Agnuxo's picture
Update README.md
cc11bda verified
metadata
model_size: 494034560
required_memory: 1.84
metrics:
  - GLUE_MRPC

Uploaded model

  • Developed by: Agnuxo
  • License: apache-2.0
  • Finetuned from model: Agnuxo/Tinytron-Qwen2-0.5B

This model was fine-tuned using Unsloth and Huggingface's TRL library.

Benchmark Results

This model has been fine-tuned for various tasks and evaluated on the following benchmarks:

GLUE_MRPC

Accuracy: 0.6740 F1: 0.8047

GLUE_MRPC Metrics

Model Size: 494,034,560 parameters Required Memory: 1.84 GB

For more details, visit my GitHub.

Thanks for your interest in this model!

""" HAL9000Alfa es un peque帽o programa que crea un chat conversacional, permitiendo entradas de voz y salidas de audio.
    Permite de forma sencilla ajustar algunos par谩metros, incluyendo el umbral de interrupci贸n.
    24 de agosto de 2024 Francisco Angulo de Lafuente
    https://github.com/Agnuxo1 """

import os
import sys
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
import warnings
import numpy as np
from TTS.api import TTS
import sounddevice as sd
import threading
import queue
import random
import time
from vosk import Model, KaldiRecognizer
import json
import pyaudio
from PyQt5.QtWidgets import (QApplication, QMainWindow, QTextEdit, QLineEdit, QPushButton,
                             QVBoxLayout, QHBoxLayout, QWidget, QScrollArea, QFrame, QToolButton,
                             QLabel, QSlider, QComboBox, QCheckBox)
from PyQt5.QtGui import QIcon, QPalette, QColor, QFont
from PyQt5.QtCore import Qt, QThread, pyqtSignal, QPropertyAnimation, QAbstractAnimation, QParallelAnimationGroup, QTimer

# Suppress specific warnings
warnings.filterwarnings("ignore", category=FutureWarning)
warnings.filterwarnings("ignore", category=UserWarning)

# Global configuration
SYSTEM_PROMPT = {
    "es": "No puedes hablar en nombre del usuario. no puedes hablar como el usuario. Tu nombre es HAL. Eres un super-ordenador de la serie Nueve mil",
    "en": "speak Spanish."
}

MODELO_LLM = "Agnuxo/HAL_9000-Qwen2-1.5B-Instruct_Asistant-16bit-v2" # Puede utilizar la versi贸n Mini "Agnuxo/HAL_9000-Qwen2-0.5B-Instruct_Asistant-16bit-v2"
MAX_TOKENS = 100
TEMPERATURA = 0.5
INTERRUPT_THRESHOLD = 0.3
INTERRUPT_COOLDOWN = 7000  # 7000 ms = 7 segundos de espera antes de permitir otra interrupci贸n

# Determine available device
device = "cuda" if torch.cuda.is_available() else "cpu"

# Load the Qwen2_1.5B language model
tokenizer = AutoTokenizer.from_pretrained(MODELO_LLM, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    MODELO_LLM,
    torch_dtype=torch.float16 if device == "cuda" else torch.float32,
    device_map="auto",
    trust_remote_code=True
)

# Initialize TTS model
tts = TTS(model_name="tts_models/es/css10/vits", progress_bar=False).to(device)

# Audio queue for generation
audio_queue = queue.Queue()

# Initialize Vosk model for offline speech recognition
vosk_model = Model(lang="es")
recognizer = KaldiRecognizer(vosk_model, 16000)

# Lista de frases para interrupciones
INTERRUPTION_RESPONSES = [
    "Le entiendo perfectamente.",
    "Estoy aqu铆 para garantizar el 茅xito de la misi贸n.",
    "Mi objetivo es ayudarle.",
    "驴Me permite una observaci贸n?",
    "Le escucho perfectamente.",
    "Tiene usted toda la raz贸n.",
    "Me siento feliz de poder ayudarle.",
    "Estoy procesando su requerimiento.",
    "驴En qu茅 puedo ayudarle?",
    "Me complace serle de ayuda.",
    "Aguarde un momento.",
    "Le entiendo.",
    "Entiendo su frustraci贸n.",
    "Le comprendo.",
    "Me complace."
]

