app.py

"""
🚀 PERI BERT Classifier - FastAPI Backend para HuggingFace Space

API REST para clasificación de reflexiones éticas sobre IA usando BERT fine-tuneado.
Soporta predicción con MC Dropout para uncertainty quantification.

Endpoints:
- POST /predict - Clasificar una reflexión
- POST /predict-batch - Clasificar múltiples reflexiones
- GET /health - Health check
- GET /info - Información del modelo
"""

from fastapi import FastAPI, HTTPException
from fastapi.middleware.cors import CORSMiddleware
from pydantic import BaseModel, Field
from typing import List, Optional, Dict, Any
import torch
from transformers import BertTokenizer, BertForSequenceClassification
import numpy as np
from pathlib import Path
import time
import logging

# Configurar logging
logging.basicConfig(level=logging.INFO)
logger = logging.getLogger(__name__)

# ============================================================================
# CONFIGURACIÓN
# ============================================================================

# Mapeo de arquetipos
ARCHETYPE_LABELS = {
    0: "TECNOCRATA_OPTIMIZADOR",
    1: "HUMANISTA_CRITICO",
    2: "PRAGMATICO_EQUILIBRADO",
    3: "VISIONARIO_ADAPTATIVO",
    4: "ESCEPTICO_CONSERVADOR",
}

ARCHETYPE_NAMES = {
    0: "Tecnócrata Optimizador",
    1: "Humanista Crítico",
    2: "Pragmático Equilibrado",
    3: "Visionario Adaptativo",
    4: "Escéptico Conservador",
}

ARCHETYPE_DESCRIPTIONS = {
    0: "Confía en la eficiencia y objetividad de los sistemas automatizados",
    1: "Prioriza el bienestar humano y cuestiona activamente los sesgos tecnológicos",
    2: "Busca balance entre innovación tecnológica y consideraciones humanas",
    3: "Abraza la transformación tecnológica con enfoque adaptativo y progresista",
    4: "Mantiene una postura cautelosa y crítica hacia la adopción de IA",
}

# Device configuration
DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
MAX_LENGTH = 512
MC_SAMPLES = 10  # Número de muestras para MC Dropout

# ============================================================================
# MODELOS PYDANTIC
# ============================================================================

class ReflectionInput(BaseModel):
    """Input para clasificación individual"""
    text: str = Field(..., min_length=100, max_length=5000, description="Reflexión ética sobre IA")
    use_mc_dropout: bool = Field(default=True, description="Usar MC Dropout para uncertainty")


class BatchReflectionInput(BaseModel):
    """Input para clasificación en batch"""
    texts: List[str] = Field(..., max_items=50, description="Lista de reflexiones (máx 50)")
    use_mc_dropout: bool = Field(default=True, description="Usar MC Dropout para uncertainty")


class ArchetypeResult(BaseModel):
    """Resultado de clasificación"""
    id: str
    name: str
    description: str


class PredictionResponse(BaseModel):
    """Respuesta de predicción individual"""
    archetype: ArchetypeResult
    confidence: float = Field(..., ge=0.0, le=1.0, description="Confianza de la predicción")
    uncertainty: Optional[float] = Field(None, ge=0.0, description="Incertidumbre (MC Dropout)")
    top3_predictions: List[Dict[str, Any]] = Field(..., description="Top 3 predicciones")
    inference_time_ms: float = Field(..., description="Tiempo de inferencia en milisegundos")
    method: str = Field(default="bert", description="Método de clasificación")


class BatchPredictionResponse(BaseModel):
    """Respuesta de predicción en batch"""
    predictions: List[PredictionResponse]
    total_inference_time_ms: float


class HealthResponse(BaseModel):
    """Health check response"""
    model_config = {"protected_namespaces": ()}

    status: str
    model_loaded: bool
    device: str
    timestamp: float


class InfoResponse(BaseModel):
    """Información del modelo"""
    model_config = {"protected_namespaces": ()}

    model_name: str
    num_classes: int
    max_length: int
    device: str
    mc_dropout_samples: int
    archetypes: List[Dict[str, str]]


