Skip to content

Latest commit

 

History

History
17 lines (12 loc) · 1.39 KB

index_search.md

File metadata and controls

17 lines (12 loc) · 1.39 KB
Raw

Extracting features & creating a faiss index

  1. conda env create
  2. conda install -c pytorch faiss-cpu
  3. To extract features for a directory containing images (as well as subdirectory of images):
    • Edit the field [dataloader][args][data_dir] field in the config file ./configs/image_feat_extract.json or ./configs/video_feat_extract.json for image or video data respectively. N.B. the code assumes all images,videos are .jpg,.mp4, if this is not the case you can edit the file data_loader/ImageDirectory_dataset.py
    • Edit the field "load_checkpoint" to the path of the downloaded pretrained model.
  4. then run feature extraction python test.py --config configs/image_feat_extract.json --save_feats SAVE_FTR_DIR --save_type video. Assumes there is a gpu available.

csv file(s) and numpy array(s) are saved to SAVE_FTR_PATH. When there's more than 1000 images/videos it creates multiple files to avoid OOM. In this case we need aggregate these into a single numpy array and single csv file:

  1. python scripts/agg_ids_embeds.py --path_to_saved_features SAVE_FTR_DIR

Next, create a FAISS index from the single numpy array and csv file. If your images/videos have unique integer IDs then we can create and index with ids (so you dont need to carry around the csv mapping, use the --add_index_with_ids in this case):

  1. python scripts/create_faiss_index.py --path_to_saved_features SAVE_FTR_DIR