Skip to content

visiongrep 0.4.0

Choose a tag to compare

@udayj udayj released this 08 Nov 10:46
· 12 commits to main since this release
1cff617

Image queries and re-indexing are about 2.5× faster with about 65% less memory. The vision model now uses mixed-precision int8 (dynamic quantization, with the final projection and first transformer block kept in fp32).
Measured on AWS c7a.xlarge against v0.3.0:
image query: 871 → 336 ms, peak memory 596 → 207 MiB;
search after 1% of images change: 990 → 401 ms, peak memory 603 → 234 MiB;
text queries unchanged.
Accuracy: on COCO val2017 text-to-image retrieval (5,000 images, 25,014 captions), recall changed by −0.0035, −0.0031 and −0.0044 at 1, 5 and 10.
Smaller download: the vision model is now 119 MB instead of 352 MB.
Existing indexes re-embed their images once on first use.