app.py

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

"""

import argparse
import logging

logging.basicConfig(level=logging.INFO, format="%(asctime)s %(message)s")

import os
import sys
import time
import warnings
from os.path import dirname
from pathlib import Path

import gradio as gr
import nltk
import torch
from cleantext import clean
from gradio.inputs import Slider, Textbox, Radio
from transformers import pipeline

from converse import discussion
from grammar_improve import (
    build_symspell_obj,
    detect_propers,
    fix_punct_spacing,
    load_ns_checker,
    neuspell_correct,
    remove_repeated_words,
    remove_trailing_punctuation,
    symspeller,
    synthesize_grammar,
)
from utils import corr, setup_logging

nltk.download("stopwords")  # download stopwords

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

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


def chat(
    prompt_message,
    temperature: float = 0.5,
    top_p: float = 0.95,
    top_k: int = 20,
    constrained_generation: str = "False",
) -> str:
    """
    chat - the main function for the chatbot. This is the function that is called when the user

    :param _type_ prompt_message: the message to send to the model
    :param float temperature: the temperature value for the model, defaults to 0.6
    :param float top_p: the top_p value for the model, defaults to 0.95
    :param int top_k: the top_k value for the model, defaults to 25
    :param bool constrained_generation: whether to use constrained generation or not, defaults to False
    :return str: the response from the model
    """
    history = []
    response = ask_gpt(
        message=prompt_message,
        chat_pipe=my_chatbot,
        top_p=top_p,
        top_k=top_k,
        temperature=temperature,
        constrained_generation=constrained_generation,
    )
    history = [prompt_message, response]
    html = ""
    for item in history:
        html += f"<b>{item}</b> <br><br>"

    html += ""

    return html


def ask_gpt(
    message: str,
    chat_pipe,
    speaker="person alpha",
    responder="person beta",
    min_length=12,
    max_length=48,
    top_p=0.95,
    top_k=25,
    temperature=0.5,
    constrained_generation=False,
    max_input_length=128,
) -> str:
    """
    ask_gpt - helper function that asks the GPT model a question and returns the response

    :param str message: the question to ask the model
    :param chat_pipe: the pipeline object for the model, created by the pipeline() function
    :param str speaker: the name of the speaker, defaults to "person alpha"
    :param str responder: the name of the responder, defaults to "person beta"
    :param int min_length: the minimum length of the response, defaults to 12
    :param int max_length: the maximum length of the response, defaults to 64
    :param float top_p: the top_p value for the model, defaults to 0.95
    :param int top_k: the top_k value for the model, defaults to 25
    :param float temperature: the temperature value for the model, defaults to 0.6
    :param bool constrained_generation: whether to use constrained generation or not, defaults to False
    :return str: the response from the model
    """
    st = time.perf_counter()
    prompt = clean(message)  # clean user input
    prompt = prompt.strip()  # get rid of any extra whitespace
    in_len = len(chat_pipe.tokenizer(prompt).input_ids)
    if in_len > max_input_length:
        # truncate to last max_input_length tokens
        tokens = chat_pipe.tokenizer(prompt).input_ids
        trunc_tokens = tokens[-max_input_length:]
        prompt = chat_pipe.tokenizer.decode(trunc_tokens)
        print(f"truncated prompt to {len(trunc_tokens)} tokens, input length: {in_len}")
    logging.info(f"prompt: {prompt}")
    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_length,
        min_length=min_length,
        constrained_beam_search=constrained_generation,
    )
    gpt_et = time.perf_counter()
    gpt_rt = round(gpt_et - st, 2)
    rawtxt = resp["out_text"]
    # check for proper nouns
    if basic_sc:
        cln_resp = symspeller(rawtxt, sym_checker=basic_spell)
    else:
        cln_resp = synthesize_grammar(corrector=grammarbot, message=rawtxt)
    bot_resp_a = corr(remove_repeated_words(cln_resp))
    bot_resp = fix_punct_spacing(bot_resp_a)
    corr_rt = round(time.perf_counter() - gpt_et, 4)
    print(f"{gpt_rt + corr_rt} to respond, {gpt_rt} 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="ethzanalytics/ai-msgbot-gpt2-XL",  # 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(
        "-gm",
        "--gram-model",
        required=False,
        type=str,
        default="pszemraj/grammar-synthesis-base",
        help="text2text generation model ID from huggingface for the model to correct grammar",
    )

