Abdelrahman Shaker*1, Muhammad Maaz1, Hanoona Rasheed1, Salman Khan1, Ming-Hsuan Yang2,3 and Fahad Shahbaz Khan1,4
Mohamed Bin Zayed University of Artificial Intelligence1, University of California Merced2, Google Research3, Linkoping University4
Comparison of our SwiftFormer Models with state-of-the-art on ImgeNet-1K. The latency is measured on iPhone 14 Neural Engine (iOS 16).
Comparison with different self-attention modules. (a) is a typical self-attention. (b) is the transpose self-attention, where the self-attention operation is applied across channel feature dimensions (d×d) instead of the spatial dimension (n×n). (c) is the separable self-attention of MobileViT-v2, it uses element-wise operations to compute the context vector from the interactions of Q and K matrices. Then, the context vector is multiplied by V matrix to produce the final output. (d) Our proposed efficient additive self-attention. Here, the query matrix is multiplied by learnable weights and pooled to produce global queries. Then, the matrix K is element-wise multiplied by the broadcasted global queries, resulting the global context representation.
Abstract Self-attention has become a defacto choice for capturing global context in various vision applications. However, its quadratic computational complexity with respect to image resolution limits its use in real-time applications, especially for deployment on resource-constrained mobile devices. Although hybrid approaches have been proposed to combine the advantages of convolutions and self-attention for a better speed-accuracy trade-off, the expensive matrix multiplication operations in self-attention remain a bottleneck. In this work, we introduce a novel efficient additive attention mechanism that effectively replaces the quadratic matrix multiplication operations with linear element-wise multiplications. Our design shows that the key-value interaction can be replaced with a linear layer without sacrificing any accuracy. Unlike previous state-of-the-art methods, our efficient formulation of self-attention enables its usage at all stages of the network. Using our proposed efficient additive attention, we build a series of models called "SwiftFormer" which achieves state-of-the-art performance in terms of both accuracy and mobile inference speed. Our small variant achieves 78.5% top-1 ImageNet-1K accuracy with only 0.8~ms latency on iPhone 14, which is more accurate and 2x faster compared to MobileViT-v2.The latency reported in SwiftFormer for iPhone 14 (iOS 16) uses the benchmark tool from XCode 14.
SwiftFormer meets AndroidCommunity-driven results with Samsung Galaxy S23 Ultra, with Qualcomm Snapdragon 8 Gen 2:
Export & profiler results of SwiftFormer_L1:
Export & profiler results of SwiftFormerEncoder block:
QNN 2.16 2.17 2.18 Latency (msec) 2.17 1.69 1.7Refer to the script above for details of the input & block parameters.
❓ Interested in reproducing the results above?
Refer to Issue #14 for details about exporting & profiling.
conda virtual environment is recommended.
conda create --name=swiftformer python=3.9 conda activate swiftformer pip install torch==1.11.0+cu113 torchvision==0.12.0+cu113 --extra-index-url https://download.pytorch.org/whl/cu113 pip install timm pip install coremltools==5.2.0
Download and extract ImageNet train and val images from http://image-net.org. The training and validation data are expected to be in the train folder and val folder respectively:
|-- /path/to/imagenet/
|-- train
|-- val
We provide training script for all models in dist_train.sh using PyTorch distributed data parallel (DDP).
To train SwiftFormer models on an 8-GPU machine:
sh dist_train.sh /path/to/imagenet 8
Note: specify which model command you want to run in the script. To reproduce the results of the paper, use 16-GPU machine with batch-size of 128 or 8-GPU machine with batch size of 256. Auto Augmentation, CutMix, MixUp are disabled for SwiftFormer-XS, and CutMix, MixUp are disabled for SwiftFormer-S.
On a Slurm-managed cluster, multi-node training can be launched as
sbatch slurm_train.sh /path/to/imagenet SwiftFormer_XS
Note: specify slurm specific parameters in slurm_train.sh script.
We provide an example test script dist_test.sh using PyTorch distributed data parallel (DDP). For example, to test SwiftFormer-XS on an 8-GPU machine:
sh dist_test.sh SwiftFormer_XS 8 weights/SwiftFormer_XS_ckpt.pth
if you use our work, please consider citing us:
@InProceedings{Shaker_2023_ICCV,
author = {Shaker, Abdelrahman and Maaz, Muhammad and Rasheed, Hanoona and Khan, Salman and Yang, Ming-Hsuan and Khan, Fahad Shahbaz},
title = {SwiftFormer: Efficient Additive Attention for Transformer-based Real-time Mobile Vision Applications},
booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
year = {2023},
}
If you have any questions, please create an issue on this repository or contact at abdelrahman.youssief@mbzuai.ac.ae.
Our code base is based on LeViT and EfficientFormer repositories. We thank the authors for their open-source implementation.
I'd like to express my sincere appreciation to Victor Escorcia for measuring & reporting the latency of SwiftFormer on Android (Samsung Galaxy S23 Ultra, with Qualcomm Snapdragon 8 Gen 2). Check SwiftFormer Meets Android for more details!
RetroSearch is an open source project built by @garambo | Open a GitHub Issue
Search and Browse the WWW like it's 1997 | Search results from DuckDuckGo
HTML:
3.2
| Encoding:
UTF-8
| Version:
0.7.4