Awesome

Class-Balanced Active Learning for Image Classification:

Description

This repository contains code for the paper Class-Balanced Active Learning for Image Classification:

http://arxiv.org/abs/2110.04543

Dependencies

Install Anaconda environment: https://docs.anaconda.com/anaconda/install/linux/

Install PyTorch :

conda install pytorch torchvision torchaudio cudatoolkit=10.2 -c pytorch

Install CVXPY python package: https://www.cvxpy.org/install/

Install Gurobi optimizer and its licenese: https://www.gurobi.com/gurobi-and-anaconda-for-linux/

Getting Started

Before starting AL cycles, execute run_cycl_0.py on cifar10/cifar100 dataset to obtain the imbalance dataset (only once) and train the model on initial sampels:

CUDA_VISIBLE_DEVICES=0 python run_cycle_0.py --method RandomSampling --dataset cifar100

Executing program

To run Active Learning cycles use python run.py with the following arguments:

--imb_type (To specify the imbalance type of the dataset)

--imb_factor (To specify the imbalance factor)

--dataset (To specify the dataset)

Examples:

To run standard EntropySampling method on CIFAR100 dataset and step imbalance and imbalance factor=0.1:

CUDA_VISIBLE_DEVICES=0 python run.py --method EntropySampling_imbalance --dataset cifar100 --imb_factor 0.1 --imb_type step

To run EntropySampling with optimal class balancing add "optimal" in the method name, for example:

CUDA_VISIBLE_DEVICES=0 python run.py --method EntropySampling_optimal --dataset cifar100 --imb_factor 0.1 --imb_type step

Other sampling strategies are available by --mehtod argument:

CUDA_VISIBLE_DEVICES=0 python run.py --method RandomSampling --dataset cifar100 --imb_factor 0.1 --imb_type step

Contributors

Javad Zolfaghari Bengar (djavad.z@gmail.com)

Laura Lopez Fuentes (lopezfuenteslaura@gmail.com )