    parser.add_argument(
        "--basic-sc",
        required=False,
        default=False,
        action="store_true",
        help="use symspell (statistical spelling correction) instead of neural spell correction",
    )

    parser.add_argument(
        "--test",
        action="store_true",
        default=False,
        help="load the smallest model for simple testing (ethzanalytics/distilgpt2-tiny-conversational)",
    )

    parser.add_argument(
        "--verbose",
        action="store_true",
        default=False,
        help="turn on verbose printing",
    )
    parser.add_argument(
        "-q",
        "--quiet",
        dest="loglevel",
        help="set loglevel to WARNING (reduce output)",
        action="store_const",
        const=logging.WARNING,
    )
    parser.add_argument(
        "-vv",
        "--very-verbose",
        dest="loglevel",
        help="set loglevel to DEBUG",
        action="store_const",
        const=logging.DEBUG,
    )
    return parser


if __name__ == "__main__":
    args = get_parser().parse_args()
    loglevel = args.loglevel or logging.INFO
    setup_logging(loglevel)
    logging.info("\n\n\nStarting app.py\n\n\n")
    logging.info(f"args: {args}")

    default_model = str(args.model)

    if args.test:
        logging.info("loading the smallest model for testing")
        default_model = "ethzanalytics/distilgpt2-tiny-conversational"

    model_loc = Path(default_model)  # if the model is a path, use it
    basic_sc = args.basic_sc  # whether to use the baseline spellchecker
    gram_model = str(args.gram_model)
    device = 0 if torch.cuda.is_available() else -1

    logging.info(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
    logging.info(f"using model {my_chatbot.model}")

    if basic_sc:
        logging.info("Using the baseline spellchecker")
        basic_spell = build_symspell_obj()
    else:
        logging.info("using neural spell checker")
        grammarbot = pipeline("text2text-generation", gram_model, device=device)

    logging.debug(f"using model stored here: \n {model_loc} \n")
    iface = gr.Interface(
        chat,
        inputs=[
            Textbox(
                default="Why is everyone here eating chocolate cake?",
                label="prompt_message",
                placeholder="Start a conversation with the bot",
                lines=2,
            ),
            Slider(
                minimum=0.0, maximum=1.0, step=0.05, default=0.4, label="temperature"
            ),
            Slider(minimum=0.0, maximum=1.0, step=0.01, default=0.95, label="top_p"),
            Slider(minimum=0, maximum=100, step=5, default=20, label="top_k"),
            Radio(
                choices=[True, False],
                default=False,
                label="constrained_generation",
            ),
        ],
        outputs="html",
        examples_per_page=8,
        examples=[
            ["Point Break or Bad Boys II?", 0.75, 0.95, 50, False],
            ["So... you're saying this wasn't an accident?", 0.6, 0.95, 40, False],
            ["Hi, my name is Reginald", 0.6, 0.95, 100, False],
            ["Happy birthday!", 0.9, 0.95, 50, False],
            ["I have a question, can you help me?", 0.6, 0.95, 50, False],
            ["Do you know a joke?", 0.8, 0.85, 50, False],
            ["Will you marry me?", 0.9, 0.95, 100, False],
            ["Are you single?", 0.95, 0.95, 100, False],
            ["Do you like people?", 0.7, 0.95, 25, False],
            ["You never took a shortcut before?", 0.7, 0.95, 100, False],
        ],
        title=f"GPT Chatbot Demo: {default_model} Model",
        description=f"A Demo of a Chatbot trained for conversation with humans. Size XL= 1.5B parameters.\n\n"
        "**Important Notes & About:**\n\n"
        "You can find a link to the model card **[here](https://huggingface.co/ethzanalytics/ai-msgbot-gpt2-XL-dialogue)**\n\n"
        "1. responses can take up to 60 seconds to respond sometimes, patience is a virtue.\n"
        "2. the model was trained on several different datasets.  fact-check responses instead of regarding as a true statement.\n"
        "3. Try adjusting the **[generation parameters](https://huggingface.co/blog/how-to-generate)** to get a better understanding of how they work!\n"
        "4. New - try using [constrained beam search](https://huggingface.co/blog/constrained-beam-search) decoding to generate more coherent responses. _(experimental, feedback welcome!)_\n",
        css="""
            .chatbox {display:flex;flex-direction:row}
            .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_flagging="never",
        theme="dark",
    )

    # launch the gradio interface and start the server
    iface.launch(
        enable_queue=True,
    )