app.py

"""
app.py - the main file for the app. This creates the flask app and handles the routes.

"""

import torch
from transformers import pipeline
from cleantext import clean
from pathlib import Path
import warnings
import time
import argparse
import logging
import gradio as gr
import os
import sys
from os.path import dirname
import nltk
from converse import discussion
from grammar_improve import (
    detect_propers,
    load_ns_checker,
    neuspell_correct,
    remove_repeated_words,
    remove_trailing_punctuation,
    build_symspell_obj,
    symspeller,
    fix_punct_spacing,
)

from utils import (
    cleantxt_wrap,
    corr,
)

nltk.download("stopwords")  # TODO: find where this requirement originates from

sys.path.append(dirname(dirname(os.path.abspath(__file__))))
warnings.filterwarnings(action="ignore", message=".*gradient_checkpointing*")
import transformers

transformers.logging.set_verbosity_error()
logging.basicConfig()
cwd = Path.cwd()
my_cwd = str(cwd.resolve())  # string so it can be passed to os.path() objects


def chat(trivia_query):
    """
    chat - helper function that makes the whole gradio thing work.

    Args:
        trivia_query (str): the question to ask the bot

    Returns:
        [str]: the bot's response
    """
    history = []
    response = ask_gpt(message=trivia_query, chat_pipe=my_chatbot)
    history = [trivia_query, response]
    html = ""
    for item in history:
        html += f"<b>{item}</b> <br>"

    html += ""

    return html


def ask_gpt(
    message: str,
    chat_pipe,
    speaker="person alpha",
    responder="person beta",
    max_len=196,
    top_p=0.95,
    top_k=50,
    temperature=0.6,
):
    """

    ask_gpt - a function that takes in a prompt and generates a response using the pipeline. This interacts the discussion function.

    Parameters:
        message (str): the question to ask the bot
        chat_pipe (str): the chat_pipe to use for the bot (default: "pszemraj/Ballpark-Trivia-XL")
        speaker (str): the name of the speaker (default: "person alpha")
        responder (str): the name of the responder (default: "person beta")
        max_len (int): the maximum length of the response (default: 128)
        top_p (float): the top probability threshold (default: 0.95)
        top_k (int): the top k threshold (default: 50)
        temperature (float): the temperature of the response (default: 0.7)
    """

    st = time.perf_counter()
    prompt = clean(message)  # clean user input
    prompt = prompt.strip()  # get rid of any extra whitespace
    in_len = len(prompt)
    if in_len > 512:
        prompt = prompt[-512:]  # truncate to 512 chars
        print(f"Truncated prompt to last 512 chars: started with {in_len} chars")
        max_len = min(max_len, 512)

    resp = discussion(
        prompt_text=prompt,
        pipeline=chat_pipe,
        speaker=speaker,
        responder=responder,
        top_p=top_p,
        top_k=top_k,
        temperature=temperature,
        max_length=max_len,
    )
    gpt_et = time.perf_counter()
    gpt_rt = round(gpt_et - st, 2)
    rawtxt = resp["out_text"]
    # check for proper nouns
    if basic_sc and not detect_propers(rawtxt):
        cln_resp = symspeller(rawtxt, sym_checker=schnellspell)
    elif not detect_propers(rawtxt):
        cln_resp = neuspell_correct(rawtxt, checker=ns_checker)
    else:
        # no correction needed
        cln_resp = rawtxt.strip()
    bot_resp_a = corr(remove_repeated_words(cln_resp))
    bot_resp = fix_punct_spacing(bot_resp_a)
    print(f"the prompt was:\n\t{message}\nand the response was:\n\t{bot_resp}\n")
    corr_rt = round(time.perf_counter() - gpt_et, 4)
    print(
        f"took {gpt_rt + corr_rt} sec to respond, {gpt_rt} for GPT, {corr_rt} for correction\n"
    )
    return remove_trailing_punctuation(bot_resp)


