This repo is the official implementation of the CVPR2023 paper: Joint Token Pruning and Squeezing Towards More Aggressive Compression of Vision Transformers.
conda env create -f environment.yml
Train dTPS-DeiT on a 8-gpu machine using shell scripts in ./scripts:
bash scripts/finetune_dtps_deit_s.sh
you can modify hyperparams in the .sh scripts, including the location index of pruned layers and token keep ratio.
TPS-CVPR2023 is released under the Apache 2.0 license. See LICENSE for details.