joakimbh/atc_tiny-nab-single

Fine-Tuned Agglomerative Token Clustering - DeiT-Tiny-Single - NABirds

Model Details

Agglomerative Token Clustering (ATC), a novel hierarchical hard-merging based token reduction method.

• Developed by: Joakim Bruslund Haurum, Sergio Escalera, Graham W. Taylor, and Thomas B. Moeslund
• Model type: Vision Transformer
• License: MIT
• Task: Image Classification

Model Card

• Backbone: DeiT-Tiny
• Token Reduction Method: ATC
• Linkage Function: Single
• Reduction Ratio: {0.25, 0.5, 0.7, 0.9}
• Reduction Stages: 3, 6, 9

More Resources

• Repository: https://github.com/JoakimHaurum/ATC
• Paper: https://arxiv.org/abs/2409.11923
• Project Page: https://vap.aau.dk/atc
• HuggingFace Collection: https://huggingface.co/collections/joakimbh/agglomerative-token-clustering-66e94dfb313e85ec97590fe4

Use

The model files contain both standard and EMA model parameters. The version which gave the best performance is indicated with the "ema_best" boolean.