supported_models.py

from typing import Generator, Set, Union, List, Optional

import requests
from bs4 import BeautifulSoup, Tag, NavigableString, PageElement
from concurrent.futures import ThreadPoolExecutor, as_completed

SUPPORTED_MODEL_NAME_PAGES_FORMAT = "https://huggingface.co/models?pipeline_tag=text-generation&library=pytorch"
MAX_WORKERS = 10
BLACKLISTED_MODEL_NAMES = {
    "ykilcher/gpt-4chan",
    "bigscience/mt0-xxl",
    "bigscience/mt0-xl",
    "bigscience/mt0-large",
    "bigscience/mt0-base",
    "bigscience/mt0-small",
}
BLACKLISTED_ORGANIZATIONS = {
    "huggingtweets"
}
DEFAULT_MIN_NUMBER_OF_DOWNLOADS = 100
DEFAULT_MIN_NUMBER_OF_LIKES = 20


def get_model_name(model_card: Tag) -> str:
    """returns the model name from the model card tag"""
    h4_class = "text-md truncate font-mono text-black dark:group-hover:text-yellow-500 group-hover:text-indigo-600"
    h4_tag = model_card.find("h4", class_=h4_class)
    return h4_tag.text


def is_a_number(element: Union[PageElement, Tag]) -> bool:
    """returns True if the element is a number, False otherwise"""
    if isinstance(element, Tag):
        return False
    text = element.text
    lowered_text = text.strip().lower()
    no_characters_text = lowered_text.replace("k", "").replace("m", "").replace("b", "")
    element = no_characters_text.replace(",", "").replace(".", "")
    try:
        float(element)
    except ValueError:
        return False
    return True


def get_numeric_contents(model_card: Tag) -> List[PageElement]:
    """returns the number of likes and downloads from the model card tag it they exist in the model card"""
    div: Union[Tag | NavigableString] = model_card.find(
        "div",
        class_="mr-1 flex items-center overflow-hidden whitespace-nowrap text-sm leading-tight text-gray-400",
        recursive=True
    )
    contents: List[PageElement] = div.contents
    number_contents: List[PageElement] = [content for content in contents if is_a_number(content)]
    return number_contents


def convert_to_int(element: PageElement) -> int:
    """converts the element to an int"""
    element_str = element.text.strip().lower()
    if element_str.endswith("k"):
        return int(float(element_str[:-1]) * 1_000)
    elif element_str.endswith("m"):
        return int(float(element_str[:-1]) * 1_000_000)
    elif element_str.endswith("b"):
        return int(float(element_str[:-1]) * 1_000_000_000)
    return int(element_str)


def get_page(page_index: int) -> Optional[BeautifulSoup]:
    """returns the page with the given index if it exists, None otherwise"""
    curr_page_url = f"{SUPPORTED_MODEL_NAME_PAGES_FORMAT}&p={page_index}"
    response = requests.get(curr_page_url)
    if response.status_code == 200:
        soup = BeautifulSoup(response.content, "html.parser")
        return soup
    return None


def card_filter(
        model_card: Tag,
        model_name: str,
        min_number_of_downloads: int,
        min_number_of_likes: int,
) -> bool:
    """returns True if the model card is valid, False otherwise"""
    if model_name in BLACKLISTED_MODEL_NAMES:
        return False
    organization = model_name.split("/")[0]
    if organization in BLACKLISTED_ORGANIZATIONS:
        return False
    numeric_contents = get_numeric_contents(model_card)
    if len(numeric_contents) < 2:
        # If the model card doesn't have at least 2 numeric contents,
        # It means that he doesn't have any downloads/likes, so it's not a valid model card.
        return False
    number_of_downloads = convert_to_int(numeric_contents[0])
    if number_of_downloads < min_number_of_downloads:
        return False
    number_of_likes = convert_to_int(numeric_contents[1])
    if number_of_likes < min_number_of_likes:
        return False
    return True


def get_model_names(
        soup: BeautifulSoup,
        min_number_of_downloads: int,
        min_number_of_likes: int,
) -> Generator[str, None, None]:
    """Scrapes the model names from the given soup"""
    model_cards: List[Tag] = soup.find_all("article", class_="overview-card-wrapper group", recursive=True)
    for model_card in model_cards:
        model_name = get_model_name(model_card)
        if card_filter(
                model_card=model_card,
                model_name=model_name,
                min_number_of_downloads=min_number_of_downloads,
                min_number_of_likes=min_number_of_likes
        ):
            yield model_name


def generate_supported_model_names(
        min_number_of_downloads: int,
        min_number_of_likes: int,
) -> Generator[str, None, None]:
    with ThreadPoolExecutor(max_workers=MAX_WORKERS) as executor:
        future_to_index = {executor.submit(get_page, index): index for index in range(300)}
        for future in as_completed(future_to_index):
            soup = future.result()
            if soup:
                yield from get_model_names(
                    soup=soup,
                    min_number_of_downloads=min_number_of_downloads,
                    min_number_of_likes=min_number_of_likes,
                )


def get_supported_model_names(
        min_number_of_downloads: int = DEFAULT_MIN_NUMBER_OF_DOWNLOADS,
        min_number_of_likes: int = DEFAULT_MIN_NUMBER_OF_LIKES,
) -> Set[str]:
    return set(generate_supported_model_names(
        min_number_of_downloads=min_number_of_downloads,
        min_number_of_likes=min_number_of_likes,
    ))


if __name__ == "__main__":
    supported_model_names = get_supported_model_names(1, 1)
    print(f"Number of supported model names: {len(supported_model_names)}")
    print(f"Supported model names: {supported_model_names}")