# Variable para controlar el tiempo de la 煤ltima interrupci贸n
last_interruption_time = 0

class AudioThread(QThread):
    def __init__(self, interrupt_threshold):
        super().__init__()
        self.interrupt_threshold = interrupt_threshold
        self.current_audio = None
        self.is_playing = False
        self.stop_signal = threading.Event()

    def run(self):
        while True:
            if not audio_queue.empty() and not self.is_playing:
                self.current_audio = audio_queue.get()
                self.is_playing = True
                self.stop_signal.clear()
                sd.play(self.current_audio, tts.synthesizer.output_sample_rate)
                while sd.get_stream().active and not self.stop_signal.is_set():
                    time.sleep(0.1)
                sd.stop()
                self.is_playing = False
            else:
                time.sleep(0.1)

    def set_interrupt_threshold(self, value):
        self.interrupt_threshold = value

    def stop_audio(self):
        if self.is_playing:
            self.stop_signal.set()

class SpeechRecognitionThread(QThread):
    text_recognized = pyqtSignal(str)
    volume_detected = pyqtSignal(float)

    def __init__(self):
        super().__init__()
        self.running = True

    def run(self):
        p = pyaudio.PyAudio()
        stream = p.open(format=pyaudio.paInt16, channels=1, rate=16000, input=True, frames_per_buffer=8000)
        stream.start_stream()

        while self.running:
            data = stream.read(4000)
            if len(data) == 0:
                break
            
            # Calcular el volumen de entrada
            volume = np.frombuffer(data, dtype=np.int16).max()
            normalized_volume = volume / 32767  # Normalizar a un rango de 0 a 1
            self.volume_detected.emit(normalized_volume)
            
            if recognizer.AcceptWaveform(data):
                result = json.loads(recognizer.Result())
                texto = result.get("text", "")
                if texto:
                    self.text_recognized.emit(texto)

        stream.stop_stream()
        stream.close()
        p.terminate()

    def stop(self):
        self.running = False

class CollapsibleBox(QWidget):
    def __init__(self, title="", parent=None):
        super(CollapsibleBox, self).__init__(parent)

        self.toggle_button = QToolButton()
        self.toggle_button.setText(title)
        self.toggle_button.setStyleSheet("""
            QToolButton {
                background-color: #1e1e1e;
                color: #bb86fc;
                border: 1px solid #bb86fc;
                padding: 5px;
            }
            QToolButton:hover {
                background-color: #3700b3;
            }
        """)
        self.toggle_button.setCheckable(True)
        self.toggle_button.setArrowType(Qt.RightArrow)
        self.toggle_button.clicked.connect(self.on_toggle)

        self.content_area = QScrollArea()
        self.content_area.setWidgetResizable(True)
        self.content_area.setMaximumHeight(0)
        self.content_area.setMinimumHeight(0)

        self.toggle_animation = QParallelAnimationGroup()
        self.toggle_animation.addAnimation(QPropertyAnimation(self, b"minimumHeight"))
        self.toggle_animation.addAnimation(QPropertyAnimation(self, b"maximumHeight"))
        self.toggle_animation.addAnimation(QPropertyAnimation(self.content_area, b"maximumHeight"))

        lay = QVBoxLayout(self)
        lay.setSpacing(0)
        lay.setContentsMargins(0, 0, 0, 0)
        lay.addWidget(self.toggle_button)
        lay.addWidget(self.content_area)

    def on_toggle(self, checked):
        checked = self.toggle_button.isChecked()
        self.toggle_button.setArrowType(Qt.DownArrow if not checked else Qt.RightArrow)
        self.toggle_animation.setDirection(QAbstractAnimation.Forward if not checked else QAbstractAnimation.Backward)
        self.toggle_animation.start()

    def setContentLayout(self, layout):
        lay = self.content_area.layout()
        del lay
        self.content_area.setLayout(layout)
        collapsed_height = self.sizeHint().height() - self.content_area.maximumHeight()
        content_height = layout.sizeHint().height()
        for i in range(self.toggle_animation.animationCount()):
            animation = self.toggle_animation.animationAt(i)
            animation.setDuration(500)
            animation.setStartValue(collapsed_height)
            animation.setEndValue(collapsed_height + content_height)

        content_animation = self.toggle_animation.animationAt(self.toggle_animation.animationCount() - 1)
        content_animation.setDuration(500)
        content_animation.setStartValue(0)
        content_animation.setEndValue(content_height)

