activelearning2022/adversarial_active_learning

Here's the code for paper Adaptive Adversarial samples based Active learning for medical image classification

Adaptive adversarial

Introduction

This project is the code for paper Adaptive Adversarial samples based Active learning for medical image classification based on python and pytorch framework.

Requirements

The main package and version of the python environment are as follows

# Name                    Version         
python                    3.8.5                    
pytorch                   1.10.1         
torchvision               0.11.2         
cudatoolkit               10.2.89       
cudnn                     7.6.5           
matplotlib                3.3.2              
numpy                     1.19.2        
opencv                    4.6.0.66         
pandas                    1.1.3               
scikit-learn              0.23.2                
tqdm                      4.50.2             

The above environment is successful when running the code of the project. Pytorch has very good compatibility. Thus, I suggest that try to use the existing pytorch environment firstly.

---

Usage

1) Download Project

Runninggit clone https://github.com/activelearning2022/adversarial_active_learning.git
The project structure and intention are as follows :

Adversarial active learning			# Source code		
    ├── seed.py			 	                                          # Set up random seed
    ├── query_strategies		                                    # All query_strategies
    │   ├── adaptive_adversarial_sample.py                      # Our method
    │   ├── adversarial_deepfool.py                             # The method we based on
    │   ├── bayesian_active_learning_disagreement_dropout.py	  # Deep bayesian query method
    │   ├── entropy_sampling.py		                              # Entropy based query method
    │   ├── entropy_sampling_dropout.py		                      # Entropy based MC dropout query method
    │   ├── random_sampling.py		                              # Random selection
    │   ├── strategy.py                                         # Functions needed for query strategies
    ├── data.py	                                                # Prepare the dataset & initialization and update for training dataset
    ├── handlers.py                                             # Get dataloader for the dataset
    ├── main.py			                                            # An example for code utilization, including the whole process of active learning
    ├── nets.py		                                              # Training models and methods needed for query method
    ├── supervised_baseline.py	                                # An example for supervised learning traning process
    └── utils.py			                                          # Important setups including network, dataset, hyperparameters...

2) Datasets preparation

1. Download the datasets from the official address:

   Messidor: https://www.adcis.net/en/third-party/messidor/

   Breast Cancer Diagnosis: https://iciar2018-challenge.grand-challenge.org/

2. Modify the data folder path for specific dataset in data.py

3) Run Active learning process

Please confirm the configuration information in the [utils.py]

  python main.py \
      --n_round 34 \
      --n_query 20 \
      --n_init_labeled 100 \
      --dataset_name Messidor \
      --traning_method supervised_train_acc \
      --strategy_name RandomSampling \
      --seed 2022

The training results will be saved to the corresponding directory(save name) in performance.csv.
You can also run supervised_baseline.py by

python supervised_baseline.py

Visualization

1 Active learning performance visualization
After you got the performance.csv, you can run visualization.py to visualize the whole process