import warnings import numpy as np import pandas as pd import os import json import random import gradio as gr import torch from sklearn.ensemble import RandomForestClassifier from sklearn.model_selection import train_test_split from sklearn.preprocessing import OneHotEncoder from transformers import AutoModelForSequenceClassification, AutoTokenizer, T5ForConditionalGeneration, T5Tokenizer, pipeline from deap import base, creator, tools, algorithms import gc warnings.filterwarnings('ignore', category=FutureWarning, module='huggingface_hub.file_download') # Initialize Example Emotions Dataset data = { 'context': [ 'I am happy', 'I am sad', 'I am angry', 'I am excited', 'I am calm', 'I am feeling joyful', 'I am grieving', 'I am feeling peaceful', 'I am frustrated', 'I am determined', 'I feel resentment', 'I am feeling glorious', 'I am motivated', 'I am surprised', 'I am fearful', 'I am trusting', 'I feel disgust', 'I am optimistic', 'I am pessimistic', 'I feel bored', 'I am envious' ], 'emotion': [ 'joy', 'sadness', 'anger', 'joy', 'calmness', 'joy', 'grief', 'calmness', 'anger', 'determination', 'resentment', 'glory', 'motivation', 'surprise', 'fear', 'trust', 'disgust', 'optimism', 'pessimism', 'boredom', 'envy' ] } df = pd.DataFrame(data) # Encoding the contexts using One-Hot Encoding (memory-efficient) encoder = OneHotEncoder(handle_unknown='ignore', sparse=True) contexts_encoded = encoder.fit_transform(df[['context']]) # Encoding emotions emotions_target = pd.Categorical(df['emotion']).codes emotion_classes = pd.Categorical(df['emotion']).categories # Load pre-trained BERT model for emotion prediction emotion_prediction_model = AutoModelForSequenceClassification.from_pretrained("bhadresh-savani/distilbert-base-uncased-emotion") emotion_prediction_tokenizer = AutoTokenizer.from_pretrained("bhadresh-savani/distilbert-base-uncased-emotion") # Lazy loading for the fine-tuned language model (FLAN-T5) _finetuned_lm_tokenizer = None _finetuned_lm_model = None def get_finetuned_lm_model(): global _finetuned_lm_tokenizer, _finetuned_lm_model if _finetuned_lm_tokenizer is None or _finetuned_lm_model is None: _finetuned_lm_tokenizer = T5Tokenizer.from_pretrained("google/flan-t5-base") _finetuned_lm_model = T5ForConditionalGeneration.from_pretrained("google/flan-t5-base", device_map="auto", low_cpu_mem_usage=True) return _finetuned_lm_tokenizer, _finetuned_lm_model # Enhanced Emotional States emotions = { 'joy': {'percentage': 10, 'motivation': 'positive', 'intensity': 0}, 'pleasure': {'percentage': 10, 'motivation': 'selfish', 'intensity': 0}, 'sadness': {'percentage': 10, 'motivation': 'negative', 'intensity': 0}, 'grief': {'percentage': 10, 'motivation': 'negative', 'intensity': 0}, 'anger': {'percentage': 10, 'motivation': 'traumatic or strong', 'intensity': 0}, 'calmness': {'percentage': 10, 'motivation': 'neutral', 'intensity': 0}, 'determination': {'percentage': 10, 'motivation': 'positive', 'intensity': 0}, 'resentment': {'percentage': 10, 'motivation': 'negative', 'intensity': 0}, 'glory': {'percentage': 10, 'motivation': 'positive', 'intensity': 0}, 'motivation': {'percentage': 10, 'motivation': 'positive', 'intensity': 0}, 'ideal_state': {'percentage': 100, 'motivation': 'balanced', 'intensity': 0}, 'fear': {'percentage': 10, 'motivation': 'defensive', 'intensity': 0}, 'surprise': {'percentage': 10, 'motivation': 'unexpected', 'intensity': 0}, 'anticipation': {'percentage': 10, 'motivation': 'predictive', 'intensity': 0}, 'trust': {'percentage': 10, 'motivation': 'reliable', 'intensity': 0}, 