huggingface/user.py

from dataclasses import dataclass
from huggingface_hub import HfApi, hf_hub_download, upload_file
import pyarrow.parquet as pq
from datetime import datetime, timezone
import tempfile
import hashlib
import toml
import os

_ROOT = os.path.dirname(os.path.abspath(__file__))
ROOT = os.path.dirname(_ROOT)

GR_CONF = {
    "title": "OpenVideo",
    "css": f"{ROOT}/assets/design.css",
    "js": f"{ROOT}/assets/design.js",
    "theme": f"{ROOT}/assets/theme.json",
}

@dataclass
class HFUser:
    token: str
    name: str
    avatar: str
    hf_api: HfApi

    repo: str = None

    @classmethod
    def from_token(cls, _token: str) -> "HFUser":
        _api = HfApi(token=_token, endpoint="https://huggingface.co")
        raw = _api.whoami()
        _name = raw["name"]
        _avatar = f"https://huggingface.co{raw["avatarUrl"]}"
        return cls(token=_token, name=_name, avatar=_avatar, hf_api=_api)

    @classmethod
    def empty(cls) -> "HFUser":
        return cls(token=None, name=None, avatar=None, hf_api=None)

    def auth(self, _token: str) -> "HFUser":
        _api = HfApi(token=_token, endpoint="https://huggingface.co")
        raw = _api.whoami()

        self.token = _token
        self.hf_api = _api
        self.name = raw["name"]
        self._avatar = f"https://huggingface.co{raw["avatarUrl"]}"
        return self.name

    def ping(self, name) -> str:
        if(name == self.name):
            _utc = datetime.now(timezone.utc)
            return _utc.strftime('%Y-%m-%d')
        return "Unauthorized OP"

    def list_dataset(self, repo, path="data"):
        self.repo = repo
        _raw = self.hf_api.list_repo_files(repo, repo_type="dataset")
        if repo == "OpenVideo/Sample-2k":
            files = filter(lambda f: f.endswith(".mp4"), _raw)
        else:
            files = filter(lambda f: f.startswith(path + os.sep), _raw)
        return list(files)

    def fetch_file(self, fname, local_dir="/dev/shm"):
        os.environ["TOKIO_WORKER_THREADS"] = "8"
        os.environ["HF_HUB_ENABLE_HF_TRANSFER"] = "1"
        file = hf_hub_download(repo_id=self.repo, filename=fname, repo_type="dataset", local_dir=local_dir, token=self.token)
        return file

    def split_parquet(self, pq_path, batch_size=1):
        tasks = []
        print(pq_path, batch_size)
        pf = pq.ParquetFile(pq_path)
        for batch in pf.iter_batches(batch_size):
            _chunk = []
            df = batch.to_pandas()
            for binary in df["video"]:
                if(binary):
                    _v = tempfile.NamedTemporaryFile(suffix=".mp4", delete=False)
                    with open(_v.name, "wb") as f:
                        _ = f.write(binary)
                    _chunk.append(_v.name)
            tasks.append(_chunk)
            del df
        return tasks

    def fetch_file_slice(self, fname, batch_size):
        pq_path = self.fetch_file(self.repo, fname, local_dir="/tmp")
        file_slice = self.split_parquet(pq_path, batch_size)
        return file_slice

    def dump_video(self, binary):
        vt = tempfile.NamedTemporaryFile(suffix=".mp4")
        with open(vt.name, "wb") as f:
            _ = f.write(binary)
        return vt.name
    
    def bytes_md5(self, binary):
        md5_hash = hashlib.md5(binary).hexdigest()
        return md5_hash

    def tag(self, binary, tags):
        md5_hash = self.bytes_md5(binary)
        _info = {
            "md5": md5_hash,
            "tags": tags,
            "operator_name": self.name,
        }
        return _info

    def push_tag(self, info, target_repo):
        md5 = info["md5"]
        _raw = f"raw/{md5}/{md5}.mp4"
        idx_path = f"index/{md5}/{md5}.toml"

        with open(idx_path, "w") as f:
            toml.dump(info, f)

        _repo = upload_file(
            path_or_fileobj=idx_path,
            path_in_repo=idx_path,
            repo_id = target_repo,
            token=self.token, 
            repo_type="dataset"
        )
        return _repo

if __name__ == "__main__":
    u = HFUser.from_token("hf_xxxxx")
    _patch = u.split_parquet("/opt/000000.parquet", batch_size=4)