Update quantum_inference.py

quantum_inference.py

@@ -1,234 +1 @@
-#!/usr/bin/env python3
-"""
-Hypnos-i3-1.5B - Quantum-Classical Hybrid Model
-================================================
-Sentiment Analysis с квантовым ядром
 
-Версия: 1.0
-Релиз: December 2024
-"""
-
-import numpy as np
-import json
-import torch
-from transformers import AutoModel, AutoTokenizer
-from sklearn.decomposition import PCA
-import time
-
-print("="*70)
-print("🌙 HYPNOS-i3-1.5B - QUANTUM-CLASSICAL MODEL")
-print("="*70)
-print("Version: 1.0")
-print("Type: Quantum Kernel-Enhanced Sentiment Analysis")
-print("Base: VibeThinker-1.5B + 2-qubit Quantum Kernel\n")
-
-# Проверка наличия модели
-try:
-    with open('hypnos_i3_results.json', 'r') as f:
-        results = json.load(f)
-    
-    K_train = np.load('K_train_quantum.npy')
-    K_test = np.load('K_test_quantum.npy')
-    
-    print("✅ Модель загружена из кэша")
-    print(f"📊 Точность (quantum): {results['accuracy_quantum']:.1%}")
-    print(f"📊 Точность (classical): {results['accuracy_baseline']:.1%}")
-    
-except FileNotFoundError:
-    print("⚠️  Модель не найдена. Запустите обучение:")
-    print("   python3 hypnos_i3.py")
-    exit(1)
-
-# Загрузка VibeThinker
-print("\n🔄 Загрузка базовой модели VibeThinker-1.5B...")
-device = torch.device("mps" if torch.backends.mps.is_available() else "cpu")
-
-tokenizer = AutoTokenizer.from_pretrained("WeiboAI/VibeThinker-1.5B")
-model = AutoModel.from_pretrained(
-    "WeiboAI/VibeThinker-1.5B",
-    torch_dtype=torch.float16
-).to(device).eval()
-
-print(f"✅ VibeThinker загружен на {device}")
-
-# Обучающие данные (для квантового ядра)
-TRAIN_DATA = [
-    ("I absolutely love quantum computing! It's amazing!", 1),
-    ("This is the worst experience ever, terrible.", 0),
-    ("Quantum neural networks are fascinating and powerful.", 1),
-    ("I hate bugs in my code, so frustrating!", 0),
-    ("The future of AI is quantum, incredible potential!", 1),
-    ("This product is garbage, waste of money.", 0),
-    ("Machine learning combined with quantum is brilliant!", 1),
-    ("Awful customer service, never coming back.", 0)
-]
-
-print(f"\n📚 База знаний: {len(TRAIN_DATA)} примеров")
-
-# Функция предсказания
-def predict(text, verbose=True):
-    """
-    Предсказывает sentiment текста
-    
-    Pipeline:
-    1. VibeThinker embeddings (1536D)
-    2. Normalization (важно!)
-    3. Quantum kernel similarity
-    4. Classification
-    """
-    
-    if verbose:
-        print(f"\n{'='*70}")
-        print(f"🔮 Анализ текста")
-        print(f"{'='*70}")
-        print(f"📝 Входной текст: '{text}'")
-    
-    start = time.time()
-    
-    # Шаг 1: Embedding через VibeThinker
-    inputs = tokenizer(
-        text, 
-        return_tensors="pt", 
-        padding=True, 
-        truncation=True,
-        max_length=128
-    ).to(device)
-    
-    with torch.no_grad():
-        outputs = model(**inputs)
-        embedding = outputs.last_hidden_state.mean(dim=1).cpu().numpy()[0]
-    
-    # ИСПРАВЛЕНИЕ: Нормализация L2
-    from sklearn.preprocessing import normalize
-    embedding = normalize([embedding])[0]
-    
-    if verbose:
-        print(f"   1️⃣ VibeThinker embedding: {len(embedding)}D (normalized)")
-    
-    # Шаг 2: Вычисляем embeddings для обучающих данных
-    train_embeddings = []
-    train_labels = []
-    
-    for train_text, label in TRAIN_DATA:
-        t_inputs = tokenizer(
-            train_text,
-            return_tensors="pt",
-            padding=True,
-            truncation=True,
-            max_length=128
-        ).to(device)
