ATM

Official PyTorch implementation from our paper: Lossless Token Merging Even Without Fine-Tuning in Vision Transformers (ECAI 2025)

Running the code

To reproduce our experiments, run the following command:

Inference

python main.py --eval --resume /path/to/model --model deit_small_patch16_224 --data-path /path/to/ILSVRC2012 --batch-size 1024 --alpha 0.99 --beta 0.04 --lower-bound 0.88

The optimal hyperparameter combination can be found in the Appendix.

Visualization results

Our proposed ATM adaptively merges similar token pairs based on their image characteristics.

Acknowledgement

Our code is based on DeiT and ToMe. Thanks for their great works.