handler.py

import torch
import torch.nn.functional as F
import torchvision.transforms as transforms
from PIL import Image
import numpy as np
import pandas as pd
from pathlib import Path
import json
from io import BytesIO
import boto3
from botocore.exceptions import ClientError
from huggingface_hub import hf_hub_download

# Imports desde el código de entrenamiento actualizado
from redes_prototipicas_tvt5 import ConvEncoder, PrototypicalNetwork, SmartPadResize

def load_image_from_s3_url(s3_url, s3_client):
    """Cargar imagen desde S3 extrayendo bucket y key de la URL"""
    try:
        url_parts = s3_url.replace('https://', '').split('/')
        bucket = url_parts[0].split('.s3.amazonaws.com')[0]
        key = '/'.join(url_parts[1:])
        
        response = s3_client.get_object(Bucket=bucket, Key=key)
        image_data = response['Body'].read()
        return Image.open(BytesIO(image_data)).convert('RGB')
    except Exception as e:
        print(f"❌ Error cargando imagen: {e}")
        return None

def model_selector(self, model_category):
    """Seleccionar modelo según categoría"""
    models = {
        182: (self.encoder_detergentes, self.class_names_detergentes, self.prototypes_detergentes, self.eval_transform_detergentes),
        175: (self.encoder_mascotas, self.class_names_mascotas, self.prototypes_mascotas, self.eval_transform_mascotas),
        202: (self.encoder_vinos, self.class_names_vinos, self.prototypes_vinos, self.eval_transform_vinos),
        161: (self.encoder_cecinas, self.class_names_cecinas, self.prototypes_cecinas, self.eval_transform_cecinas),
        198: (self.encoder_licores, self.class_names_licores, self.prototypes_licores, self.eval_transform_licores)
    }
    return models.get(model_category)

def get_ood_thresholds(model_category):
    """Umbrales OOD para modelos 512px"""
    config = {
        182: {'similarity_threshold': 0.70, 'distance_threshold': 0.80},  # detergentes
        175: {'similarity_threshold': 0.68, 'distance_threshold': 0.85},  # mascotas
        202: {'similarity_threshold': 0.72, 'distance_threshold': 0.75},  # vinos
        161: {'similarity_threshold': 0.69, 'distance_threshold': 0.82},  # cecinas
        198: {'similarity_threshold': 0.71, 'distance_threshold': 0.78}   # licores
    }
    return config.get(model_category, {'similarity_threshold': 0.70, 'distance_threshold': 0.80})

def detect_out_of_distribution(query_features, prototypes, ood_config):
    """Detección OOD simplificada"""
    similarities = torch.mm(query_features, prototypes.t()).squeeze(0)
    max_similarity = similarities.max().item()
    
    distances = torch.cdist(query_features, prototypes).squeeze(0)
    min_distance = distances.min().item()
    
    # Criterios OOD
    is_ood = (max_similarity < ood_config['similarity_threshold'] or 
              min_distance > ood_config['distance_threshold'])
    
    # Score combinado
    similarity_score = max_similarity
    distance_score = max(0, (ood_config['distance_threshold'] - min_distance) / ood_config['distance_threshold'])
    ood_score = (0.7 * similarity_score + 0.3 * distance_score)
    
    if is_ood:
        ood_score = max(0, ood_score - 0.05)
    
    return is_ood, ood_score

def load_classification_model_optimized(model_path, device):
    """Cargar modelo 512px únicamente"""
    checkpoint = torch.load(model_path, map_location=device, weights_only=False)
    
    if 'prototypes' not in checkpoint or 'class_names' not in checkpoint:
        raise ValueError("❌ Modelo sin prototipos. Re-entrena con código actualizado.")
    
    # Configuración del modelo
    model_config = checkpoint.get('model_config', {})
    hidden_dim = model_config.get('hidden_dim', 64)
    output_dim = model_config.get('output_dim', 256)
    image_size = model_config.get('image_size', 512)
    
    print(f"📊 Cargando modelo {image_size}px: {len(checkpoint['class_names'])} clases")
    
    # Cargar arquitectura y pesos
    encoder = ConvEncoder(hidden_dim=hidden_dim, output_dim=output_dim).to(device)
    model = PrototypicalNetwork(encoder).to(device)
    
    encoder.load_state_dict(checkpoint['encoder_state_dict'])
    model.load_state_dict(checkpoint['model_state_dict'])
    encoder.eval()
    model.eval()
    
    # Prototipos y clases
    prototypes = checkpoint['prototypes'].to(device)
    class_names = checkpoint['class_names']
    
    # Transformaciones 512px con SmartPadResize
    eval_transform = transforms.Compose([
        SmartPadResize(target_size=image_size, fill_value=128),
        transforms.ToTensor(),
        transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
    ])
    
    return encoder, class_names, prototypes, eval_transform

def load_json_from_s3(json_s3_url):
    """Cargar JSON desde S3"""
    session = boto3.Session(
        aws_access_key_id='AKIA6BH4GPXQCUZ3PAX5',
        aws_secret_access_key='VMcl897FpEeakLb2mzm3Nfi5FJBIDh9on1yhNFGr',
        region_name='us-east-1'
    )
    s3_client = session.client('s3')

    try:
        url_parts = json_s3_url.replace('https://', '').split('/')
        bucket = url_parts[0].split('.s3.amazonaws.com')[0]
        key = '/'.join(url_parts[1:])
        
        response = s3_client.get_object(Bucket=bucket, Key=key)
        json_content = response['Body'].read().decode('utf-8')
        return json.loads(json_content), s3_client
    except Exception as e:
        print(f"❌ Error cargando JSON: {e}")
        return None, None

def classify_saved_bboxes(saved_images, encoder, class_names, prototypes, eval_transform, device, minimal_accuracy, s3_client, model_category):
    """Clasificación con detección OOD"""
    if not saved_images:
        return pd.DataFrame()
    
    print(f"🔄 Clasificando {len(saved_images)} imágenes...")
    
