|
| 1 | +import io |
| 2 | +from typing import Any |
| 3 | +import shutil |
| 4 | +from videohash import VideoHash |
| 5 | +from uuid import uuid4 |
| 6 | +import os |
| 7 | + |
| 8 | +from DPF.types import ModalityToDataMapping |
| 9 | + |
| 10 | +from .video_filter import VideoFilter |
| 11 | + |
| 12 | + |
| 13 | + |
| 14 | +class VideohashFilter(VideoFilter): |
| 15 | + |
| 16 | + def __init__( |
| 17 | + self, |
| 18 | + workers: int = 16, |
| 19 | + pbar: bool = True, |
| 20 | + _pbar_position: int = 0 |
| 21 | + ): |
| 22 | + super().__init__(pbar, _pbar_position) |
| 23 | + self.num_workers = workers |
| 24 | + |
| 25 | + @property |
| 26 | + def result_columns(self) -> list[str]: |
| 27 | + return [f"video_hash"] |
| 28 | + |
| 29 | + @property |
| 30 | + def dataloader_kwargs(self) -> dict[str, Any]: |
| 31 | + return { |
| 32 | + "num_workers": self.num_workers, |
| 33 | + "batch_size": 1, |
| 34 | + "drop_last": False, |
| 35 | + } |
| 36 | + |
| 37 | + def preprocess_data( |
| 38 | + self, |
| 39 | + modality2data: ModalityToDataMapping, |
| 40 | + metadata: dict[str, Any] |
| 41 | + ) -> Any: |
| 42 | + key = metadata[self.key_column] |
| 43 | + video_file = modality2data['video'] |
| 44 | + |
| 45 | + uid = str(uuid4()) |
| 46 | + tmp_dir = os.path.join(os.path.abspath(os.getcwd()), uid) + os.path.sep |
| 47 | + os.makedirs(tmp_dir, exist_ok=True) |
| 48 | + |
| 49 | + video_path = os.path.join(tmp_dir, 'video.mp4') |
| 50 | + with open(video_path, 'wb') as f: |
| 51 | + f.write(video_file) |
| 52 | + |
| 53 | + hash_obj = VideoHash(path=video_path, storage_path=tmp_dir) |
| 54 | + shutil.rmtree(hash_obj.storage_path) |
| 55 | + os.remove(video_path) |
| 56 | + os.rmdir(tmp_dir) |
| 57 | + return key, hash_obj.hash_hex |
| 58 | + |
| 59 | + def process_batch(self, batch: list[Any]) -> dict[str, list[Any]]: |
| 60 | + df_batch_labels = self._get_dict_from_schema() |
| 61 | + for key, hash_hex in batch: |
| 62 | + df_batch_labels[self.key_column].append(key) |
| 63 | + df_batch_labels[self.result_columns[0]].append(hash_hex) |
| 64 | + return df_batch_labels |
0 commit comments