Official implementation of CVPR 2024 paper "Multi-criteria Token Fusion with One-step-ahead Attention for Efficient Vision Transformers".
- Clone repository
git clone https://github.com/mlvlab/MCTF.git
cd MCTF
- Setup conda environment
conda env create --file env.yaml
conda activate mctf
python -m torch.distributed.launch --nproc_per_node=8 --use_env main.py --data_path /path/to/imagenet --dataset IMNET --output_dir ./output --batch-size 128 --min_lr_times 0.1 --lr_times 0.03 --epochs 30 --activate_layer [1,2,3,4,5,6,7,8,9,10,11] --model deit_small --use_mctf True --mctf_type [16,8,0,1,1,1,20,40,1,1,0] --task_type [1,1,0.4] --task_weight [1.0,1.0,3.0]
To evaluate a pre-trained MCTF (DeiT-S) model on ImageNet val with a single GPU run:
python main.py --eval --data_path /path/to/imagenet --dataset IMNET --resume /path/to/checkpoint/ --use_mctf True --mctf_type [16,0,0,1,1,1,20,40,1,1,0] --task_type [1,0,0,0] --r_evals [16] --activate_layer [1,2,3,4,5,6,7,8,9,10,11] --model deit_small
Flops and Throughput are measured with single RTX 3090. MCTF-Fast is advanced version of MCTF that boost the throughput while keeping the accuracy and FLOPs by changing the drop ratio per layer.
| Baseline Model | Top 1 Accuracy | GFLOPs | Throughput | script | Checkpoint |
|---|---|---|---|---|---|
| DeiT-T | 72.2 | 1.26 | - | - | - |
| MCTF (DeiT-T) | 72.7 | 0.71 | 4639 | script | drive |
| MCTF-Fast (DeiT-T) | 72.7 | 0.75 | 7386 | script | drive |
| DeiT-S | 79.8 | 4.61 | - | - | - |
| MCTF (DeiT-S) | 80.1 | 2.60 | 2302 | script | drive |
| MCTF-Fast (DeiT-S) | 80.1 | 2.76 | 3302 | script | drive |
@inproceedings{lee2024multi,
title={Multi-criteria Token Fusion with One-step-ahead Attention for Efficient Vision Transformers},
author={Lee, Sanghyeok and Choi, Joonmyung and Kim, Hyunwoo J.},
booktitle={Conference on Computer Vision and Pattern Recognition},
year={2024}
}