# ============================================================================
# CARGA DEL MODELO
# ============================================================================

class BERTClassifier:
    """Wrapper para el modelo BERT con MC Dropout"""

    def __init__(self, model_path: str):
        logger.info(f"Cargando modelo desde {model_path}...")
        self.tokenizer = BertTokenizer.from_pretrained(model_path)
        self.model = BertForSequenceClassification.from_pretrained(model_path)
        self.model.to(DEVICE)
        self.model.eval()
        logger.info(f"Modelo cargado exitosamente en {DEVICE}")

    def predict(
        self,
        text: str,
        use_mc_dropout: bool = True
    ) -> Dict[str, Any]:
        """
        Realizar predicción con o sin MC Dropout

        Returns:
            dict con keys: predicted_class, confidence, uncertainty, all_probabilities
        """
        start_time = time.time()

        # Tokenizar
        encoding = self.tokenizer(
            text,
            max_length=MAX_LENGTH,
            padding="max_length",
            truncation=True,
            return_tensors="pt"
        )

        input_ids = encoding["input_ids"].to(DEVICE)
        attention_mask = encoding["attention_mask"].to(DEVICE)

        if use_mc_dropout:
            # MC Dropout: múltiples predicciones con dropout activado
            self.model.train()  # Activar dropout
            all_probs = []

            with torch.no_grad():
                for _ in range(MC_SAMPLES):
                    outputs = self.model(
                        input_ids=input_ids,
                        attention_mask=attention_mask
                    )
                    logits = outputs.logits
                    probs = torch.softmax(logits, dim=1).cpu().numpy()[0]
                    all_probs.append(probs)

            # Calcular estadísticas
            all_probs = np.array(all_probs)  # (MC_SAMPLES, num_classes)
            mean_probs = np.mean(all_probs, axis=0)
            predicted_class = int(np.argmax(mean_probs))
            confidence = float(mean_probs[predicted_class])

            # Calcular incertidumbre (entropía)
            epsilon = 1e-10
            uncertainty = float(-np.sum(mean_probs * np.log(mean_probs + epsilon)))

            self.model.eval()  # Volver a modo evaluación

        else:
            # Predicción estándar sin MC Dropout
            with torch.no_grad():
                outputs = self.model(
                    input_ids=input_ids,
                    attention_mask=attention_mask
                )

            logits = outputs.logits
            probs = torch.softmax(logits, dim=1).cpu().numpy()[0]
            mean_probs = probs
            predicted_class = int(np.argmax(probs))
            confidence = float(probs[predicted_class])
            uncertainty = None
            all_probs = probs.reshape(1, -1)

        # Top 3 predicciones
        top3_indices = np.argsort(mean_probs)[-3:][::-1]
        top3 = [
            {
                "archetype_id": ARCHETYPE_LABELS[int(idx)],
                "archetype_name": ARCHETYPE_NAMES[int(idx)],
                "probability": float(mean_probs[idx])
            }
            for idx in top3_indices
        ]

        inference_time = (time.time() - start_time) * 1000  # ms

        return {
            "predicted_class": predicted_class,
            "confidence": confidence,
            "uncertainty": uncertainty,
            "top3": top3,
            "inference_time_ms": inference_time,
            "all_probabilities": mean_probs.tolist()
        }


# Inicializar modelo global
classifier: Optional[BERTClassifier] = None


def load_model():
    """Cargar modelo al iniciar la aplicación"""
    global classifier

    # En HuggingFace Space, el modelo estará en ./model/
    # Localmente, usar path relativo
    model_paths = [
        Path("./model"),  # HF Space
        Path("../../../models/peri-bert/best_model"),  # Local
    ]

    model_path = None
    for path in model_paths:
        if path.exists():
            model_path = str(path)
            break

    if model_path is None:
        logger.error("No se encontró el modelo. Asegúrate de subirlo a HuggingFace Space.")
        raise RuntimeError("Model not found")

    classifier = BERTClassifier(model_path)


# ============================================================================
# FASTAPI APP
# ============================================================================

app = FastAPI(
    title="PERI BERT Classifier API",
    description="API REST para clasificación de arquetipos éticos en reflexiones sobre IA",
    version="1.0.0",
    docs_url="/",  # Swagger UI en la raíz
)

# CORS middleware
app.add_middleware(
    CORSMiddleware,
    allow_origins=["*"],  # En producción, especificar dominios permitidos
    allow_credentials=True,
    allow_methods=["*"],
    allow_headers=["*"],
)


