Requirements and Installation

• A computer running macOS or Linux
• For training new models, you'll also need a NVIDIA GPU and NCCL
• Python version 3.6
• A PyTorch installation

Basic Installation Package

pip install opencv-python
pip install easydict
pip install pyyaml
pip install matplotlib
pip install scipy
git clone https://github.com/NVIDIA/apex.git
cd apex
python setup.py install
pip install gco-wrapper

FILE description

• ./checkpoint : Stored with trained models

• ./results : Stored training log

• ./mixup.py : original mixup function: Beta Distribution with vector

• ./mixup_v2.py : modified mixup function: Matrix-Mixup + Gaussian Distribution

• ./train.py : one_third concatenation(matrix mix-up images, original mix-up images and original images in one iteration

Training

Use python train.py to train a new model. Here is an example setting:

$ CUDA_VISIBLE_DEVICES=0 python train.py --lr=0.1 --seed=20220103 --decay=1e-4

Generate mix-up images

Uncomment Line :63,64,66,67 in train.py & Uncomment Line 30-33 in mixup_v2.py

$ python train.py --lr=0.1 --seed=20220103 --decay=1e-4 --epoch=1

Add adv samples for test

1、install torchattacks pip install torchattacks

2、PGD_eval.py run with

    CUDA_VISIBLE_DEVICES=0 python PGD_eval.py

Generate High and Low frequency data：frequency.py

• All code are from : frequencyHelper.py ，(https://github.com/HaohanWang/HFC/tree/master/utility) , only do some minor modifications for data store location
• adds a line of code to generate test data labels to facilitate subsequent model testing
• Run with：

  python frequency.py
  

  data is kept in: ./data/CIFAR10/

Create new dataset： new_dataset.py

• used for parse the generated data from frequency.py, and processed into a form that can be loaded by dataloader

Use the generated data for training in train.py

• add a button for using frequency or not in training process