SLAB

SLAB: Efficient Transformers with Simplified Linear Attention and Progressive Re-parameterized Batch Normalization

Jialong Guo*, Xinghao Chen*, Yehui Tang, Yunhe Wang (*Equal Contribution)

ICML 2024

[arXiv] [BibTeX]

🔥 Updates

• 2024/08/23: Unofficial pretrained checkpoints for Llama-350M-PRepBN from huggingface.
• 2024/05/13: Pre-trained models and codes of SLAB are released both in Pytorch and Mindspore.

📸 Overview

This is an official pytorch implementation of our paper "SLAB: Efficient Transformers with Simplified Linear Attention and Progressive Re-parameterized Batch Normalization". In this paper, we investigate the computational bottleneck modules of efficient transformer, i.e., normalization layers and attention modules. Layer normalization is commonly used in transformer architectures but is not computational friendly due to statistic calculation during inference. However, replacing Layernorm with more efficient batch normalization in transformer often leads to inferior performance and collapse in training. To address this problem, we propose a novel method named PRepBN to progressively replace LayerNorm with re-parameterized BatchNorm in training. During inference, the proposed PRepBN could be simply re-parameterized into a normal BatchNorm, thus could be fused with linear layers to reduce the latency. Moreover, we propose a simplified linear attention (SLA) module that is simply yet effective to achieve strong performance. Extensive experiments on image classification as well as object detection demonstrate the effectiveness of our proposed method. For example, powered by the proposed methods, our SLAB-Swin obtains 83.6% top-1 accuracy on ImageNet with 16.2ms latency, which is 2.4ms less than that of Flatten-Swin with 0.1 higher accuracy.

<p align="center"> <img src="docs/framework.png" width="90%"/> <br /> <em> Figure 1: The framework of our proposed Progressive Re-parameterized BatchNorm. </em> </p> <p align="center"> <img src="docs/attn.png" width="90%"/> <br /> <em> Figure 2: Visualization of attention map for different methods. </em> </p> <p align="center"> <img src="docs/results.png" width="90%"/> <br /> <em> Figure 3: Results of our method for classification and detection. </em> </p> <p align="center"> <img src="docs/llama.png" width="90%"/> <br /> <em> Figure 4: Results of our method for LLaMA-350M on various benchmarks. </em> </p>

1️⃣ Image Classification

Dependenices

- torch
- torchvision
- numpy
- einops
- timm==0.4.12
- opencv-python==4.4.0.46
- termcolor==1.1.0
- yacs==0.1.8
- apex

Training

Train models from scratch using the following command:

python -m torch.distributed.launch --nproc_per_node=8 main.py --cfg <config-path> --data-path <imagenet-path> --output <output-path>

Evaluation

Merge PRepBN for Swin Transformer: For a Swin-T model, we provide the implementation of PRepBN fusion. You can convert the whole model by simply calling merge_bn of the module. This is the recommended way. Examples are shown in eval.py.

for module in model.modules():
    if module.__class__.__name__ == 'SwinTransformerBlock':
        module.merge_bn()
    elif module.__class__.__name__ == 'PatchMerging':
        module.merge_bn()
    elif module.__class__.__name__ == 'PatchEmbed':
        module.merge_bn()
for module in model.modules():
    if module.__class__.__name__ == 'SwinTransformer':
        module.merge_bn()

We have also provide an example for the conversion.

python -m torch.distributed.launch --nproc_per_node=1 eval.py --cfg cfgs/swin_t_prepbn.yaml --batch-size 128 --data-path <imagenet-path>  --pretrained <pretrained-path>

Checkpoints

| Model | Top1 | config | checkpoints | | :---: | :---: | :---: | :---: | |deit_t_prepbn|73.6%|deit_t_prepbn.yaml|deit_tiny_prepbn.pth| |deit_s_prepbn|80.2%|deit_s_prepbn.yaml|deit_small_prepbn.pth| |slab_deit_t|74.3%|slab_deit_t.yaml|slab_deit_tiny.pth| |slab_deit_s|80.0%|slab_deit_s.yaml|slab_deit_small.pth| |pvt_t_prepbn|76.0%|pvt_t_prepbn.yaml|pvt_tiny_prepbn.pth| |pvt_s_prepbn|80.1%|pvt_s_prepbn.yaml|pvt_small_prepbn.pth| |pvt_m_prepbn|81.7%|pvt_m_prepbn.yaml|pvt_medium_prepbn.pth| |slab_pvt_t|76.5%|slab_pvt_t.yaml|slab_pvt_tiny.pth| |swin_t_prepbn|81.4%|swin_t_prepbn.yaml|swin_tiny_prepbn.pth| |slab_swin_t|81.8%|slab_swin_t.yaml|slab_swin_tiny.pth| |slab_swin_s|83.6%|slab_swin_s.yaml|slab_swin_small.pth| |slab_cswin_t|82.8%|slab_cswin_t.yaml|slab_cswin_tiny.pth|

