EfficientFormer
_{Vision Transformers at MobileNet Speed}

arXiv | PDF

Models are trained on ImageNet-1K and measured by iPhone 12 with CoreMLTools to get latency.

EfficientFormer: Vision Transformers at MobileNet Speed
Yanyu Li^1,2, Genge Yuan^1,2, Yang Wen¹, Eric Hu¹, Georgios Evangelidis¹,
Sergey Tulyakov¹, Yanzhi Wang², Jian Ren^1
1Snap Inc., ²Northeastern University

Abstract

Vision Transformers (ViT) have shown rapid progress in computer vision tasks, achieving promising results on various benchmarks. However, due to the massive number of parameters and model design, e.g., attention mechanism, ViT-based models are generally times slower than lightweight convolutional networks. Therefore, the deployment of ViT for real-time applications is particularly challenging, especially on resource-constrained hardware such as mobile devices. Recent efforts try to reduce the computation complexity of ViT through network architecture search or hybrid design with MobileNet block, yet the inference speed is still unsatisfactory. This leads to an important question: can transformers run as fast as MobileNet while obtaining high performance? To answer this, we first revisit the network architecture and operators used in ViT-based models and identify inefficient designs. Then we introduce a dimension-consistent pure transformer (without MobileNet blocks) as a design paradigm. Finally, we perform latency-driven slimming to get a series of final models dubbed EfficientFormer. Extensive experiments show the superiority of EfficientFormer in performance and speed on mobile devices. Our fastest model, EfficientFormer-L1, achieves 79.2% top-1 accuracy on ImageNet-1K with only 1.6 ms inference latency on iPhone 12 (compiled with CoreML), which runs as fast as MobileNetV2x1.4 (1.6 ms, 74.7% top-1), and our largest model, EfficientFormer-L7, obtains 83.3% accuracy with only 7.0 ms latency. Our work proves that properly designed transformers can reach extremely low latency on mobile devices while maintaining high performance.

Classification on ImageNet-1K

Models

Model	Top-1 Acc.	Latency on iPhone 12 (ms)	Pytorch Checkpoint	CoreML	ONNX
EfficientFormer-L1	79.2 (80.2)	1.6	L1-300 (L1-1000)	L1	L1
EfficientFormer-L3	82.4	3.0	L3	L3	L3
EfficientFormer-L7	83.3	7.0	L7	L7	L7

Latency Measurement

The latency reported is based on the open-source CoreMLTools.

coreml-performance can simply benchmark the speed of our released mlmodels. Thanks for the nice-implemented latency measurement!

Tips: MacOS+XCode and a mobile device (iPhone 12) are needed to reproduce the reported speed.

ImageNet

Prerequisites

conda virtual environment is recommended.

conda install pytorch torchvision cudatoolkit=11.3 -c pytorch
pip install timm
pip install submitit

Data preparation

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

Single machine multi-GPU training

We provide an example training script dist_train.sh using PyTorch distributed data parallel (DDP).

To train EfficientFormer-L1 on an 8-GPU machine:

sh dist_train.sh efficientformer_l1 8

Tips: specify your data path and experiment name in the script!

Multi-node training

On a Slurm-managed cluster, multi-node training can be launched through submitit, for example,

sh slurm_train.sh efficientformer_l1

Tips: specify GPUs/CPUs/memory per node in the script based on your resource!

Testing

We provide an example test script dist_test.sh using PyTorch distributed data parallel (DDP). For example, to test EfficientFormer-L1 on an 8-GPU machine:

sh dist_test.sh efficientformer_l1 8 weights/efficientformer_l1_300d.pth

Using EfficientFormer as backbone

Object Detection and Instance Segmentation
Semantic Segmentation

Acknowledgement

Classification (ImageNet) code base is partly built with LeViT and PoolFormer.

The detection and segmentation pipeline is from MMCV (MMDetection and MMSegmentation).

Thanks for the great implementations!

Citation

If our code or models help your work, please cite our paper:

@article{li2022efficientformer,
  title={EfficientFormer: Vision Transformers at MobileNet Speed},
  author={Li, Yanyu and Yuan, Geng and Wen, Yang and Hu, Eric and Evangelidis, Georgios and Tulyakov, Sergey and Wang, Yanzhi and Ren, Jian},
  journal={arXiv preprint arXiv:2206.01191},
  year={2022}
}