Skip to content

FeatureStore.md

Latest commit

 

History

History
29 lines (22 loc) · 1.34 KB

FeatureStore.md

File metadata and controls

29 lines (22 loc) · 1.34 KB

Feature Store

The FeatureStore provides storage for features extracted from video frames and audio samples. WISE currently supports the following two types of FeatureStore:

  • NumpySaveStore : stores features across multiple .npz files
  • WebdatasetStore : stores features across multiple .tar files

The NumpySaveStore uses numpy.savez() to store extracted features as numpy ndarray in .npz files. The WebdatasetStore uses the Webdataset format to store extracted features across multiple shards maitained as .tar files.

Here is an example of how the features stores in these stores can be accessed.

# source: src/search_index.py
from pathlib import Path
from .feature.store.feature_store_factory import FeatureStoreFactory

...
media_type = 'video' # or 'audio'
feature_dir = Path('/data/projects/Kinetics-7/store/mlfoundations/open_clip/xlm-roberta-large-ViT-H-14/frozen_laion5b_s13b_b90k/features/')
feature_store = FeatureStoreFactory.load_store(media_type, feature_dir)
feature_store.enable_read(shard_shuffle = False)

for feature_id, feature_vector in feature_store:
    print(f'feature_id={feature_id}, feature = {feature_vector.shape}')

...