app.py

import gradio as gr
import torch
import torch.nn.functional as F
from transformers import AutoImageProcessor, AutoModelForImageClassification
from PIL import Image
import requests
from io import BytesIO
import numpy as np
import os
from pathlib import Path
import tempfile

# 🔥 MODÈLE SPÉCIALISÉ DANS LA MODE
MODEL_NAME = "google/vit-base-patch16-224"  # Modèle fiable et rapide

print("🔄 Chargement du modèle de mode...")

try:
    processor = AutoImageProcessor.from_pretrained(MODEL_NAME)
    model = AutoModelForImageClassification.from_pretrained(MODEL_NAME)
    
    device = "cuda" if torch.cuda.is_available() else "cpu"
    model.to(device)
    model.eval()
    
    print(f"✅ Modèle chargé sur {device}")

except Exception as e:
    print(f"❌ Erreur chargement: {e}")
    processor = None
    model = None

# 🎯 LABELS COMPRÉHENSIBLES POUR LA MODE (adaptés au modèle)
FASHION_LABELS = {
    0: "T-shirt", 1: "Pantalon", 2: "Pull", 3: "Robe", 4: "Manteau",
    5: "Sandale", 6: "Chemise", 7: "Sneaker", 8: "Sac", 9: "Botte",
    10: "Veste", 11: "Jupe", 12: "Short", 13: "Chaussures", 14: "Accessoire"
}

def convert_heic_to_jpeg(image_path):
    """Convertit les HEIC en JPEG si nécessaire"""
    try:
        if isinstance(image_path, str) and image_path.lower().endswith('.heic'):
            # Conversion HEIC → JPEG
            img = Image.open(image_path)
            jpeg_path = image_path.replace('.heic', '.jpeg')
            img.convert('RGB').save(jpeg_path, 'JPEG')
            return jpeg_path
    except:
        pass
    return image_path

def preprocess_image(image):
    """Prétraitement robuste des images"""
    try:
        # Si c'est un chemin de fichier (HEIC)
        if isinstance(image, str):
            image = convert_heic_to_jpeg(image)
            image = Image.open(image)
        
        # Conversion en RGB
        if image.mode != 'RGB':
            image = image.convert('RGB')
        
        # Redimensionnement
        image = image.resize((224, 224), Image.Resampling.LANCZOS)
        
        return image
        
    except Exception as e:
        raise Exception(f"Erreur prétraitement: {str(e)}")

def classify_fashion(image):
    """Classification avec gestion robuste des formats"""
    try:
        if image is None:
            return "❌ Veuillez uploader une image de vêtement"
        
        if processor is None or model is None:
            return "⚠️ Modèle en cours de chargement... Patientez 30s"
        
        # 📸 Gestion spéciale HEIC et formats complexes
        try:
            # Si l'image est un chemin temporaire (format HEIC)
            if isinstance(image, str) and ('gradio' in image or 'tmp' in image):
                if image.lower().endswith('.heic'):
                    # Conversion HEIC → JPEG
                    img = Image.open(image)
                    with tempfile.NamedTemporaryFile(suffix='.jpg', delete=False) as tmp:
                        img.convert('RGB').save(tmp.name, 'JPEG', quality=95)
                        processed_image = Image.open(tmp.name)
                        os.unlink(tmp.name)  # Nettoyage
                else:
                    processed_image = Image.open(image)
            else:
                # Image normale
                processed_image = image
            
            # Conversion en RGB si nécessaire
            if processed_image.mode != 'RGB':
                processed_image = processed_image.convert('RGB')
                
        except Exception as e:
            return f"❌ Format d'image non supporté: {str(e)}\n\n💡 Utilisez JPEG, PNG ou WebP"
        
        # 🔥 PRÉTRAITEMENT CORRECT
        processed_image = processed_image.resize((224, 224), Image.Resampling.LANCZOS)
        
        # Transformation pour le modèle
        inputs = processor(images=processed_image, return_tensors="pt")
        inputs = {k: v.to(device) for k, v in inputs.items()}
        
        # 🔥 INFÉRENCE
        with torch.no_grad():
            outputs = model(**inputs)
        
        # 📊 POST-TRAITEMENT
        probabilities = F.softmax(outputs.logits, dim=-1)
        top_probs, top_indices = torch.topk(probabilities, 5)
        
        # Conversion en résultats
        results = []
        for i in range(len(top_indices[0])):
            label_idx = top_indices[0][i].item()
            label_name = FASHION_LABELS.get(label_idx, f"Catégorie {label_idx}")