Adaptive Split-Fusion Transformer

Update:

19/02/2024

Release the pretrained model weights.

31/08/2022

Add the code for ImageNet-21k pretrained weights loading.
Add the code for the pyramid version model ASF-former_p.

27/04/2022

The paper is posted on arXiv and the code is released.

1. Requirements

timm==0.3.4

torch==1.8.0

torchvision

pyyaml

Data Preparation

ImageNet with following folder structure:

│imagenet/
├──train/
│  ├── n01440764
│  │   ├── n01440764_10026.JPEG
│  │   ├── n01440764_10027.JPEG
│  │   ├── ......
│  ├── ......
├──val/
│  ├── n01440764
│  │   ├── ILSVRC2012_val_00000293.JPEG
│  │   ├── ILSVRC2012_val_00002138.JPEG
│  │   ├── ......
│  ├── ......

2. Pretrained Models

ImageNet-1K Training:

Model	Params	Top-1
ASF-former-S	19.3M	82.7%
ASF-former-B	56.7M	83.9%
ASF-former_p-S	21.3M	83.0%
ASF-former_p-B	58.9M	83.9%

ImageNet-22K Pretraining + ImageNet-1K Fine-tuning:

Model	Params	Top-1
ASF-former-B	56.7M	85.2%

3. Training

Train ASF-former-S with 8 GPUs:

CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 ./distributed_train.sh 8 path/to/data --model ASF_former_S -b 64 --lr 5e-4 --weight-decay .05 --amp --img-size 224

Train ASF-former-B with 8 GPUs:

CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 ./distributed_train.sh 8 path/to/data --model ASF_former_B -b 64 --lr 5e-4 --weight-decay .065 --amp --img-size 224

Train ASF-former_p-S with 8 GPUs:

CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 ./distributed_train.sh 8 /data/imagenet --model ASF_former_p_S -b 64 --lr 5e-4 --weight-decay .05 --amp --img-size 224

Train ASF-former_p-B with 8 GPUs:

CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 ./distributed_train.sh 8 /data/imagenet --model ASF_former_p_B -b 64 --lr 5e-4 --weight-decay .065 --amp --img-size 22

4. Evaluation

Evaluate the ASF-former-S model:

CUDA_VISIBLE_DEVICES=0 python main.py path/to/data --model ASF_former_S -b 100 --eval_checkpoint path/to/checkpoint

Evaluate the ASF-former-B model:

CUDA_VISIBLE_DEVICES=0 python main.py path/to/data --model ASF_former_B -b 100 --eval_checkpoint path/to/checkpoint

5. Transfer Learning

Transfer ASF-former-S to CIFAR-10:

CUDA_VISIBLE_DEVICES=0,1 python transfer_learning.py --lr 0.025 --b 64 --num-classes 10 --img-size 224 --transfer-learning True --transfer-model path/to/model

Transfer ASF-former-B to CIFAR-10:

CUDA_VISIBLE_DEVICES=0,1 python transfer_learning.py --lr 0.025 --b 64 --num-classes 10 --img-size 224 --transfer-learning True --transfer-model path/to/model --model ASF_former_B

Transfer ASF-former-S to CIFAR-100:

CUDA_VISIBLE_DEVICES=0,1 python transfer_learning.py --lr 0.05 --b 64 --num-classes 100 --img-size 224 --transfer-learning True --transfer-model path/to/model --dataset cifar100

Transfer ASF-former-B to CIFAR-100:

CUDA_VISIBLE_DEVICES=0,1 python transfer_learning.py --lr 0.05 --b 64 --num-classes 100 --img-size 224 --transfer-learning True --transfer-model path/to/model --dataset cifar100 --model ASF_former_B

Citation

@article{Su2022AdaptiveST,
  title={Adaptive Split-Fusion Transformer},
  author={Zixuan Su and Hao Zhang and Jingjing Chen and Lei Pang and Chong-Wah Ngo and Yu-Gang Jiang},
  journal={ArXiv},
  year={2022},
  volume={abs/2204.12196}
}

Our codes are based on T2T-ViT.

szx503045266 / ASF-former