A Study of Zero-cost proxies for Remote Sensing Image Segmentation

This repository contains code for paper [A Study of Zero-cost proxies for Remote Sensing Image Segmentation].

Prerequisites

• Python 3.8
• Pytorch 1.8.1
• torch-scatter pip install torch-scatter==2.0.9 -f https://data.pyg.org/whl/torch-1.8.0+${CUDA}.html
• torch-sparse pip install torch-sparse==0.6.10 -f https://data.pyg.org/whl/torch-1.8.0+${CUDA}.html
• torch-cluster pip install torch-cluster==1.6.0 -f https://data.pyg.org/whl/torch-1.8.0+${CUDA}.html
• torch-spline-conv pip install torch-spline-conv==1.2.1 -f https://data.pyg.org/whl/torch-1.8.0+${CUDA}.html
• torch-geometric pip install torch-geometric==2.0.4

${CUDA} refers to cpu, cu101, cu102, cu111

Environments

• Ubuntu 18.04
• cuda 10.2
• cudnn 8.1.1

Installation

1. See detectron2 installation instructions to install required packages.
2. git clone https://github.com/auroua/nas_seg_detectron2.git
3. python -m pip install -e nas_seg_detectron2

Getting Started

1. Datasets

This project uses the WHU Building and Massachusetts road datasets.

2. Running Segmentation Algorithms DeepLab V3+

   # Train model  
   # As using the sync bn layer, the parameter --num-gpus must be 2
   python train_net_seg.py --config-file /home/albert_wei/WorkSpaces/nas_seg_detectron2/configs/Segmentation/PSPNet/Remote-SemanticSegmentation/Base-PSPNet-OS16-Semantic.yaml --num-gpus 2 --num-machines 1

   # Evaluate model
   python train_net_seg.py --config-file /home/albert_wei/WorkSpaces/nas_seg_detectron2/configs/Segmentation/PSPNet/Remote-SemanticSegmentation/Base-PSPNet-OS16-Semantic.yaml --num-gpus 2 --num-machines 1 --eval-only --resume

3. Running Segmentation Algorithms PSPNet

   # Train model 
   # As using the sync bn layer, the parameter --num-gpus must be 2
   python train_net_seg.py --config-file /home/albert_wei/WorkSpaces/nas_seg_detectron2/configs/Segmentation/DeepLab/Remote-SemanticSegmentation/Base-DeepLabV3-plus-OS16-Semantic.yaml --num-gpus 2 --num-machines 1

   # Evaluate model
   python train_net_seg.py --config-file /home/albert_wei/WorkSpaces/nas_seg_detectron2/configs/Segmentation/DeepLab/Remote-SemanticSegmentation/Base-DeepLabV3-plus-OS16-Semantic.yaml --num-gpus 2 --num-machines 1 --eval-only --resume

4. Using NPENAS-NP to search architecture from SEG101

   export DETECTRON2_DATASETS='/home/albert_wei/fdisk_a/datasets_train_seg/'
   cd nas_seg_detectron2/tools_nas
   # --config_file: config file path for neural architecture search
   # --gpus: using how many to finish search
   # --save_dir: save search results
   python train_nas_open_domain.py --config_file '/home/albert_wei/WorkSpaces/nas_seg_detectron2/configs_nas/seg/seg_101_npenas.yaml' --gpus 2 --save_dir '/home/albert_wei/fdisk_b/'

5. Train the searched architecture

   cd nas_seg_detectron2/tools_nas
   # Train the searched architecture
   # You have to modify the following parameters in the config-file
   # SEARCHED_ARCHITECTURE: point to the searched architecture's folder, e.g., '/home/albert_wei/fdisk_a/train_output_remote/npenas_seg101_mass_road_20_epochs_2/1e4d24eb52f67424eabfe070ffbaee7ac2f31ca4f2e19a3c87680fbb4ed8167a'
   # OUTPUT_DIR: point to the folder where to save the training results
   python train_searched_architecture.py --config-file /home/albert_wei/WorkSpaces/nas_seg_detectron2/configs/Segmentation/Seg101/Remote-SemanticSegmentation/Base-Seg101-OS16-Semantic.yaml --num-gpus 1 --num-machines 1
   
   # Evaluate searched architecture
   python train_net_seg.py --config-file /home/albert_wei/WorkSpaces/nas_seg_detectron2/configs/Segmentation/PSPNet/Remote-SemanticSegmentation/Base-PSPNet-OS16-Semantic.yaml --num-gpus 2 --num-machines 1 --eval-only --resume

Modifications to detectron2

1. Adding python file detectron2/data/datasets/builtin_remote to include the mass road and whu building datasets.

Acknowledge

1. detectron2
2. NPENAS
3. pytorch_geometric
4. Zero-Cost-NAS

Citation

If you find this project useful for your research, please cite our paper:

    @inproceedings{weistudy,
      title={A Study of Zero-Cost Proxies for Remote Sensing Image Segmentation},
      author={Wei, Chen and Guo, Tai Kai and Tang, Ping Yi and Ge, Yao Jun and Liang, Jimin},
      booktitle={First Conference on Automated Machine Learning (Late-Breaking Workshop)}
    }

Contact

Chen Wei

email: weichen_3@stu.xidian.edu.cn