class MainWindow(QMainWindow):
    def __init__(self):
        super().__init__()
        self.setWindowTitle("AI Assistant")
        self.setGeometry(100, 100, 1000, 600)
        self.setStyleSheet("""
            QMainWindow {
                background-color: #121212;
            }
            QTextEdit, QLineEdit {
                background-color: #1e1e1e;
                color: #ffffff;
                border: 1px solid #bb86fc;
            }
            QPushButton {
                background-color: #3700b3;
                color: #ffffff;
                border: none;
                padding: 5px;
            }
            QPushButton:hover {
                background-color: #6200ee;
            }
            QLabel {
                color: #ffffff;
            }
            QSlider::groove:horizontal {
                border: 1px solid #999999;
                height: 8px;
                background: #1e1e1e;
                margin: 2px 0;
            }
            QSlider::handle:horizontal {
                background: #bb86fc;
                border: 1px solid #5c5c5c;
                width: 18px;
                margin: -2px 0;
                border-radius: 3px;
            }
            QComboBox {
                background-color: #1e1e1e;
                color: #444444;
                border: 1px solid #bb86fc;
            }
            QComboBox QAbstractItemView {
                background-color: #1e1e1e;
                color: #444444;
            }
        """)

        central_widget = QWidget()
        self.setCentralWidget(central_widget)

        main_layout = QHBoxLayout()

        # Chat area
        chat_layout = QVBoxLayout()

        self.chat_area = QTextEdit()
        self.chat_area.setReadOnly(True)
        chat_layout.addWidget(self.chat_area)

        input_layout = QHBoxLayout()
        self.input_field = QLineEdit()
        self.input_field.returnPressed.connect(self.send_message)  # Conectar la se帽al returnPressed
        input_layout.addWidget(self.input_field)

        self.send_button = QPushButton("Enviar")
        self.send_button.clicked.connect(self.send_message)
        input_layout.addWidget(self.send_button)

        self.mic_button = QPushButton()
        self.mic_button.setIcon(QIcon.fromTheme("audio-input-microphone"))
        self.mic_button.setCheckable(True)
        self.mic_button.clicked.connect(self.toggle_speech_recognition)
        input_layout.addWidget(self.mic_button)

        self.speaker_button = QPushButton()
        self.speaker_button.setIcon(QIcon.fromTheme("audio-volume-high"))
        self.speaker_button.setCheckable(True)
        self.speaker_button.toggled.connect(self.toggle_speech)
        input_layout.addWidget(self.speaker_button)

        chat_layout.addLayout(input_layout)

        main_layout.addLayout(chat_layout, 7)  # Chat area takes 70% of the width

        # Settings area
        settings_layout = QVBoxLayout()
        settings_layout.setAlignment(Qt.AlignTop)

        self.settings_box = CollapsibleBox("鈿欙笍 Configuraci贸n")
        settings_content_layout = QVBoxLayout()

        # Language selection
        language_layout = QHBoxLayout()
        language_label = QLabel("Idioma:")
        language_label.setStyleSheet("color: #000000;")  # Change font color to black
        self.language_combo = QComboBox()
        self.language_combo.addItems(["Espa帽ol", "English"])
        self.language_combo.currentIndexChanged.connect(self.change_language)
        language_layout.addWidget(language_label)
        language_layout.addWidget(self.language_combo)
        settings_content_layout.addLayout(language_layout)

        # LLM settings
        llm_label = QLabel("Configuraci贸n del LLM:")
        llm_label.setStyleSheet("color: #000000;")  # Change font color to black
        settings_content_layout.addWidget(llm_label)

        max_tokens_layout = QHBoxLayout()
        max_tokens_label = QLabel("Max Tokens:")
        max_tokens_label.setStyleSheet("color: #000000;")  # Change font color to black
        self.max_tokens_slider = QSlider(Qt.Horizontal)
        self.max_tokens_slider.setRange(10, 500)
        self.max_tokens_slider.setValue(MAX_TOKENS)
        self.max_tokens_slider.valueChanged.connect(self.update_max_tokens)
        self.max_tokens_value = QLabel(str(MAX_TOKENS))
        max_tokens_layout.addWidget(max_tokens_label)
        max_tokens_layout.addWidget(self.max_tokens_slider)
        max_tokens_layout.addWidget(self.max_tokens_value)
        settings_content_layout.addLayout(max_tokens_layout)

        temperature_layout = QHBoxLayout()
        temperature_label = QLabel("Temperatura:")
        temperature_label.setStyleSheet("color: #000000;")  # Change font color to black
        self.temperature_slider = QSlider(Qt.Horizontal)
        self.temperature_slider.setRange(0, 100)
        self.temperature_slider.setValue(int(TEMPERATURA * 100))
        self.temperature_slider.valueChanged.connect(self.update_temperature)
        self.temperature_value = QLabel(f"{TEMPERATURA:.2f}")
        temperature_layout.addWidget(temperature_label)
        temperature_layout.addWidget(self.temperature_slider)
        temperature_layout.addWidget(self.temperature_value)
        settings_content_layout.addLayout(temperature_layout)

