DEBRA

Project Introduction

This project primarily introduces a novel on-orbit self-supervised split learning framework designed to achieve real-time on-orbit feature extraction from satellite monitoring data. The framework mainly includes the following components: self-supervised contrastive learning, split learning, on-orbit distributed training optimization algorithms, and model split optimization algorithms. Details are provided in the file structure.

ENVS

• python>=3.8
• torch 2.1.1+cu121
• torchaudio 2.1.1+cu121
• torchvision 0.16.1+cu121
• numpy 1.26.4
• pandas 2.2.1

File Structure

DEBRA/ 
▾ checkpoints/
▾ configs/ # Configuration files
▾ data_transform/
    __init__.py
    eval_aug.py # Test set data augmentation
    simsiam_aug.py # Training set data augmentation
▾ dataset/ # Dataset path
    __init__.py
    compute_std_mean.py 
    dataset_split.py # Dataset split for test set and training set
▾ networks/ # System model
    __init__.py
    resnet.py # Backbone
    simsiam.py # Contrastive module
▾ optimize_algorithm/
    tools/ # Computing latency tools
    MWO.py # Proposed optimization algorithm
▾ optimizers/
    __init__.py
    lr_scheduler.py # Learning rate optimization based on cosine annealing
▾ split_train/ # Model split training
    __init__.py
    network/ The split model
    GA_ssl_train.py # Gradient accumulation training
    satellite_train.py 
    set_GPU_clock.py
▾ log/
▾ tools/ Self-supervised learning tools
    __init__.py
__init__.py 
arguments.py # Parameter configuration file
linear_training.py # Model fine-tuning and evaluation
README.md
ssl_training.py # Self-supervised training

Datasets

• UCMerced
• AID
• EuroSAT
• Directory for storing datasets: DEBRA/dataset/

RUN

• ssl running

  python ssl_training.py
  

• optimization algorithm running

You need to add satellite_positions.csv to the \DEBRA\optimize_algorithm\tools\ directory. This file contains the coordinates of LEO satellites specific to your runtime environment.

Acknowledgment

Parts of our implementation of SimSiam and related utilities were adapted from or inspired by Facebook Research and Patrick Hua. We would like to thank the original authors for making the code publicly available.