DASO: Distribution-Aware Semantics-Oriented Pseudo-label for Imbalanced Semi-Supervised Learning.

This repo provides the official pytorch implementation of our paper:

DASO: Distribution-Aware Semantics-Oriented Pseudo-label for Imbalanced Semi-Supervised Learning

Youngtaek Oh, Dong-Jin Kim, and In So Kweon

CVPR 2022

[Paper] [Project Page]

This includes the experiments on CIFAR-10, CIFAR-100, and STL-10 datasets with varying number of data-points and imbalance ratio in semi-supervised learning.

Abstract

The capability of the traditional semi-supervised learning (SSL) methods is far from real-world application due to severely biased pseudo-labels caused by (1) class imbalance and (2) class distribution mismatch between labeled and unlabeled data.

This paper addresses such a relatively under-explored problem. First, we propose a general pseudo-labeling framework that class-adaptively blends the semantic pseudo-label from a similarity-based classifier to the linear one from the linear classifier, after making the observation that both types of pseudo-labels have complementary properties in terms of bias. We further introduce a novel semantic alignment loss to establish balanced feature representation to reduce the biased predictions from the classifier. We term the whole framework as Distribution-Aware Semantics-Oriented (DASO) Pseudo-label.

We conduct extensive experiments in a wide range of imbalanced benchmarks: CIFAR10/100-LT, STL10-LT, and large-scale long-tailed Semi-Aves with open-set class, and demonstrate that, the proposed DASO framework reliably improves SSL learners with unlabeled data especially when both (1) class imbalance and (2) distribution mismatch dominate.

Figure 2: Analysis on recall and precision of pseudo-labels and the corresponding test accuracy obtained from FixMatch, USADTM, and DASO (Ours).

Installation

The code is tested with:

CUDA 10.1
Python 3.6
PyTorch 1.4.0
Single Titan Xp GPU (Mem: 12GB)

To install requirements:

(Option 1) Install with conda

#  installing environments via conda
conda create -n daso python=3.6 -y
conda activate daso

# installing required library
conda install pytorch==1.4.0 torchvision==0.5.0 cudatoolkit=10.0 -c pytorch  # pytorch for CUDA 10.0

pip install yacs
pip install future tensorboard

git clone https://github.com/ytaek-oh/daso.git && cd daso

(Option 2) Install with Docker

git clone https://github.com/ytaek-oh/daso.git && cd daso
bash docker/run_docker.sh  # see Dockfile therein if needed.

Code Structure

configs/: includes all the pre-defined configuration files for each baseline method and our DASO across all the benchmarks.
lib/algorithm/: includes the implementations of each baseline method.
lib/config/: holds all the configurations that can be manipulated.

Training

As note, each dataset is automatically downloaded to ./data/ when you firstly run the training.
You may refer to the pre-defined configuration files on ./configs/, or all the available default parameters are defined in lib/config/defaults.py.

To train the model(s) in the paper, run the command as:

python main.py --config-file {PATH_TO_CONFIG_FILE} {OTHER OPTIONS}

Example (1):

# run FixMatch + DASO on CIFAR10-LT with \N_1=500, \M_1=4000 and \gamma_l = \gamma_u = 100.
python main.py --config-file configs/cifar10/fixmatch_daso.yaml \
                DATASET.CIFAR10.NUM_LABELED_HEAD 500 DATASET.CIFAR10.NUM_UNLABELED HEAD 4000 \
                # Below is optional, if required
                SEED {SEED} GPU_ID {DEVICE} \
                {OTHER OPTIONS}

Example (2):

# run FixMatch + DASO under CIFAR100-LT with \N_1=150, \gamma_l = \gamma_u = 20.
python main.py --config-file configs/cifar100/fixmatch_daso.yaml \
                DATASET.CIFAR100.IMB_FACTOR_L 20 DATASET.CIFAR100.IMB_FACTOR_UL 20 \
                {OTHER OPTIONS}

Example (3):

# run FixMatch + DASO under STL10-LT with \N_1=150, \gamma_l = 20.
python main.py --config-file configs/stl10/fixmatch_daso.yaml \
                DATASET.STL10.NUM_LABELED_HEAD 150 DATASET.STL10.IMB_FACTOR_L 20 \
                {OTHER OPTIONS}

A commonly used template for the commands would be:

python main.py --config-file configs/{DATASET}/{ALGORITHM}.yaml \
        DATASET.{DATASET}.NUM_LABELED_HEAD {VALUE} DATASET.{DATASET}.NUM_UNLABELED_HEAD {VALUE} \
        DATASET.{DATASET}.IMB_FACTOR_L {VALUE} DATASET.{DATASET}.IMB_FACTOR_UL {VALUE} \
        (ALGORITHM.{ALGORITHM}.{OPTIONS} {VALUE} ...)

For data-specific setups and hyper-parameters in detail, please refer to the Implementation Details section in Appendix of our paper.

Monitoring Performance

Final configuration file, checkpoints, and all the logged results and metrics with tensorboard are produced in the path {OUTPUT_DIR} designated in the config file.

Citation

If you find our work useful for your research, please cite with the following bibtex:

@inproceedings{oh2021distribution,
  title = {DASO: Distribution-aware semantics-oriented pseudo-label for imbalanced semi-supervised learning},
  author = {Oh, Youngtaek and Kim, Dong-Jin and Kweon, In So},
  booktitle = {{IEEE} Conference on Computer Vision and Pattern Recognition (CVPR)},
  year = {2022},
  pages = {9786-9796}
}

ytaek-oh / daso