duplex.py

import os
import json
import random
import string

import numpy as np
import gradio as gr
import requests
import soundfile as sf

from transformers import pipeline, set_seed
from transformers import AutoTokenizer, AutoModelForCausalLM
import logging

import sys
import gradio as gr
from transformers import pipeline, AutoModelForCTC, Wav2Vec2Processor, Wav2Vec2ProcessorWithLM

DEBUG = os.environ.get("DEBUG", "false")[0] in "ty1"
MAX_LENGTH = int(os.environ.get("MAX_LENGTH", 1024))
DEFAULT_LANG = os.environ.get("DEFAULT_LANG", "English")
HF_AUTH_TOKEN = os.environ.get("HF_AUTH_TOKEN", None)

HEADER = """
# Poor Man's Duplex

Talk to a language model like you talk on a Walkie-Talkie! Well, with larger latencies.
The models are [EleutherAI's GPT-J-6B](https://huggingface.co/EleutherAI/gpt-j-6B) for English, and [BERTIN GPT-J-6B](https://huggingface.co/bertin-project/bertin-gpt-j-6B) for Spanish.
""".strip()

FOOTER = """
<div align=center>
<img src="https://visitor-badge.glitch.me/badge?page_id=versae/poor-mans-duplex"/>
<div align=center>
""".strip()

asr_model_name_es = "jonatasgrosman/wav2vec2-large-xlsr-53-spanish"
model_instance_es = AutoModelForCTC.from_pretrained(asr_model_name_es, use_auth_token=HF_AUTH_TOKEN)
processor_es = Wav2Vec2ProcessorWithLM.from_pretrained(asr_model_name_es, use_auth_token=HF_AUTH_TOKEN)
asr_es = pipeline(
    "automatic-speech-recognition",
    model=model_instance_es,
    tokenizer=processor_es.tokenizer,
    feature_extractor=processor_es.feature_extractor,
    decoder=processor_es.decoder
)
tts_model_name = "facebook/tts_transformer-es-css10"
speak_es = gr.Interface.load(f"huggingface/{tts_model_name}", api_key=HF_AUTH_TOKEN)
transcribe_es = lambda input_file: asr_es(input_file, chunk_length_s=5, stride_length_s=1)["text"]
def generate_es(text, **kwargs):
    # text="Promtp", max_length=100, top_k=100, top_p=50, temperature=0.95, do_sample=True, do_clean=True
    api_uri = "https://hf.space/embed/bertin-project/bertin-gpt-j-6B/+/api/predict/"
    response = requests.post(api_uri, data=json.dumps({"data": [text, kwargs["max_length"], 100, 50, 0.95, True, True]}))
    if response.ok:
        if DEBUG:
            print("Spanish response >", response.json())
        return response.json()["data"][0]
    else:
        return ""

asr_model_name_en = "jonatasgrosman/wav2vec2-large-xlsr-53-english"
model_instance_en = AutoModelForCTC.from_pretrained(asr_model_name_en)
processor_en = Wav2Vec2ProcessorWithLM.from_pretrained(asr_model_name_en)
asr_en = pipeline(
    "automatic-speech-recognition",
    model=model_instance_en,
    tokenizer=processor_en.tokenizer,
    feature_extractor=processor_en.feature_extractor,
    decoder=processor_en.decoder
)
tts_model_name = "facebook/fastspeech2-en-ljspeech"
speak_en = gr.Interface.load(f"huggingface/{tts_model_name}", api_key=HF_AUTH_TOKEN)
transcribe_en = lambda input_file: asr_en(input_file, chunk_length_s=5, stride_length_s=1)["text"]
# generate_iface = gr.Interface.load("huggingface/EleutherAI/gpt-j-6B", api_key=HF_AUTH_TOKEN)

empty_audio = 'empty.flac'
sf.write(empty_audio, [], 16000)
deuncase = gr.Interface.load("huggingface/pere/DeUnCaser", api_key=HF_AUTH_TOKEN)

def generate_en(text, **kwargs):
    api_uri = "https://api.eleuther.ai/completion"
    #--data-raw '{"context":"Promtp","top_p":0.9,"temp":0.8,"response_length":128,"remove_input":true}'
    response = requests.post(api_uri, data=json.dumps({"context": text, "top_p": 0.9, "temp": 0.8, "response_length": kwargs["max_length"], "remove_input": True}))
    if response.ok:
        if DEBUG:
            print("English response >", response.json())
        return response.json()[0]["generated_text"].lstrip()
    else:
        return ""


def select_lang(lang):
    if lang.lower() == "spanish":
        return generate_es, transcribe_es, speak_es
    else:
        return generate_en, transcribe_en, speak_en


def select_lang_vars(lang):
    if lang.lower() == "spanish":
        AGENT = "BERTIN"
        USER = "ENTREVISTADOR"
        CONTEXT = """La siguiente conversación es un extracto de una entrevista a {AGENT} celebrada en Madrid para Radio Televisión Española:

{USER}: Bienvenido, {AGENT}. Un placer tenerlo hoy con nosotros.
{AGENT}: Gracias. El placer es mío."""
    else:
        AGENT = "ELEUTHER"
        USER = "INTERVIEWER"
        CONTEXT = """The next conversation is an excerpt from an interview to {AGENT} that appeared in the New York Times:

