VITA-Group / AsViT

[ICLR 2022] "As-ViT: Auto-scaling Vision Transformers without Training" by Wuyang Chen, Wei Huang, Xianzhi Du, Xiaodan Song, Zhangyang Wang, Denny Zhou

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

As-ViT: Auto-scaling Vision Transformers without Training [PDF]

MIT licensed

Wuyang Chen, Wei Huang, Xianzhi Du, Xiaodan Song, Zhangyang Wang, Denny Zhou

In ICLR 2022.

Note: We implemented topology search (sec. 3.3) and scaling (sec. 3.4) in this code base in PyTorch. Our training code is based on Tensorflow and Keras on TPU, which will be released soon.

Overview

We present As-ViT, a framework that unifies the automatic architecture design and scaling for ViT (vision transformer), in a training-free strategy.

Highlights:

  • Trainig-free ViT Architecture Design: we design a "seed" ViT topology by leveraging a training-free search process. This extremely fast search is fulfilled by our comprehensive study of ViT's network complexity (length distorsion), yielding a strong Kendall-tau correlation with ground-truth accuracies.
  • Trainig-free ViT Architecture Scaling: starting from the "seed" topology, we automate the scaling rule for ViTs by growing widths/depths to different ViT layers. This will generate a series of architectures with different numbers of parameters in a single run.
  • Efficient ViT Training via Progressive Tokenization: we observe that ViTs can tolerate coarse tokenization in early training stages, and further propose to train ViTs faster and cheaper with a progressive tokenization strategy.

teaser
Left: Length Distortion shows a strong correlation with ViT's accuracy. Middle: Auto scaling rule of As-ViT. Right: Progressive re-tokenization for efficient ViT training.

Prerequisites

  • Ubuntu 18.04
  • Python 3.6.9
  • CUDA 11.0 (lower versions may work but were not tested)
  • NVIDIA GPU + CuDNN v7.6

This repository has been tested on V100 GPU. Configurations may need to be changed on different platforms.

Installation

  • Clone this repo:
git clone https://github.com/VITA-Grou/AsViT.git
cd AsViT
  • Install dependencies:
pip install -r requirements.txt

1. Seed As-ViT Topology Search

CUDA_VISIBLE_DEVICES=0 python ./search/reinforce.py --save_dir ./output/REINFORCE-imagenet --data_path /path/to/imagenet

This job will return you a seed topology. For example, our search seed topology is 8,2,3|4,1,2|4,1,4|4,1,6|32, which can be explained as below:

Stage1Stage2Stage3Stage4Head
Kernel K1Split S1Expansion E1Kernel K2Split S2Expansion E2Kernel K3Split S3Expansion E3Kernel K4Split S4Expansion E4
82341241441632

2. Scaling

CUDA_VISIBLE_DEVICES=0 python ./search/grow.py --save_dir ./output/GROW-imagenet \
--arch "[arch]" --data_path /path/to/imagenet

Here [arch] is the seed topology (output from step 1 above). This job will return you a series of topologies. For example, our largest topology (As-ViT Large) is 8,2,3,5|4,1,2,2|4,1,4,5|4,1,6,2|32,180, which can be explained as below:

Stage1Stage2Stage3Stage4HeadInitial Hidden Size
Kernel K1Split S1Expansion E1Layers L1Kernel K2Split S2Expansion E2Layers L2Kernel K3Split S3Expansion E3Layers L3Kernel K4Split S4Expansion E4Layers L4
823541224145416232180

3. Evaluation

Tensorflow and Keras code for training on TPU. To be released soon.

Citation

@inproceedings{chen2021asvit,
  title={Auto-scaling Vision Transformers without Training},
  author={Chen, Wuyang and Huang, Wei and Du, Xianzhi and Song, Xiaodan and Wang, Zhangyang and Zhou, Denny},
  booktitle={International Conference on Learning Representations},
  year={2022}
}

About

[ICLR 2022] "As-ViT: Auto-scaling Vision Transformers without Training" by Wuyang Chen, Wei Huang, Xianzhi Du, Xiaodan Song, Zhangyang Wang, Denny Zhou

License:MIT License


Languages

Language:Python 100.0%