FFDS-Loss

Official pytorch implementation of the paper:

- Rethinking Long-Tailed Visual Recognition with Dynamic Probability Smoothing and Frequency Weighted Focusing

ICIP2023 | Project Page | Paper | Video | Poster | Slides | Bibtex

(Released on August 28, 2023)

Introduction:

This paper highlights the limitations of existing solutions that combine class- and instance-level re-weighting loss in a naive manner. Specifically, we demonstrate that such solutions result in overfitting the training set, significantly impacting the rare classes. To address this issue, we propose a novel loss function that dynamically reduces the influence of outliers and assigns class-dependent focusing parameters. We also introduce a new long-tailed dataset, ICText-LT, featuring various image qualities and greater realism than artificially sampled datasets. Our method has proven effective, outperforming existing methods through superior quantitative results on CIFAR-LT, Tiny ImageNet-LT, and our new ICText-LT datasets.

Codebase Reference: MiSLAS and IB-Loss.

Installation

Requirements

• PyTorch 1.10.0
• torchvision 0.15.0
• yacs 0.1.8
• wandb 0.13.11

Create Virtual Environment

conda create -n FFDS python==3.9
source activate FFDS

Install FFDS

git clone https://github.com/nwjun/FFDS-Loss.git
cd FFDS-Loss
pip install -r requirements.txt

Dataset

• Imbalanced CIFAR
• Imbalanced Tiny-ImageNet
• Imbalanced ICText

Download all the datasets (except CIFAR) and place them in the ./data folder. If you prefer a different location for the datasets, you can modify the data_path parameter in the config/*/*.yaml accordingly.

The default configuration expects the following folder structure:

./data
    ├── cifar10   # This dataset will be downloaded automatically
    ├── cifar100  # This dataset will be downloaded automatically
    ├── ictext2021
    │   ├── train/
    │   ├── val/
    │   └── char_2_idx.json
    └── tiny-imagenet-200
        ├── train/
        ├── val/
        ├── val_annotations.txt
        ├── wnids.txt
        └── words.txt

Training

To train a model, run the following command: python train.py --cfg ./config/DFFDS/<DATASETNAME>/<DATASETNAME>_imb<IF>.yaml

DATASETNAME: Can be selected from {cifar10, cifar100, tiny_imagenet, ictext} \ IF: Can be selected from {001, 002, 01} for CIFAR and {001, 01} for Tiny-ImageNet and {001,0} for ICText

ICText IF=0 is the natural distribution with imbalance ratio of 18 after removing lower-case letters. Read the paper or visit ICText Dataset Repository for more details.

Evaluation

To evaluate a trained model from checkpoint, run the following command: python eval.py --cfg ./config/DFFDS/<DATASETNAME>/<DATASETNAME>_imb<IF>.yaml resume /path/to/checkpoint/

Results and Models

CIFAR-10

IF	Loss	Top-1 Acc.	Model
0.01	FFDS	75.60
	D-FFDS	79.93	Link
0.02	FFDS	79.82
	D-FFDS	82.94	Link
0.1	FFDS	87.46
	D-FFDS	88.48	Link

CIFAR-100

IF	Loss	Top-1 Acc.	Model
0.01	FFDS	40.74
	D-FFDS	43.46	Link
0.02	FFDS	45.67
	D-FFDS	48.48
0.1	FFDS	58.66
	D-FFDS	58.82	Link

Tiny-ImageNet

IF	Loss	Top-1 Acc.	Model
0.01	FFDS	42.34
	D-FFDS	43.86	Link
0.1	FFDS	56.11
	D-FFDS	58.31

ICText

IF	Loss	Top-1 Acc.
0.01	FFDS	76.89
	D-FFDS	79.56
0.0	FFDS	85.40
	D-FFDS	85.98

Citation

If you find our paper and repository useful, please cite

@inproceedings{icip2023_ffds,
  author = {Nah, Wan Jun and Ng, Chun Chet and Lin, Che-Tsung and Lee, Yeong Khang and Kew, Jie Long and Tan, Zhi Qin and Chan, Chee Seng and Zach, Christopher and Lai, Shang-Hong},
  booktitle = {2023 30th IEEE International Conference on Image Processing (ICIP)}, 
  title = {Rethinking Long-Tailed Visual Recognition with Dynamic Probability Smoothing and Frequency Weighted Focusing}, 
  year = {2023}
}

Feedback

Suggestions and opinions on this work (both positive and negative) are greatly welcomed. Please contact the authors by sending an email to nicolenahwj at gmail.com or cs.chan at um.edu.my.

License and Copyright

The project is open source under BSD-3 license (see the LICENSE file).

© 2023 Universiti Malaya.