moucheng2017/mnist_robust

CMIC Hackathon

News

[2022 11 09] This repo was born.

[2022 11 10] Team assembled.

Introduction

This code base is for the hackathon project.

How to download the data and prepare for it:

1. download the data from the link (https://drive.google.com/drive/folders/1kKKxA0F8Vcm3042TB-Qc_oF-RifNaYoT) and unzip everything
2. unzip the data folder and all of the npy files in each subfolder and move your data to the same folder where the code is, like this:

your/path/to/cloned/github/repo/
└───data
│     └──np
│        └──train
│        └──test
│───code
│───visulisation

Install the following libraries:

argparse, pytorch1.4, numpy

Run the code in your directory of mnist_robust/code/

python run.py 
--seed 1024 # random seed
--net 'moe' # calling mixture of model based UNet
--lr 0.001 # learning rate
--temp 2,0 # temperature scaling to soft the output prob
--width 8 # number of filters in the 1st conv/mlp layer
--loss_fun 'dice' # loss function between dice or ce
--train_noise 0 # 0 for no aug on train, 1 for gaussian noise on train
--test_noise 1 # 0 for no aug on test; 
                 1 for gaussian noise sigma 0.5; 
                 2 for gaussian noise sigma 1.0; 
                 3 for gaussian blurr with sigma 0.5;
                 4 for gaussian blurr with sigma 0.7;
                 5 for jigsaw images;
                 6 for new class image
--epsilon 5 # when larger than 0, it will use gradient attack,
              this is the strength of the FGSM attack
              

Hyper-parameters:

Under constructions.