{USER}: Welcome, {AGENT}. It is a pleasure to have you here today.
{AGENT}: Thanks. The pleasure is mine."""

    return AGENT, USER, CONTEXT


def format_chat(history):
    interventions = []
    for user, bot in history:
        interventions.append(f"""
            <div data-testid="user" style="background-color:#16a34a" class="px-3 py-2 rounded-[22px] rounded-bl-none place-self-start text-white ml-7 text-sm">{user}</div>
            <div data-testid="bot" style="background-color:gray" class="px-3 py-2 rounded-[22px] rounded-br-none  text-white ml-7 text-sm">{bot}</div>
        """)
    return f"""<details><summary>Conversation log</summary>
    <div class="overflow-y-auto h-[40vh]">
        <div class="flex flex-col items-end space-y-4 p-3">
        {"".join(interventions)}
        </div>
    </div>
    </summary>"""


def chat_with_gpt(lang, agent, user, context, audio_in, history):
    if not audio_in:
        return history, history, empty_audio, format_chat(history)
    generate, transcribe, speak = select_lang(lang)
    AGENT, USER, _ = select_lang_vars(lang)
    user_message = deuncase(transcribe(audio_in))
    # agent = AGENT
    # user = USER
    generation_kwargs = {
        "max_length": 50,
        # "top_k": top_k,
        # "top_p": top_p,
        # "temperature": temperature,
        # "do_sample": do_sample,
        # "do_clean": do_clean,
        # "num_return_sequences": 1,
        # "return_full_text": False,
    }
    message = user_message.split(" ", 1)[0].capitalize() + " " + user_message.split(" ", 1)[-1]
    history = history or []  #[(f"{user}: Bienvenido. Encantado de tenerle con nosotros.", f"{agent}: Un placer, muchas gracias por la invitación.")]
    context = context.format(USER=user or USER, AGENT=agent or AGENT).strip()
    if context[-1] not in ".:":
        context += "."
    context_length = len(context.split())
    history_take = 0
    history_context = "\n".join(f"{user}: {history_message.capitalize()}.\n{agent}: {history_response}." for history_message, history_response in history[-len(history) + history_take:])
    while len(history_context.split()) > MAX_LENGTH - (generation_kwargs["max_length"] + context_length):
        history_take += 1
        history_context = "\n".join(f"{user}: {history_message.capitalize()}.\n{agent}: {history_response}." for history_message, history_response in history[-len(history) + history_take:])
        if history_take >= MAX_LENGTH:
            break
    context += history_context
    for _ in range(5):
        prompt = f"{context}\n\n{user}: {message}.\n"
        response = generate(prompt, context_length=context_length, **generation_kwargs)
        if DEBUG:
            print("\n-----\n" + response + "\n-----\n")
        # response = response.split("\n")[-1]
        # if agent in response and response.split(agent)[-1]:
        #     response = response.split(agent)[-1]
        # if user in response and response.split(user)[-1]:
        #     response = response.split(user)[-1]
        # Take the first response
        response = [
            r for r in response.replace(prompt, "").split(f"{AGENT}:") if r.strip()
        ][0].split(USER)[0].replace(f"{AGENT}:", "\n").strip()
        if response and response[0] in string.punctuation:
            response = response[1:].strip()
        if response.strip().startswith(f"{user}: {message}"):
            response = response.strip().split(f"{user}: {message}")[-1]
        if response.replace(".", "").strip() and message.replace(".", "").strip() != response.replace(".", "").strip():
            break
    if DEBUG:
        print()
        print("CONTEXT:")
        print(context)
        print()
        print("MESSAGE")
        print(message)
        print()
        print("RESPONSE:")
        print(response)
    if not response.strip():
        response = "Lo siento, no puedo hablar ahora" if lang.lower() == "Spanish" else "Sorry, can't talk right now"
    history.append((user_message, response))
    return history, history, speak(response), format_chat(history)


with gr.Blocks() as demo:
    gr.Markdown(HEADER)
    lang = gr.Radio(label="Language", choices=["English", "Spanish"], value=DEFAULT_LANG, type="value")
    AGENT, USER, CONTEXT = select_lang_vars(DEFAULT_LANG)
    context = gr.Textbox(label="Context", lines=5, value=CONTEXT)
    with gr.Row():
        audio_in = gr.Audio(label="User", source="microphone", type="filepath")
        audio_out = gr.Audio(label="Agent", interactive=False, value=empty_audio)
        # chat_btn = gr.Button("Submit")
    with gr.Row():
        user = gr.Textbox(label="User", value=USER)
        agent = gr.Textbox(label="Agent", value=AGENT)
    lang.change(select_lang_vars, inputs=[lang], outputs=[agent, user, context])
    history = gr.Variable(value=[])
    chatbot = gr.Variable()  # gr.Chatbot(color_map=("green", "gray"), visible=False)
    # chat_btn.click(chat_with_gpt, inputs=[lang, agent, user, context, audio_in, history], outputs=[chatbot, history, audio_out])
    log = gr.HTML()
    audio_in.change(chat_with_gpt, inputs=[lang, agent, user, context, audio_in, history], outputs=[chatbot, history, audio_out, log])
    gr.Markdown(FOOTER)

demo.launch()