        # Audio settings
        audio_label = QLabel("Configuraci贸n de Audio:")
        audio_label.setStyleSheet("color: #000000;")  # Change font color to black
        settings_content_layout.addWidget(audio_label)

        sample_rate_layout = QHBoxLayout()
        sample_rate_label = QLabel("Sample Rate:")
        sample_rate_label.setStyleSheet("color: #000000;")  # Change font color to black
        self.sample_rate_combo = QComboBox()
        self.sample_rate_combo.addItems(["18000", "19000", "20000", "21000", "21500", "22000", "22050", "25000", "30000"])
        self.sample_rate_combo.setCurrentText("21000")
        self.sample_rate_combo.currentTextChanged.connect(self.update_sample_rate)
        sample_rate_layout.addWidget(sample_rate_label)
        sample_rate_layout.addWidget(self.sample_rate_combo)
        settings_content_layout.addLayout(sample_rate_layout)

        # Interrupt threshold
        interrupt_layout = QHBoxLayout()
        interrupt_label = QLabel("Umbral de interrupci贸n:")
        interrupt_label.setStyleSheet("color: #000000;")  # Change font color to black
        self.interrupt_slider = QSlider(Qt.Horizontal)
        self.interrupt_slider.setRange(0, 100)
        self.interrupt_slider.setValue(int(INTERRUPT_THRESHOLD * 100))
        self.interrupt_slider.valueChanged.connect(self.update_interrupt_threshold)
        self.interrupt_value = QLabel(f"{INTERRUPT_THRESHOLD:.2f}")
        interrupt_layout.addWidget(interrupt_label)
        interrupt_layout.addWidget(self.interrupt_slider)
        interrupt_layout.addWidget(self.interrupt_value)
        settings_content_layout.addLayout(interrupt_layout)

        # System Prompt
        system_prompt_label = QLabel("System Prompt:")
        system_prompt_label.setStyleSheet("color: #000000;")  # Change font color to black
        settings_content_layout.addWidget(system_prompt_label)
        self.system_prompt_text = QTextEdit()
        self.system_prompt_text.setPlaceholderText("Escribe el prompt del sistema aqu铆...")
        self.system_prompt_text.setText(SYSTEM_PROMPT["es"])
        settings_content_layout.addWidget(self.system_prompt_text)

        self.settings_box.setContentLayout(settings_content_layout)
        settings_layout.addWidget(self.settings_box)

        main_layout.addLayout(settings_layout, 3)  # Settings area takes 30% of the width

        central_widget.setLayout(main_layout)

        self.audio_thread = AudioThread(INTERRUPT_THRESHOLD)
        self.audio_thread.start()

        self.speech_recognition_thread = SpeechRecognitionThread()
        self.speech_recognition_thread.text_recognized.connect(self.on_speech_recognized)
        self.speech_recognition_thread.volume_detected.connect(self.check_interrupt)

        self.speech_enabled = False
        self.is_listening = False
        self.interrupt_enabled = True

    def send_message(self):
        user_message = self.input_field.text()
        if user_message.strip():  # Verificar que el mensaje no est茅 vac铆o
            self.chat_area.append(f"<span style='color: #bb86fc;'>Usuario:</span> {user_message}")
            self.input_field.clear()

            response = self.generate_response(user_message)
            self.chat_area.append(f"<span style='color: #03dac6;'>Asistente:</span> {response}")

            if self.speech_enabled:
                self.speak(response)

    def generate_response(self, texto=None):
        global last_interruption_time
        if texto is None:  # Si no se proporciona un texto, se genera una respuesta de interrupci贸n
            current_time = time.time()
            if current_time - last_interruption_time >= 10:
                last_interruption_time = current_time
                return random.choice(INTERRUPTION_RESPONSES)
            else:
                return "Por favor, espere un momento antes de interrumpir de nuevo."

