vi_backbones.py

import os
import re
import hashlib
import requests
import datasets
import torchvision.models as models
from bs4 import BeautifulSoup


_DBNAME = os.path.basename(__file__).split('.')[0]

_HOMEPAGE = "https://huggingface.co/datasets/george-chou/" + _DBNAME

_URL = 'https://pytorch.org/vision/main/_modules/'


class vi_backbones(datasets.GeneratorBasedBuilder):

    def _info(self):
        return datasets.DatasetInfo(
            features=datasets.Features(
                {
                    "ver": datasets.Value("string"),
                    "type": datasets.Value("string"),
                    "input_size": datasets.Value("int16"),
                    "output_size": datasets.Value("int64"),
                    "url": datasets.Value("string"),
                    # "md5": datasets.Value("string"),
                }
            ),
            supervised_keys=("ver", "type"),
            homepage=_HOMEPAGE,
            license="mit"
        )

    def _get_file_md5(self, url):
        """
        Calculate the MD5 hash value of a file using its URL

        :param url: the URL address of the file
        :return: the MD5 hash value in string format
        """
        try:
            response = requests.get(url, stream=True)
            if response.status_code == 200:
                md5obj = hashlib.md5()
                for chunk in response.iter_content(chunk_size=1024*1024):
                    md5obj.update(chunk)
                return md5obj.hexdigest()
            else:
                raise ValueError(
                    f"Error downloading file from {url}. Status code: {response.status_code}")
        except Exception as e:
            raise ValueError(
                f"Error calculating MD5 of file at {url}: {str(e)}")

    def _parse_url(self, url):
        response = requests.get(url)
        html = response.text
        return BeautifulSoup(html, 'html.parser')

    def _special_type(self, m_type):
        if m_type == 'wide' or m_type == 'resnext':
            return 'resnet'

        if m_type == 'swin':
            return 'swin_transformer'

        if m_type == 'inception':
            return 'googlenet'

        return m_type

    def _info_on_dataset(self, m_ver, m_type, in1k_span):
        url_span = in1k_span.find_next_sibling('span', {'class': 's2'})
        size_span = url_span.find_next_sibling('span', {'class': 'mi'})
        m_type = self._special_type(m_type)
        m_url = str(url_span.text[1:-1])
        input_size = int(size_span.text)
        m_dict = {
            'ver': m_ver,
            'type': m_type,
            'input_size': input_size,
            'url': m_url
        }
        return m_dict, size_span

    def _generate_dataset(self, url):
        torch_page = self._parse_url(url)
        article = torch_page.find('article', {'id': 'pytorch-article'})
        ul = article.find('ul').find('ul')
        in1k_v1, in1k_v2 = [], []

        for li in ul.find_all('li'):
            name = str(li.text)
            if name.__contains__('torchvision.models.') and len(name.split('.')) == 3:

                if name.__contains__('_api') or name.__contains__('feature_extraction'):
                    continue

                href = li.find('a').get('href')
                model_page = self._parse_url(url + href)
                divs = model_page.select('div.viewcode-block')

                for div in divs:
                    div_id = str(div['id'])
                    if div_id.__contains__('_Weights'):
                        m_ver = div_id.split('_Weight')[0].lower()

                        if not hasattr(models, m_ver):
                            continue

                        m_type = re.search('[a-zA-Z]+', m_ver).group(0)

                        in1k_v1_span = div.find(
                            name='span',
                            attrs={'class': 'n'},
                            string='IMAGENET1K_V1'
                        )

                        if in1k_v1_span == None:
                            continue

                        m_dict, size_span = self._info_on_dataset(
                            m_ver,
                            m_type,
                            in1k_v1_span
                        )
                        in1k_v1.append(m_dict)

                        in1k_v2_span = size_span.find_next_sibling(
                            name='span',
                            attrs={'class': 'n'},
                            string='IMAGENET1K_V2'
                        )

                        if in1k_v2_span != None:
                            m_dict, _ = self._info_on_dataset(
                                m_ver,
                                m_type,
                                in1k_v2_span
                            )
                            in1k_v2.append(m_dict)

        return in1k_v1, in1k_v2

    def _split_generators(self, _):
        in1k_v1, in1k_v2 = self._generate_dataset(_URL)

        return [
            datasets.SplitGenerator(
                name="IMAGENET1K_V1",
                gen_kwargs={
                    "subset": in1k_v1,
                },
            ),
            datasets.SplitGenerator(
                name="IMAGENET1K_V2",
                gen_kwargs={
                    "subset": in1k_v2,
                },
            ),
        ]

    def _generate_examples(self, subset):
        for i, model in enumerate(subset):
            yield i, {
                "ver": model['ver'],
                "type": model['type'],
                "input_size": model['input_size'],
                "output_size": 1234,
                "url": model['url'],
                # "md5": self._get_file_md5(model['url']),
            }