'disgust': {'percentage': 10, 'motivation': 'repulsive', 'intensity': 0}, 'optimism': {'percentage': 10, 'motivation': 'hopeful', 'intensity': 0}, 'pessimism': {'percentage': 10, 'motivation': 'doubtful', 'intensity': 0}, 'boredom': {'percentage': 10, 'motivation': 'indifferent', 'intensity': 0}, 'envy': {'percentage': 10, 'motivation': 'jealous', 'intensity': 0}, 'neutral': {'percentage': 10, 'motivation': 'balanced', 'intensity': 0} } total_percentage = 200 emotion_history_file = 'emotion_history.json' def load_historical_data(file_path=emotion_history_file): if os.path.exists(file_path): with open(file_path, 'r') as file: return json.load(file) return [] def save_historical_data(historical_data, file_path=emotion_history_file): with open(file_path, 'w') as file: json.dump(historical_data, file) emotion_history = load_historical_data() def update_emotion(emotion, percentage, intensity): emotions['ideal_state']['percentage'] -= percentage emotions[emotion]['percentage'] += percentage emotions[emotion]['intensity'] = intensity total_current = sum(e['percentage'] for e in emotions.values()) adjustment = total_percentage - total_current emotions['ideal_state']['percentage'] += adjustment def normalize_context(context): return context.lower().strip() def evaluate(individual): emotion_values = individual[:len(emotions) - 1] intensities = individual[-21:-1] ideal_state = individual[-1] ideal_diff = abs(100 - ideal_state) sum_non_ideal = sum(emotion_values) intensity_range = max(intensities) - min(intensities) return ideal_diff, sum_non_ideal, intensity_range def evolve_emotions(): creator.create("FitnessMulti", base.Fitness, weights=(-1.0, -0.5, -0.2)) creator.create("Individual", list, fitness=creator.FitnessMulti) toolbox = base.Toolbox() toolbox.register("attr_float", random.uniform, 0, 20) toolbox.register("attr_intensity", random.uniform, 0, 10) toolbox.register("individual", tools.initCycle, creator.Individual, (toolbox.attr_float,) * (len(emotions) - 1) + (toolbox.attr_intensity,) * len(emotions) + (lambda: 100,), n(toolbox.attr_intensity,) * len(emotions) + (lambda: 100,), n=1) toolbox.register("population", tools.initRepeat, list, toolbox.individual) toolbox.register("mate", tools.cxTwoPoint) toolbox.register("mutate", tools.mutGaussian, mu=0, sigma=1, indpb=0.2) toolbox.register("select", tools.selNSGA2) toolbox.register("evaluate", evaluate) population = toolbox.population(n=100) algorithms.eaMuPlusLambda(population, toolbox, mu=50, lambda_=100, cxpb=0.7, mutpb=0.2, ngen=100, stats=None, halloffame=None, verbose=False) best_individual = tools.selBest(population, k=1)[0] emotion_values = best_individual[:len(emotions) - 1] intensities = best_individual[-21:-1] ideal_state = best_individual[-1] for i, emotion in enumerate(emotions): emotions[emotion]['percentage'] = emotion_values[i] emotions[emotion]['intensity'] = intensities[i] emotions['ideal_state']['percentage'] = ideal_state def predict_emotion(context): emotion_prediction_pipeline = pipeline('text-classification', model=emotion_prediction_model, tokenizer=emotion_prediction_tokenizer, top_k=None) predictions = emotion_prediction_pipeline(context) emotion_scores = {prediction['label']: prediction['score'] for prediction in predictions[0]} predicted_emotion = max(emotion_scores, key=emotion_scores.get) # Map the predicted emotion to our emotion categories emotion_mapping = { 'sadness': 'sadness', 'joy': 'joy', 'love': 'pleasure', 'anger': 'anger', 'fear': 'fear', 'surprise': 'surprise' } return emotion_mapping.get(predicted_emotion, 'neutral') def