app.py

import time

import torch
from peft import PeftModel, PeftConfig
from transformers import AutoModelForCausalLM, AutoTokenizer, AutoModelForSeq2SeqLM

import gradio as gr
import speech_recognition as sr
from math import log2, pow
import os

#from scipy.fftpack import fft
import gc

peft_model_id='hackathon-somos-nlp-2023/T5unami-small-v1'

config = PeftConfig.from_pretrained(peft_model_id)
model2 = AutoModelForSeq2SeqLM.from_pretrained(config.base_model_name_or_path, return_dict=True, 
                                               # load_in_8bit=True, 
                                               # load_in_8bit_fp32_cpu_offload=True,
                                               device_map='auto')
tokenizer2 = AutoTokenizer.from_pretrained(peft_model_id)

model2 = PeftModel.from_pretrained(model2, peft_model_id)

Problema_tarjetaCredito= os.path.abspath("Problema_tarjetaCredito.ogg")
list_audios= [[Problema_tarjetaCredito]]

def gen_conversation(text,max_new_tokens=100): 
  text = "<SN>instruction: " + text + "\n "
  batch = tokenizer2(text, return_tensors='pt')
  
  output_tokens = model2.generate(**batch, 
                                    max_new_tokens=max_new_tokens, 
                                    eos_token_id= tokenizer2.eos_token_id,
                                    pad_token_id= tokenizer2.pad_token_id,
                                    bos_token_id= tokenizer2.bos_token_id,
                                    early_stopping = True, 
                                    no_repeat_ngram_size=2,
                                    repetition_penalty=1.2,
                                    temperature=.69,
                                    num_beams=3
                                    )
  gc.collect()
  return tokenizer2.decode(output_tokens[0], skip_special_tokens=True).split("\n")[-1].replace("output:","")

conversacion = ""
def speech_to_text(audio_file, texto_adicional):
    global conversacion
    if audio_file is not None:
        # Lógica para entrada de audio
        r = sr.Recognizer()
        audio_data = sr.AudioFile(audio_file)
        with audio_data as source:
            audio = r.record(source)
        text_enrada=""
        
        texto_generado = r.recognize_google(audio, language="es-ES")
        texto_generado= f"[|Audio a texto|]:{texto_generado}\n" + "<br>[AGENTE]:"+gen_conversation(texto_generado,max_new_tokens=50)
        texto_generado = "<div style='color: #66b3ff;'>" + texto_generado + "</div><br>"
    else:
        texto_generado= f"[|Solo texto|]:{texto_adicional}\n" + "<br>[AGENTE]:"+gen_conversation(texto_adicional,max_new_tokens=50)
        texto_generado = "<div style='color: #66b3ff;'> " + texto_generado + "</div><br>"
    conversacion += texto_generado
    return conversacion

iface = gr.Interface(
    fn=speech_to_text,
    inputs=[gr.inputs.Audio(label="Voz", type="filepath"), gr.inputs.Textbox(label="Texto adicional")],
    outputs=gr.outputs.HTML(label=["chatbot","state"]),
    title="Chat bot para empresas.",
    description="Este modelo convierte la entrada de voz a texto e inferencia, texto a inferencia",
    examples=list_audios,
    theme="default",
    layout="vertical",
    allow_flagging=False,
    flagging_dir=None,
    server_name=None,
    server_port=None,
    live=False,
    capture_session=False
)

iface.launch()