#burn #vit #computer-vision #dinov2 #dino

bin+lib burn_dino

burn dinov2 model inference and training

3 unstable releases

new 0.2.1 Nov 23, 2024
0.2.0 Nov 23, 2024
0.1.0 Nov 22, 2024

#298 in Machine learning

45 downloads per month

MIT/Apache

64KB
1.5K SLoC

burn_dino 🔥🦖

test GitHub License GitHub Last Commit GitHub Releases GitHub Issues Average time to resolve an issue crates.io

burn dinov2 model, view the live demo

input fg pca
Alt text Alt text
Alt text Alt text
Alt text Alt text
Alt text Alt text

cargo run --example pca cargo run -p bevy_burn_dino

features

  • inference
  • pca transform layer
  • all ViT configurations
  • real-time camera demo
  • training (loss + dropout)
  • optimized attention
  • automatic weights cache/download
  • quantization

setup

  • download pre-trained model (ViT-S|B, /wo registers) from here
  • place in ./assets/models
  • run import tool cargo run --bin import

benchmarks

  • cargo bench
  • open target/criterion/report/index.html

Dependencies

~75–120MB
~2.5M SLoC