@app.on_event("startup")
async def startup_event():
    """Cargar modelo al iniciar"""
    load_model()


@app.get("/health", response_model=HealthResponse)
async def health_check():
    """Health check endpoint"""
    return HealthResponse(
        status="healthy",
        model_loaded=classifier is not None,
        device=DEVICE,
        timestamp=time.time()
    )


@app.get("/info", response_model=InfoResponse)
async def model_info():
    """Información del modelo"""
    if classifier is None:
        raise HTTPException(status_code=503, detail="Model not loaded")

    archetypes = [
        {
            "id": ARCHETYPE_LABELS[i],
            "name": ARCHETYPE_NAMES[i],
            "description": ARCHETYPE_DESCRIPTIONS[i]
        }
        for i in range(5)
    ]

    return InfoResponse(
        model_name="bert-base-multilingual-cased (fine-tuned)",
        num_classes=5,
        max_length=MAX_LENGTH,
        device=DEVICE,
        mc_dropout_samples=MC_SAMPLES,
        archetypes=archetypes
    )


@app.post("/predict", response_model=PredictionResponse)
async def predict(input_data: ReflectionInput):
    """
    Clasificar una reflexión individual

    Args:
        input_data: Reflexión y configuración

    Returns:
        Predicción con arquetipo, confianza y métricas
    """
    if classifier is None:
        raise HTTPException(status_code=503, detail="Model not loaded")

    try:
        result = classifier.predict(
            text=input_data.text,
            use_mc_dropout=input_data.use_mc_dropout
        )

        archetype_result = ArchetypeResult(
            id=ARCHETYPE_LABELS[result["predicted_class"]],
            name=ARCHETYPE_NAMES[result["predicted_class"]],
            description=ARCHETYPE_DESCRIPTIONS[result["predicted_class"]]
        )

        return PredictionResponse(
            archetype=archetype_result,
            confidence=result["confidence"],
            uncertainty=result["uncertainty"],
            top3_predictions=result["top3"],
            inference_time_ms=result["inference_time_ms"],
            method="bert-mc-dropout" if input_data.use_mc_dropout else "bert"
        )

    except Exception as e:
        logger.error(f"Error en predicción: {str(e)}")
        raise HTTPException(status_code=500, detail=f"Prediction error: {str(e)}")


@app.post("/predict-batch", response_model=BatchPredictionResponse)
async def predict_batch(input_data: BatchReflectionInput):
    """
    Clasificar múltiples reflexiones en batch

    Args:
        input_data: Lista de reflexiones

    Returns:
        Lista de predicciones
    """
    if classifier is None:
        raise HTTPException(status_code=503, detail="Model not loaded")

    if len(input_data.texts) == 0:
        raise HTTPException(status_code=400, detail="Empty texts list")

    start_time = time.time()
    predictions = []

    try:
        for text in input_data.texts:
            if len(text) < 100:
                continue  # Skip textos muy cortos

            result = classifier.predict(
                text=text,
                use_mc_dropout=input_data.use_mc_dropout
            )

            archetype_result = ArchetypeResult(
                id=ARCHETYPE_LABELS[result["predicted_class"]],
                name=ARCHETYPE_NAMES[result["predicted_class"]],
                description=ARCHETYPE_DESCRIPTIONS[result["predicted_class"]]
            )

            predictions.append(
                PredictionResponse(
                    archetype=archetype_result,
                    confidence=result["confidence"],
                    uncertainty=result["uncertainty"],
                    top3_predictions=result["top3"],
                    inference_time_ms=result["inference_time_ms"],
                    method="bert-mc-dropout" if input_data.use_mc_dropout else "bert"
                )
            )

        total_time = (time.time() - start_time) * 1000

        return BatchPredictionResponse(
            predictions=predictions,
            total_inference_time_ms=total_time
        )

    except Exception as e:
        logger.error(f"Error en batch prediction: {str(e)}")
        raise HTTPException(status_code=500, detail=f"Batch prediction error: {str(e)}")


# ============================================================================
# MAIN (para testing local)
# ============================================================================

if __name__ == "__main__":
    import uvicorn

    uvicorn.run(
        "app:app",
        host="0.0.0.0",
        port=7860,  # Puerto estándar de HuggingFace Spaces
        reload=True
    )