generate_text(prompt, chat_history, emotion=None, max_length=100): finetuned_lm_tokenizer, finetuned_lm_model = get_finetuned_lm_model() # Prepare the input by concatenating the chat history and the new prompt full_prompt = "You are Adam, an AI assistant. Respond to the following conversation in a natural and engaging way, considering the emotion: " + emotion + "\n\n" for turn in chat_history[-5:]: # Consider last 5 turns for context full_prompt += f"Human: {turn[0]}\nAdam: {turn[1]}\n" full_prompt += f"Human: {prompt}\nAdam:" input_ids = finetuned_lm_tokenizer(full_prompt, return_tensors="pt", max_length=512, truncation=True).input_ids if torch.cuda.is_available(): input_ids = input_ids.cuda() finetuned_lm_model = finetuned_lm_model.cuda() # Generate the response outputs = finetuned_lm_model.generate( input_ids, max_length=max_length + input_ids.shape[1], num_return_sequences=1, no_repeat_ngram_size=2, do_sample=True, temperature=0.7, top_k=50, top_p=0.95 ) generated_text = finetuned_lm_tokenizer.decode(outputs[0], skip_special_tokens=True) return generated_text.strip() def update_emotion_history(emotion, intensity): global emotion_history emotion_history.append({ 'emotion': emotion, 'intensity': intensity, 'timestamp': pd.Timestamp.now().isoformat() }) save_historical_data(emotion_history) def get_dominant_emotion(): return max(emotions, key=lambda x: emotions[x]['percentage'] if x != 'ideal_state' else 0) def get_emotion_summary(): summary = [] for emotion, data in emotions.items(): if emotion != 'ideal_state': summary.append(f"{emotion.capitalize()}: {data['percentage']:.1f}% (Intensity: {data['intensity']:.1f})") return "\n".join(summary) def reset_emotions(): global emotions for emotion in emotions: if emotion != 'ideal_state': emotions[emotion]['percentage'] = 10 emotions[emotion]['intensity'] = 0 emotions['ideal_state']['percentage'] = 100 return get_emotion_summary() def respond_to_user(user_input, chat_history): # Predict the emotion from the user input predicted_emotion = predict_emotion(user_input) # Ensure the predicted emotion exists in our emotions dictionary if predicted_emotion not in emotions: predicted_emotion = 'neutral' # Update the emotional state update_emotion(predicted_emotion, 5, random.uniform(0, 10)) # Get the current dominant emotion dominant_emotion = get_dominant_emotion() # Generate a response considering the dominant emotion response = generate_text(user_input, chat_history, dominant_emotion) # Update emotion history update_emotion_history(predicted_emotion, emotions[predicted_emotion]['intensity']) # Update chat history chat_history.append((user_input, response)) # Evolve emotions periodically (e.g., every 10 interactions) if len(chat_history) % 10 == 0: evolve_emotions() return response, chat_history, get_emotion_summary() # Gradio interface with gr.Blocks() as demo: gr.Markdown("# Adam: Emotion-Aware AI Chatbot") gr.Markdown("Chat with Adam, an AI that understands and responds to emotions.") chatbot = gr.Chatbot() msg = gr.Textbox(label="Type your message here...") clear = gr.Button("Clear") emotion_state = gr.Textbox(label="Current Emotional State", lines=10) reset_button = gr.Button("Reset Emotions") def user(user_message, history): response, updated_history, emotion_summary = respond_to_user(user_message, history) return "", updated_history, emotion_summary msg.submit(user, [msg, chatbot], [msg, chatbot, emotion_state]) clear.click(lambda: None, None, chatbot, queue=False) reset_button.click(reset_emotions, None, emotion_state, queue=False) if __name__ == "__main__": demo.launch()