Difficulty-Net: Learning to Predict Difficulties for Long-Tailed Recognition

---

This contains the code for our WACV 2023 paper available at arXiv.

Requirements

---

The environment required to successfully reproduce our results mainly needs

- Python >= 3.6
- PyTorch == 1.6.0
- yacs == 0.1.8

How to run the code

---

CIFAR-LT

Our implementation on CIFAR100-LT dataset is built on the code of CDB-CE [2].

To run code for CIFAR-LT, go to folder CIFAR-LT.

Download the dataset from here and save it in the data/ folder.

To start stage-1 training,

python cifar_train_exaugment.py --imbalance 200 --loss_type Difficulty_net --n_gpus 1 --lamda 0.3

This starts the stage-1 training with Difficulty-Net based weighted CE loss. The imbalance variable can take values in {10, 20, 50, 100, 200}.

To use Balanced Softmax with our method, use --loss_type Balanced_Softmax . The best model will be saved in saved_model/ .

To start stage-2 training,

python cifar_train_exaugment_stage2.py --imbalance 200 --stage_2_method LAS --stage1_trained_model ... --n_gpus 1

For evaluation,

python cifar_test.py --saved_model_path ... --imbalance 100 --n_gpus 1

ImageNet-LT/Places-LT

This code is built on that of MiSLAS [1].

Change directory to ImageNet_Places-LT/ .

After downloading ImageNet or Places, update the data path in the respective config/{dataset}/***.yaml files

To start our stage-1 training,

python train_stage1.py --cfg config/{dataset}/{dataset}_{model}_stage1.yaml

For stage-2 training,

python train_stage2.py --cfg config/{dataset}/{dataset}_{model}_stage2.yaml resume saved/{dataset}_{model}_stage1_{yyyymmddhhmm}/ckps/model_best.pth.tar

Change the stage-1 path to resume from as needed.

To evaluate the best model, use

python eval.py --cfg config/{dataset}/{dataset}_{model}_stage2.yaml resume saved/{dataset}_{model}_stage2_{yyyymmddhhmm}/ckps/model_best.pth.tar

Put the respective dataset and model name in {dataset} and {model} respectively. Currently, this code is only supported for resnet10 and resnet50 on imagenet and resnet152 on places.

If you use our code, please consider citing us. For queries, contact at saptarshi.sinha@bristol.ac.uk

References:

1. Zhong et. al., Improving Calibration for Long-Tailed Recognition, CVPR 2021
2. Sinha et. al., Class-Wise Difficulty-Balanced Loss for Solving Class-Imbalance, ACCV 2020