ICM: Data Augmentation for Complementary-Label Learning.

Overview

• In this example, we can run a collection of benchmark for research purpose.

How to use

Environment

• Python version: 3.10
• GPU: Tesla V100-SXM2

Quick Start: ICM for CIFAR10 Training with balance scenario

• To reproduce SCL-NL training with balance scenario, utilizing ICM method

python train.py --algo=scl-nl --dataset_name CIFAR10 --model resnet18 --imb_type exp --imb_factor 1 --mixup true --alpha 0.2 --k_cluster 50 --new_data_aug icm --data_aug true --aug_type flipflop

• To reproduce SCL-NL training with balance scenario, utilizing MICM method

python train.py --algo=scl-nl --dataset_name CIFAR10 --model resnet18 --imb_type exp --imb_factor 1 --mixup true --alpha 0.2 --k_cluster 50 --new_data_aug micm --data_aug true --aug_type flipflop

Quick Start: ICM for CIFAR10 Training with imbalanced scenario with Setup 1.

• To reproduce SCL-NL training with imbalance scenario, utilizing ICM method

python train.py --algo=scl-nl --dataset_name CIFAR10 --model resnet18 --imb_type exp --imb_factor 0.1 --mixup true --alpha 0.2 --k_cluster 50 --new_data_aug icm --data_aug true --aug_type flipflop --setup_type "setup 1" 

• To reproduce SCL-NL training with imbalance scenario, utilizing MICM method

python train.py --algo=scl-nl --dataset_name CIFAR10 --model resnet18 --imb_type exp --imb_factor 0.1 --mixup true --alpha 0.2 --k_cluster 50 --new_data_aug micm --data_aug true --aug_type flipflop --setup_type "setup 1"

• As explanation in the paper, Setup 1: the imbalanced CLL comes from ordinary itself.

Quick Start: ICM for CIFAR10 Training with imbalanced scenario with Setup 2.

• To reproduce SCL-NL training with imbalance scenario, utilizing ICM method

python train.py --algo=scl-nl --dataset_name CIFAR10 --model resnet18 --imb_type exp --imb_factor 1 --mixup true --alpha 0.2 --k_cluster 50 --new_data_aug icm --data_aug true --aug_type flipflop --setup_type "setup 2" --transition_bias 10

• To reproduce SCL-NL training with balance scenario, utilizing MICM method

python train.py --algo=scl-nl --dataset_name CIFAR10 --model resnet18 --imb_type exp --imb_factor 1 --mixup true --alpha 0.2 --k_cluster 50 --new_data_aug micm --data_aug true --aug_type flipflop --setup_type "setup 2" --transition_bias 10

• Setup 2: The imbalanced CLL is from biased transition matrix.

Quick Start: ICM for CIFAR10 Training with imbalanced scenario with Setup 3.

• To reproduce SCL-NL training with imbalance scenario, utilizing ICM method

python train.py --algo=scl-nl --dataset_name CIFAR10 --model resnet18 --imb_type exp --imb_factor 0.1 --mixup true --alpha 0.2 --k_cluster 50 --new_data_aug icm --data_aug true --aug_type flipflop --setup_type "setup 2" --transition_bias 10

• To reproduce SCL-NL training with balance scenario, utilizing MICM method

python train.py --algo=scl-nl --dataset_name CIFAR10 --model resnet18 --imb_type exp --imb_factor 0.1 --mixup true --alpha 0.2 --k_cluster 50 --new_data_aug micm --data_aug true --aug_type flipflop --setup_type "setup 2" --transition_bias 10

• Setup 3: The imbalanced CLL is a combined of imbalanced ordinary dataset and biased transition matrix. Therefore, imb_factor 0.1 and setup_type "setup 2"

Parameter	Description
--config	Path to config file (specify by different dataset)
--algorithm	SCL-NL, FWD, DM, SCL_EXP
--model	resnet18, m-resnet18, linear, mlp
--new_data_aug	icm, micm, cl_aug, orig_mixup, none
--aug_type	randaug, autoaug, cutout, flipflop
--dataset	CIFAR10, CIFAR20, PCLCIFAR10, PCLCIFAR20, MNIST, KMNIST, FashionNIST
--imb_factor	1, 0.1, 0.02, 0.01
--imb_exp	exp, step
--k_cluster	The number of clustering