ann-unsplash-25k / README.md
Ashot Vardanian
Add: Dataset validation and docs
95669d7
|
raw
history blame
873 Bytes
metadata
license: apache-2.0

25K Unsplash Images for Search

This is a derivative work based on two existing datasets.

  • images.csv metadata from Unsplash, sorted and converted to CSV.
  • images/ in 250x250 resolution by kaggle/@jettchentt.
  • images.fbin is a binary file with UForm image embeddings.
  • images.usearch is a binary file with a serialized USearch index.

The original images.tsv from Unsplash has been filtered to avoid missing images. The embeddings and the index can be reconstructed with the main.py script. On the Apple M2 Pro CPU:

  • Image vectorization takes 100ms/image, or 10 inferences/second.
  • Indexing vectors one-by-one happens at 700 vectors/second speed.

To rebuild the indexes:

./main.py