-        
-        with torch.no_grad():
-            t_outputs = model(**t_inputs)
-            t_emb = t_outputs.last_hidden_state.mean(dim=1).cpu().numpy()[0]
-            # ИСПРАВЛЕНИЕ: Нормализация
-            t_emb = normalize([t_emb])[0]
-            train_embeddings.append(t_emb)
-            train_labels.append(label)
-    
-    # Шаг 3: Косинусное сходство (безопасное)
-    from sklearn.metrics.pairwise import cosine_similarity
-    
-    # ИСПРАВЛЕНИЕ: Клиппинг для избежания NaN
-    similarities = cosine_similarity([embedding], train_embeddings)[0]
-    similarities = np.clip(similarities, -1.0, 1.0)  # Ограничиваем
-    
-    if verbose:
-        print(f"   2️⃣ Квантовая похожесть вычислена")
-    
-    # Шаг 4: Взвешенное голосование с проверкой
-    positive_scores = []
-    negative_scores = []
-    
-    for i, sim in enumerate(similarities):
-        if np.isnan(sim):  # ИСПРАВЛЕНИЕ: Проверка NaN
-            sim = 0.0
-        
-        if train_labels[i] == 1:
-            positive_scores.append(sim)
-        else:
-            negative_scores.append(sim)
-    
-    # ИСПРАВЛЕНИЕ: Используем среднее вместо суммы
-    positive_avg = np.mean(positive_scores) if positive_scores else 0
-    negative_avg = np.mean(negative_scores) if negative_scores else 0
-    
-    # ИСПРАВЛЕНИЕ: Порог для нейтральных текстов
-    diff = positive_avg - negative_avg
-    
-    if abs(diff) < 0.05:  # Очень маленькая разница = нейтральный
-        prediction = -1  # Нейтральный
-        confidence = 0.0
-        sentiment = "😐 NEUTRAL"
-    elif positive_avg > negative_avg:
-        prediction = 1
-        confidence = abs(diff)
-        sentiment = "😊 POSITIVE"
-    else:
-        prediction = 0
-        confidence = abs(diff)
-        sentiment = "😞 NEGATIVE"
-    
-    elapsed = time.time() - start
-    
-    if verbose:
-        print(f"   3️⃣ Классификация: {sentiment}")
-        print(f"   💯 Уверенность: {confidence*100:.1f}%")
-        print(f"   📊 Positive avg: {positive_avg:.3f}, Negative avg: {negative_avg:.3f}")
-        print(f"   ⏱️  Время: {elapsed:.2f}s")
-        print(f"{'='*70}")
-    
-    return {
-        'prediction': prediction,
-        'sentiment': sentiment,
-        'confidence': confidence,
-        'time': elapsed,
-        'scores': {
-            'positive': float(positive_avg),
-            'negative': float(negative_avg)
-        }
-    }
-# ДЕМО
-print("\n" + "="*70)
-print("🧪 ДЕМОНСТРАЦИЯ")
-print("="*70)
-
-demo_texts = [
-    "Quantum computing will revolutionize everything!",
-    "This is absolutely horrible, terrible quality.",
-    "I'm amazed by the quantum algorithms!",
-    "Worst purchase ever, complete waste."
-]
-
-print("\nТестирование на новых примерах:\n")
-
-for text in demo_texts:
-    result = predict(text, verbose=False)
-    print(f"{result['sentiment']:<15} ({result['confidence']:>4.0%}) | {text[:50]}")
-
-# ИНТЕРАКТИВНЫЙ РЕЖИМ
-print("\n" + "="*70)
-print("💬 ИНТЕРАКТИВНЫЙ РЕЖИМ")
-print("="*70)
-print("Введите текст для анализа (или 'exit' для выхода)\n")
-
-while True:
-    try:
-        user_input = input("📝 Текст: ")
-        
-        if user_input.lower() in ['exit', 'quit', 'q']:
-            print("\n👋 Завершение Hypnos-i3-1.5B")
-            break
-        
-        if user_input.strip():
-            predict(user_input)
-    
-    except KeyboardInterrupt:
-        print("\n\n👋 Завершение Hypnos-i3-1.5B")
-        break
-    except Exception as e:
-        print(f"❌ Ошибка: {e}")
-
-print("\n" + "="*70)
-print("✨ Спасибо за использование Hypnos-i3-1.5B!")
-print("="*70)