        for j in range(len(batch["image"])):
            item = {"image": batch["image"][j]}
            image_url = item["image"]

            if not isinstance(image_url, str) or not image_url.startswith("http") or image_url.strip().lower() == "image":
                print(f"⚠️ Saltando {i + j} - URL inválida: {image_url}")
                continue

            name = f"image_{i + j}"
            filename = LOCAL_EMB_DIR / f"{name}.pkl"

            # Verificar si ya existe en HF
            try:
                hf_hub_download(
                    repo_id=DATASET_ID,
                    repo_type="dataset",
                    filename=f"{EMBEDDINGS_SUBFOLDER}/{name}.pkl",
                    token=HF_TOKEN
                )
                print(f"⏩ Ya existe remoto: {name}.pkl")
                continue
            except:
                pass

            try:
                response = requests.get(image_url, headers=headers, timeout=10)
                response.raise_for_status()
                img = Image.open(BytesIO(response.content)).convert("RGB")

                img_processed = preprocess_image(img)
                embedding = DeepFace.represent(
                    img_path=img_processed,
                    model_name="Facenet",
                    enforce_detection=False
                )[0]["embedding"]

                # Guardar temporal
                with open(filename, "wb") as f:
                    pickle.dump({"name": name, "img": img, "embedding": embedding}, f)

                embeddings_to_upload.append(filename)

                # Si excede límites, subir batch
                if get_folder_size(LOCAL_EMB_DIR) >= MAX_TEMP_STORAGE_GB or len(embeddings_to_upload) >= UPLOAD_EVERY:
                    flush_embeddings()

                del img_processed
                gc.collect()

            except Exception as e:
                print(f"❌ Error en {name}: {e}")
                continue

    # Subir lo que quede
    if embeddings_to_upload:
        flush_embeddings()

# 🔍 Buscar similitudes desde archivos remotos
def find_similar_faces(uploaded_image: Image.Image):
    try:
        img_processed = preprocess_image(uploaded_image)
        query_embedding = DeepFace.represent(
            img_path=img_processed,
            model_name="Facenet",
            enforce_detection=False
        )[0]["embedding"]
        del img_processed
        gc.collect()
    except Exception as e:
        return [], f"⚠ Error procesando imagen: {str(e)}"

    similarities = []

    try:
        embedding_files = [
            f for f in list_repo_files(DATASET_ID, repo_type="dataset", token=HF_TOKEN)
            if f.startswith(f"{EMBEDDINGS_SUBFOLDER}/") and f.endswith(".pkl")
        ]
    except Exception as e:
        return [], f"⚠ Error obteniendo archivos: {str(e)}"

    for file_path in embedding_files:
        try:
            file_bytes = requests.get(
                f"https://huggingface.co/datasets/{DATASET_ID}/resolve/main/{file_path}",
                headers=headers,
                timeout=10
            ).content
            record = pickle.loads(file_bytes)

            name = record["name"]
            img = record["img"]
            emb = record["embedding"]

            dist = np.linalg.norm(np.array(query_embedding) - np.array(emb))
            sim_score = 1 / (1 + dist)
            similarities.append((sim_score, name, np.array(img)))

        except Exception as e:
            print(f"⚠ Error con {file_path}: {e}")
            continue

    similarities.sort(reverse=True)
    top = similarities[:5]
    gallery = [(img, f"{name} - Similitud: {sim:.2f}") for sim, name, img in top]
    summary = "\n".join([f"{name} - Similitud: {sim:.2f}" for sim, name, _ in top])
    return gallery, summary

# 🚀 Inicializar
print("🚀 Iniciando app...")
build_database()

# 🎛️ Interfaz Gradio
demo = gr.Interface(
    fn=find_similar_faces,
    inputs=gr.Image(label="📤 Sube una imagen", type="pil"),
    outputs=[
        gr.Gallery(label="📸 Rostros similares"),
        gr.Textbox(label="🧠 Detalle", lines=6)
    ],
    title="🔍 Reconocimiento facial con DeepFace",
    description="Sube una imagen y encuentra coincidencias en el dataset privado de Hugging Face usando embeddings Facenet."
)

demo.launch()