Home

Awesome

BossNAS

PWC
PWC
PWC

This repository contains PyTorch code and pretrained models of our paper: BossNAS: Exploring Hybrid CNN-transformers with Block-wisely Self-supervised Neural Architecture Search (ICCV 2021).

<p align="center"> <img src=https://user-images.githubusercontent.com/61453811/112087629-45c29700-8bc9-11eb-8536-3485660bc7c2.png width=95%/></p> <p align="center"> Illustration of the Siamese supernets training with ensemble bootstrapping. </p> <p align="center"><img src=https://user-images.githubusercontent.com/61453811/112087643-4a874b00-8bc9-11eb-9440-757429034d81.png width=95%/></p> <p align="center"> Illustration of the fabric-like Hybrid CNN-transformer Search Space with flexible down-sampling positions. </p>

Our Results and Trained Models

Usage

1. Requirements

2. Retrain or Evaluate our BossNet-T models

<p align="center"><img src=https://user-images.githubusercontent.com/61453811/112087617-40fde300-8bc9-11eb-93ed-d043979d3e65.png width=60%/></p> <p align="center">Architecture of our BossNet-T0</p>

3. Evaluate architecture rating accuracy of BossNAS

<p align="center"><img src=https://user-images.githubusercontent.com/61453811/112087625-43603d00-8bc9-11eb-8199-402998b9c7ef.png width=90%/></p> <p align="center"><img src=https://user-images.githubusercontent.com/61453811/112087637-48bd8780-8bc9-11eb-8697-ff535cc9634b.png width=20%/> </p>

4. Search Architecture with BossNAS

First, go to the searching code directory:

cd searching

Citation

If you use our code for your paper, please cite:

@inproceedings{li2021bossnas,
  author = {Li, Changlin and
            Tang, Tao and
            Wang, Guangrun and
            Peng, Jiefeng and
            Wang, Bing and
            Liang, Xiaodan and
            Chang, Xiaojun},
  title = {{B}oss{NAS}: Exploring Hybrid {CNN}-transformers with Block-wisely Self-supervised Neural Architecture Search},
  booktitle = {ICCV},
  year = 2021,
}