app.py

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch
import gradio as gr
import openai
import pyttsx3
import subprocess
import os

tokenizer = AutoTokenizer.from_pretrained("microsoft/DialoGPT-large",padding_side = 'left')
model = AutoModelForCausalLM.from_pretrained("microsoft/DialoGPT-large")

# Load the GPT-NeoX tokenizer and model
neox_tokenizer = AutoTokenizer.from_pretrained("EleutherAI/gpt-neox-20b")
neox_model = AutoModelForCausalLM.from_pretrained("EleutherAI/gpt-neox-20b")


openai.api_key = os.getenv("OPENAI_API_KEY")


messages = [
    {"role": "system", "content": "You are a teacher"}
]

def predict(input, history=[]):
    new_user_input_ids = tokenizer.encode(input + tokenizer.eos_token, return_tensors='pt')
    bot_input_ids = torch.cat([torch.LongTensor(history), new_user_input_ids], dim=-1)
    history = model.generate(bot_input_ids, max_length=500, pad_token_id=tokenizer.eos_token_id).tolist()
    response = tokenizer.decode(history[0][len(new_user_input_ids[0]):])
    response = response.replace(tokenizer.eos_token, "").strip()

    return response, history

def transcribe(audio):
    print("Starting transcription process...")
    global messages
    input_audio = audio
    output_audio = "temp_audio.wav"

    # Delete the previous output file, if it exists
    if os.path.exists(output_audio):
        os.remove(output_audio)

    command = f"ffmpeg -i {input_audio} {output_audio}"
    subprocess.run(command, shell=True, stderr=subprocess.DEVNULL)

    with open(output_audio, "rb") as file:
        transcription = openai.Audio.transcribe("whisper-1", file)
    print("Transcription completed:", transcription)
    
    messages.append({"role": "user", "content": transcription["text"]})
    response = openai.ChatCompletion.create(
        model="gpt-3.5-turbo",
        messages=messages
    )

    AImessage = response["choices"][0]["message"]["content"]
    engine = pyttsx3.init()
    print("Speaking response...")
    engine.say(AImessage)
    engine.runAndWait()
    print("Response spoken.")
    messages.append({"role": "assistant", "content": AImessage})

    return transcription["text"]

def combined_fn(input_text, audio, mode, history_state=""):
    dialogpt_response = ""
    gpt3_response = ""
    neox_response = ""

    if history_state:
        history_list = list(map(int, history_state.strip().split(',')))
    else:
        history_list = []

    if mode == "text":
        dialogpt_response, new_history_state = predict(dialogpt_model, dialogpt_tokenizer, input_text, history_list)
        
        # GPT-3.5-turbo response for text mode
        messages.append({"role": "user", "content": input_text})
        response = openai.ChatCompletion.create(
            model="gpt-3.5-turbo",
            messages=messages
        )
        gpt3_response = response["choices"][0]["message"]["content"]
        messages.append({"role": "assistant", "content": gpt3_response})

        # GPT-NeoX response for text mode
        neox_response, _ = predict(neox_model, neox_tokenizer, input_text, history_list)
        
    elif mode == "audio":
        transcribed_text = transcribe(audio)
        dialogpt_response, new_history_state = predict(dialogpt_model, dialogpt_tokenizer, transcribed_text, history_list)
        gpt3_response = transcribed_text
        neox_response, _ = predict(neox_model, neox_tokenizer, transcribed_text, history_list)
        
    else:
        raise ValueError("Invalid mode selected")

    responses = {
        "DialoGPT-large Response": dialogpt_response,
        "GPT-3.5-turbo Response": gpt3_response,
        "GPT-NeoX Response": neox_response,
        "history_state": ",".join(map(str, new_history_state))  # Convert list of integers to a string
    }

    return responses






inputs = [
    gr.inputs.Textbox(placeholder="Write a text message as if writing a text message to a human."),
    gr.inputs.Audio(source='microphone', type='filepath'),
    gr.inputs.Radio(["text", "audio"], label="Mode", default="text"),
    gr.inputs.Textbox(label="history_state", default=""),
]

outputs = [
    gr.outputs.Textbox(label="DialoGPT-large Response"),
    gr.outputs.Textbox(label="GPT-3.5-turbo Response"),
    gr.outputs.Textbox(label="GPT-NeoX Response"),
    gr.outputs.Textbox(label="history_output"),
]

gr.Interface(fn=combined_fn, title="Chatbot Interface", inputs=inputs, outputs=outputs).launch()