2️⃣ Object Detection

Installation

pip install torch 
pip install torchvision

pip install timm==0.4.12
pip install einops
pip install opencv-python==4.4.0.46 termcolor==1.1.0 yacs==0.1.8
pip install -U openmim
pip install mmcv-full==1.4.0
pip install mmdet==2.11.0

Install apex

Training

SLAB-Swin-T

python -m torch.distributed.launch --nproc_per_node 8 --nnodes <world_size> --node_rank <rank> train.py configs/swin/mask_rcnn_slab_swin_tiny_patch4_window7_mstrain_480-800_adamw_1x_coco.py --work-dir <output_path> --launcher pytorch --init_method <init_method> --cfg-options model.pretrained=<pretrained_backbone_path>

SLAB-Swin-S

python -m torch.distributed.launch --nproc_per_node 8 --nnodes <world_size> --node_rank <rank> train.py configs/swin/mask_rcnn_slab_swin_small_patch4_window7_mstrain_480-800_adamw_1x_coco.py --work-dir <output_path> --launcher pytorch --init_method <init_method> --cfg-options model.pretrained=<pretrained_backbone_path>

Swin-T-RepBN

python -m torch.distributed.launch --nproc_per_node 8 --nnodes <world_size> --node_rank <rank> train.py configs/swin/mask_rcnn_swin_tiny_prepbn_patch4_window7_mstrain_480-800_adamw_1x_coco.py --work-dir <output_path> --launcher pytorch --init_method <init_method> --cfg-options model.pretrained=<pretrained_backbone_path>

Swin-S-RepBN

python -m torch.distributed.launch --nproc_per_node 8 --nnodes <world_size> --node_rank <rank> train.py configs/swin/mask_rcnn_swin_small_prepbn_patch4_window7_mstrain_480-800_adamw_1x_coco.py --work-dir <output_path> --launcher pytorch --init_method <init_method> --cfg-options model.pretrained=<pretrained_backbone_path>

PVT-T-RepBN

python -m torch.distributed.launch --nproc_per_node 8 --nnodes <world_size> --node_rank <rank> train.py configs/pvt/mask_rcnn_pvt_t_prepbn_fpn_1x_coco.py --work-dir <output_path> --launcher pytorch --init_method <init_method> --cfg-options model.pretrained=<pretrained_backbone_path>

PVT-S-RepBN

python -m torch.distributed.launch --nproc_per_node 8 --nnodes <world_size> --node_rank <rank> train.py configs/pvt/mask_rcnn_pvt_s_prepbn_fpn_1x_coco.py --work-dir <output_path> --launcher pytorch --init_method <init_method> --cfg-options model.pretrained=<pretrained_backbone_path>

Checkpoints

TBD

3️⃣ Language Task

Dependencies

- torch==1.13.1
- tensorboardX
- numpy
- rouge_score
- fire
- openai==0.27.6
- transformers==4.29.1
- datasets==2.17.0
- sentencepiece
- tokenizers==0.13.3
- deepspeed==0.8.3
- accelerate==0.27.2
- scikit-learn

Evaluation

Download the unofficial pretrained checkpoints for Llama-350M-PRepBN from huggingface.

python evaluation.py --ckpt <checkpoint-path>

✏️ Reference

If you find SLAB useful in your research or applications, please consider giving a star ⭐ and citing using the following BibTeX:

@inproceedings{guo2024slab,
  title={SLAB: Efficient Transformers with Simplified Linear Attention and Progressive Re-parameterized Batch Normalization},
  author={Guo, Jialong and Chen, Xinghao and Tang, Yehui  and Wang, Yunhe},
  booktitle={International Conference on Machine Learning},
  year={2024}
}