def get_parser():
    """
    get_parser - a helper function for the argparse module
    """
    parser = argparse.ArgumentParser(
        description="submit a question, GPT model responds"
    )
    parser.add_argument(
        "-m",
        "--model",
        required=False,
        type=str,
        default="pszemraj/GPT-Converse-1pt3B-Neo-WoW-DD-17",  # default model
        help="the model to use for the chatbot on https://huggingface.co/models OR a path to a local model",
    )
    parser.add_argument(
        "--basic-sc",
        required=False,
        default=True, # TODO: change this back to False once Neuspell issues are resolved.
        action="store_true",
        help="turn on symspell (baseline) correction instead of the more advanced neural net models",
    )

    parser.add_argument(
        "--verbose",
        action="store_true",
        default=False,
        help="turn on verbose logging",
    )
    return parser


if __name__ == "__main__":
    args = get_parser().parse_args()
    default_model = str(args.model)
    model_loc = Path(default_model)  # if the model is a path, use it
    basic_sc = args.basic_sc  # whether to use the baseline spellchecker
    device = 0 if torch.cuda.is_available() else -1
    print(f"CUDA avail is {torch.cuda.is_available()}")

    my_chatbot = (
        pipeline("text-generation", model=model_loc.resolve(), device=device)
        if model_loc.exists() and model_loc.is_dir()
        else pipeline("text-generation", model=default_model, device=device)
    )  # if the model is a name, use it. stays on CPU if no GPU available
    print(f"using model {my_chatbot.model}")

    if basic_sc:
        print("Using the baseline spellchecker")
        schnellspell = build_symspell_obj()
    else:
        print("using Neuspell spell checker")
        ns_checker = load_ns_checker(fast=False)

    print(f"using model stored here: \n {model_loc} \n")
    iface = gr.Interface(
        chat,
        inputs=["text"],
        outputs="html",
        examples_per_page=10,
        examples=[
            "How can you help me?",
            "what can you do?",
            "Hi, my name is……",
            "Happy birthday!",
            "I have a question, can you help me?",
            "Do you know a joke?",
            "Will you marry me?",
            "Are you single?",
            "Do you like people?",
            "Are you part of the Matrix?",
            "Do you have a hobby?",
            "You’re clever",
            "Tell me about your personality",
            "You’re annoying",
            "you suck",
            "I want to speak to a human now.",
            "Don’t you speak English?!",
            "Are you human?",
            "Are you a robot?",
            "What is your name?",
            "How old are you?",
            "What’s your age?",
            "What day is it today?",
            "Who made you?",
            "Which languages can you speak?",
            "What is your mother’s name?",
            "Where do you live?",
            "What’s the weather like today?",
            "Are you expensive?",
            "Do you get smarter?",
            "rate your overall satisfaction with the chatbot",
            "How many icebergs are in the ocean?",
        ],
        title=f"NLP template space: {default_model} Model",
        description=f"this space is used as a template. please copy the files in the space to your own space repo, AND THEN edit them ",
        article="here you can add more details about your model. \n\n"
        "**Important Notes & About:**\n\n"
        "1. the model can take up to 60 seconds to respond sometimes, patience is a virtue.\n"
        "2. the model started from a pretrained checkpoint, and was trained on several different datasets. Anything it says should be fact-checked before being regarded as a true statement.\n"
        "3. Some params are still being tweaked (in the future, will be inputs) any feedback is welcome :)\n",
        css="""
            .chatbox {display:flex;flex-direction:column}
            .user_msg, .resp_msg {padding:4px;margin-bottom:4px;border-radius:4px;width:80%}
            .user_msg {background-color:cornflowerblue;color:white;align-self:start}
            .resp_msg {background-color:lightgray;align-self:self-end}
        """,
        allow_screenshot=True,
        allow_flagging="never",
        theme="dark",
    )

    # launch the gradio interface and start the server
    iface.launch(
        # prevent_thread_lock=True,
        enable_queue=True,  # also allows for dealing with multiple users simultaneously (per newer gradio version)
    )