app.py

'''
This script calls the model from openai api to predict the next few words in a conversation.
'''
import os
import sys
import openai
import gradio as gr
os.system("pip install git+https://github.com/openai/whisper.git")
import whisper
from transformers import pipeline
import torch
from transformers import AutoModelForCausalLM
from transformers import AutoTokenizer
import time
import pandas as pd


EXAMPLE_PROMPT = """This is a tool for helping someone with memory issues remember the next word. 
The predictions follow a few rules:
1) The predictions are suggestions of ways to continue the transcript as if someone forgot what the next word was.
2) The predictions do not repeat themselves.
3) The predictions focus on suggesting nouns, adjectives, and verbs.
4) The predictions are related to the context in the transcript.
5) The predictions are ordered from most likely to least likely.
6) Five unique predictions are made per transcript.
    
EXAMPLES:
Transcript: Tomorrow night we're going out to 
Prediction: The Movies, A Restaurant, A Baseball Game, The Theater, A Party for a friend   
Transcript: I would like to order a cheeseburger with a side of
Prediction: French fries, Milkshake, Apple slices, Side salad, Extra catsup 
Transcript: My friend Savanah is
Prediction: An electrical engineer, A marine biologist, A classical musician, A developer, A product manager 
Transcript: I need to buy a birthday
Prediction: Present, Gift, Cake, Card, balloon
Transcript:  """


# whisper model specification
asr_model = whisper.load_model("tiny")

openai.api_key = os.environ["Openai_APIkey"]

    
# Transcribe function
def transcribe(audio_file):
    print("Transcribing")
    transcription = asr_model.transcribe(audio_file)["text"]
    #transcription = asr_model.transcribe(audio_file)
    return transcription

def inference(audio, latest):
    # Transcribe with Whisper
    print("The audio is:", audio)
    transcript = transcribe(audio)

    if transcript != None:
        latest.append(transcript)
        #tscript = EXAMPLE_PROMPT + str(transcript) + "\nPrediction: "
        tscript = EXAMPLE_PROMPT + latest + "\nPrediction: "
    else: tscript = EXAMPLE_PROMPT

        
    print("tscript ------- ", tscript)
    
    response = openai.Completion.create(
                        model="text-davinci-003",
                        prompt=tscript,
                        temperature=0.8,
                        max_tokens=18,
                        n=5)

    #infers = []
    #infers = []
    temp = []
    inferred=[]
    
    for i in range(5):
        print("print1 ", response['choices'][i]['text'])
        temp.append(response['choices'][i]['text'])
        print("print2: infers ", inferred)
        print("print3: Responses ", response)
        print("Object type of response: ", type(response))
        #infered = list(map(lambda x: x.split(',')[0], infers))
        #print("Infered type is: ", type(infered))
        inferred = list(map(lambda x: x.replace("\n", ""), temp))
        #infered = list(map(lambda x: x.split(','), infers))

    infers = pd.Series(inferred)
    infersNew = infers.str.split(",", n=-1, expand=True) 
    print("USEAGE: ", response['usage']['completion_tokens'])

    #infers.drop_duplicates(keep='first', inplace=True)
    print("Infers DataType ", type(infers), "Infers after drop: ", infers, "Infers at 0: ", infers[0])
    res = []
    
    op1 = infersNew[0][0]
    op2 = infersNew[1][0]
    op3 = infersNew[2][0]
    try:
        op4 = infersNew[3][0]
    except KeyError:
         op4 = infersNew[0][1]
    try:
        op5 = infersNew[4][0]
    except KeyError:
         op5 = infersNew[1][1]
    
    
    print("INFERS TYPE: ", type(infers), "INFERS ", infers)

    convoState = latest
    #infersStr = str(infers)
        

    return transcript, op1, op2, op3, op4, op5, convoState
    
def appendPrediction(val, convoState):
    convoState.append(val)
    return convoState
    
# get audio from microphone 
with gr.Blocks() as face:
    
    with gr.Row():
        convoState = gr.State([""])
        with gr.Column():
            audio = gr.Audio(source="microphone", type="filepath")
            #promptText = gr.Textbox(lines=15, placeholder="Enter a prompt here")
            #dropChoice = gr.Dropdown(choices=["text-ada-001", "text-davinci-002", "text-davinci-003", "gpt-3.5-turbo"], label="Model")
            #sliderChoice = gr.Slider(minimum=0.0, maximum=1.0, default=0.8, step=0.1, label="Temperature")
            transcribe_btn = gr.Button(value="Transcribe")
        with gr.Column():
            script = gr.Textbox(label="Transcribed text")
            #options = gr.Textbox(label="Predictions")
            option1 = gr.Button(value="  ")
            option2 = gr.Button(value="  ")
            option3 = gr.Button(value="  ")
            option4 = gr.Button(value="  ")
            option5 = gr.Button(value="  ")
            #options = gr.Dataset(components=[gr.Radio], samples=["One", "Two", "Three", "Four", "Five"])
            '''options = gr.Dataset(components=[gr.Textbox(visible=False)],
                label="Text Dataset",
                samples=[
                ["One"],
                ["Two"],
                ["Three"],
                ["Four"],
                ["Five"],
                ],
            )'''
            #options = gr.Radio(choices=["One", "Two", "Three", "Four", "Five"])
            latestConvo = gr.Textbox(label="Running conversation")
            #transcribe_btn.click(inference)
    transcribe_btn.click(fn=inference, inputs=[audio, convoState], outputs=[script, option1, option2, option3, option4, option5, latestConvo])
    option1.click(fn=appendPrediction, inputs=[option1, convoState], outputs=[latestConvo])
    option2.click(fn=appendPrediction, inputs=[option2, convoState], outputs=[latestConvo])
    option3.click(fn=appendPrediction, inputs=[option3, convoState], outputs=[latestConvo])
    option4.click(fn=appendPrediction, inputs=[option4, convoState], outputs=[latestConvo])
    option5.click(fn=appendPrediction, inputs=[option5, convoState], outputs=[latestConvo])
    #examples = gr.Examples(examples=["Sedan, Truck, SUV", "Dalmaion, Shepherd, Lab, Mutt"], inputs=[options])


face.launch()