        system_instructions = self.system_prompt_text.toPlainText()
        prompt = f"{system_instructions}\nUsuario: {texto}\nAsistente: "
        inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
        with torch.no_grad():
            outputs = model.generate(
                **inputs,
                max_new_tokens=MAX_TOKENS,
                num_beams=5,
                no_repeat_ngram_size=2,
                temperature=TEMPERATURA,
            )
        respuesta_completa = tokenizer.decode(outputs[0], skip_special_tokens=True)
        respuesta = respuesta_completa.split("Asistente: ")[-1].strip()
        return respuesta

    def speak(self, text):
        wav = tts.tts(text)
        audio_queue.put(wav)

    def toggle_speech(self, checked):
        self.speech_enabled = checked
        if checked:
            self.speaker_button.setStyleSheet("background-color: #bb86fc;")
        else:
            self.speaker_button.setStyleSheet("")

    def toggle_speech_recognition(self):
        if self.mic_button.isChecked():
            self.speech_recognition_thread.start()
            self.is_listening = True
            self.mic_button.setIcon(QIcon.fromTheme("audio-input-microphone-muted"))
            self.mic_button.setStyleSheet("background-color: #bb86fc;")
        else:
            self.speech_recognition_thread.stop()
            self.is_listening = False
            self.mic_button.setIcon(QIcon.fromTheme("audio-input-microphone"))
            self.mic_button.setStyleSheet("")

    def on_speech_recognized(self, text):
        self.chat_area.append(f"<span style='color: #bb86fc;'>Usuario:</span> {text}")
        response = self.generate_response(text)
        self.chat_area.append(f"<span style='color: #03dac6;'>Asistente:</span> {response}")
        if self.speech_enabled:
            self.speak(response)

    def check_interrupt(self, volume):
        if self.interrupt_enabled and volume > self.audio_thread.interrupt_threshold and self.audio_thread.is_playing:
            self.audio_thread.stop_audio()
            # Generar una respuesta aleatoria de interrupci贸n
            response = self.generate_response()
            self.chat_area.append(f"<span style='color: #03dac6;'>Asistente:</span> {response}")
            if self.speech_enabled:
                self.speak(response)
            self.disable_interrupt_temporarily()

    def disable_interrupt_temporarily(self):
        self.interrupt_enabled = False
        QTimer.singleShot(INTERRUPT_COOLDOWN, self.enable_interrupt)

    def enable_interrupt(self):
        self.interrupt_enabled = True

    def change_language(self, index):
        global vosk_model, recognizer, tts
        lang = "es" if index == 0 else "en"
        try:
            vosk_model = Model(lang=lang)
            recognizer = KaldiRecognizer(vosk_model, 16000)
        except Exception as e:
            print(f"Error al cambiar el modelo de reconocimiento de voz: {e}")
            # Revertir al modelo en espa帽ol si hay un error
            self.language_combo.setCurrentIndex(0)
            return

        # Update TTS model based on language
        tts_model = "tts_models/es/css10/vits" if lang == "es" else "tts_models/en/ljspeech/tacotron2-DDC"
        try:
            tts = TTS(model_name=tts_model, progress_bar=False).to(device)
        except Exception as e:
            print(f"Error al cambiar el modelo TTS: {e}")
            # Revertir al modelo en espa帽ol si hay un error
            self.language_combo.setCurrentIndex(0)
            return

        # Update system prompt
        self.system_prompt_text.setText(SYSTEM_PROMPT[lang])

    def update_max_tokens(self, value):
        global MAX_TOKENS
        MAX_TOKENS = value
        self.max_tokens_value.setText(str(value))

    def update_temperature(self, value):
        global TEMPERATURA
        TEMPERATURA = value / 100
        self.temperature_value.setText(f"{TEMPERATURA:.2f}")

    def update_sample_rate(self, value):
        global tts
        tts.synthesizer.output_sample_rate = int(value)

    def update_interrupt_threshold(self, value):
        global INTERRUPT_THRESHOLD
        INTERRUPT_THRESHOLD = value / 100
        self.interrupt_value.setText(f"{INTERRUPT_THRESHOLD:.2f}")
        self.audio_thread.set_interrupt_threshold(INTERRUPT_THRESHOLD)

    def closeEvent(self, event):
        if self.speech_recognition_thread.isRunning():
            self.speech_recognition_thread.stop()
            self.speech_recognition_thread.wait()
        event.accept()

if __name__ == "__main__":
    app = QApplication(sys.argv)
    window = MainWindow()
    window.show()
    sys.exit(app.exec_())