    ood_config = get_ood_thresholds(model_category)
    results = []
    filtered_count = 0
    ood_count = 0
    
    with torch.no_grad():
        for img_info in saved_images:
            try:
                # Cargar y transformar imagen
                bbox_image = load_image_from_s3_url(img_info['bbox_path'], s3_client)
                if bbox_image is None:
                    continue
                
                query_tensor = eval_transform(bbox_image).unsqueeze(0).to(device)
                query_features = F.normalize(encoder(query_tensor), p=2, dim=1)
                
                # Detección OOD
                is_ood, ood_score = detect_out_of_distribution(query_features, prototypes, ood_config)
                
                if is_ood:
                    ood_count += 1
                    filtered_count += 1
                    continue
                
                # Calcular similitudes
                similarities = torch.mm(query_features, prototypes.t()).cpu().numpy()[0]
                top3_indices = np.argsort(similarities)[::-1]
                
                # Filtrar por minimal_accuracy
                predictions = []
                accuracies = []
                
                for idx in top3_indices:
                    if similarities[idx] >= minimal_accuracy:
                        predictions.append(class_names[idx])
                        accuracies.append(round(similarities[idx], 4))
                
                if not predictions:
                    filtered_count += 1
                    continue
                
                # Ajustar con OOD score
                adjusted_accuracies = [round((acc * 0.9) + (ood_score * 0.1), 4) for acc in accuracies]
                
                result = {
                    'sku_bb_id': str(img_info['bbox_id']),
                    'predictions': predictions,
                    'accuracy': adjusted_accuracies,
                    'prediccion_principal': predictions[0],
                    'similarity_principal': f"{adjusted_accuracies[0]*100:.2f}%",
                    'bbox_confidence': round(float(img_info['confidence']), 4),
                    'ood_score': round(ood_score, 4),
                    'xmin': img_info['x_min'],
                    'ymin': img_info['y_min'],
                    'xmax': img_info['x_max'],
                    'ymax': img_info['y_max']
                }
                results.append(result)
                
            except Exception as e:
                print(f"❌ Error en bbox {img_info['bbox_id']}: {e}")
                continue
    
    print(f"📊 Procesadas: {len(results)}, Filtradas: {filtered_count}, OOD: {ood_count}")
    return pd.DataFrame(results)

def process_image_with_bboxes(self, image_url, picture_id, visit_id, minimal_accuracy, model_path, train_path, model_category, json_s3_url):
    """Función principal de procesamiento"""
    print(f"🚀 Procesando imagen con modelo 512px - Categoría: {model_category}")
    
    device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
    
    # Cargar bounding boxes
    saved_images, s3_client = load_json_from_s3(json_s3_url)
    if not saved_images:
        return pd.DataFrame()
    
    saved_images = saved_images['bounding_boxes']
    
    # Seleccionar modelo
    try:
        encoder, class_names, prototypes, eval_transform = model_selector(self, model_category)
    except Exception as e:
        print(f"❌ Error cargando modelo: {e}")
        return pd.DataFrame()
    
    # Clasificar
    results_df = classify_saved_bboxes(
        saved_images, encoder, class_names, prototypes, eval_transform, 
        device, minimal_accuracy, s3_client, model_category
    )
    
    if not results_df.empty:
        print(f"✅ {len(results_df)} detecciones procesadas")
        print(f"📊 Clases detectadas: {', '.join(results_df['prediccion_principal'].unique())}")
    
    return results_df

class EndpointHandler():
    def __init__(self, path=""):
        """Inicialización con modelos 512px únicamente"""
        device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
        print(f"🚀 Inicializando handler con device: {device}")
        
        # Cargar modelo de licores
        model_filename = "model_curriculum4/prototypical_model_best_licores.pth"
        local_model_path = hf_hub_download(repo_id="Drazcat-AI/redes_prototipicas", filename=model_filename)
        
        self.encoder_licores, self.class_names_licores, self.prototypes_licores, self.eval_transform_licores = load_classification_model_optimized(local_model_path, device)
        
        print("✅ Handler inicializado")

    def predict_objects(self, image_url, picture_id, visit_id, minimal_accuracy, model_category, json_s3_url):
        """Predicción con modelos 512px"""
        return process_image_with_bboxes(
            self, image_url, picture_id, visit_id, minimal_accuracy, 
            None, None, model_category, json_s3_url
        )

    def __call__(self, event):
        """Método de llamada principal"""
        if "inputs" not in event:
            return {"statusCode": 400, "body": json.dumps("Error: No 'inputs' parameter.")}
        
        event = event["inputs"]
        
        try:
            predictions = self.predict_objects(
                event["image_url"], event["picture_id"], event["visit_id"],
                event["minimal_accuracy"], event["model_category"], event["json_s3_url"]
            )
            
            return {
                "statusCode": 200,
                "body": json.dumps(predictions.to_json(orient='records'))
            }
        except Exception as e:
            return {"statusCode": 500, "body": json.dumps(f"Error: {str(e)}")}