xERM.RIDE.public

Cross-Domain Empirical Risk Minimization for Unbiased Long-tailed Classification

This project is built on top of RIDE.

Install

Please refer to RIDE to to set environment and dataset.

CIFAR100-LT-IB-100

2 experts

Step 1:

Train RIDE-PC-2 model:

python train.py -c configs/cifar100_ERM_2experts.json

Eval RIDE-PC-2 model:

python test.py -c configs/cifar100_ERM_2experts.json -r model_path

Step 2:

Train xERM-RIDE-PC-2 model, please change the balanced_model_path to the PC model path

python train.py -c "configs/config_imbalance_cifar100_ride_xERM_2experts.json"

Eval

python test.py  -c  configs/cifar100_xERM_2experts.json -r model_path

3 experts

Step 1:

Train RIDE-PC-3 model:

python train.py -c configs/cifar100_ERM_3experts.json

Eval RIDE-PC-3 model:

python test.py -c configs/cifar100_ERM_3experts.json -r model_path

Step 2:

Train xERM-RIDE-PC-3 model, please change the balanced_model_path to the PC model path

python train.py -c "configs/config_imbalance_cifar100_ride_xERM_3experts.json"

Eval

python test.py  -c  configs/cifar100_xERM_3experts.json -r model_path

Result

Model	Overall	Recall Many	Recall Med	Recall Few	Precision Many	Precision Med	Precision Few
RIDE-PC-2	46.3	62.2	45.4	28.7	59.9	49.8	27.5
xERM-RIDE-PC-2	47.8	65.5	50.5	24.0	53.9	53.5	45.0
RIDE-PC-3	47.8	64.8	46.2	29.6	61.6	51.0	28.8
xERM-RIDE-PC-3	50.5	66.7	52.2	29.6	56.5	53.0	35.6