Shuffle PatchMix (SPM)

November 14, 2025 · View on GitHub

IEEE ICIP 2025 arXiv YouTube arXiv License: MIT

This is the official implementation of the ICIP 2025 paper "Shuffle PatchMix Augmentation with Confidence-Margin Weighted Pseudo-Labels for Enhanced Source-Free Domain Adaptation", by Prasanna Reddy Pulakurthi, Majid Rabbani, Jamison Heard, Sohail A. Dianat, Celso M. de Melo, and Raghuveer Rao.

Main figure

Demo - [Hugging Face Spaces]

An interactive demo for generating SPM augmentation.

HF Spaces

Installation

  1. Clone this repository.

    git clone https://github.com/PrasannaPulakurthi/SPM.git
cd SPM

  2. Install requirements using Python 3.9.

    conda create -n spm-env python=3.9
conda activate spm-env

  3. The code is tested with Pytorch 1.7.1, CUDA 11.0.

    pip install torch==1.7.1+cu110 torchvision==0.8.2+cu110 -f https://download.pytorch.org/whl/torch_stable.html

  4. Please also make sure to install additional packages using the following command.

    pip install -r requirements.txt

Reproduce PACS

Prepare dataset

Download PACS from Kaggle or the Official Website, and put it under ${DATA_ROOT}. The .txt files for the image labels are provided under ./datasets/PACS/. The prepared directory would look like this:

${DATA_ROOT}
├── PACS
│   ├── photo
│   ├── art_painting
│   ├── cartoon
│   ├── sketch
│   ├── photo_list.txt
│   ├── art_painting_list.txt
│   ├── cartoon_list.txt
│   ├── sketch_list.txt
│   ├── acs_list.txt
│   ├── pcs_list.txt

${DATA_ROOT} is set to ./datasets/ by default, which can be modified in configs/data/basic.yaml or via the hydra command line interface data.data_root=${DATA_ROOT}.

Training

The hydra is used as the configuration system. By default, the working directory is ./output, which can be changed directly from configs/root.yaml or via the hydra command line interface workdir=${WORK_DIR}.

PACS experiments are done for 6 domain shifts constructed from combinations of Photo, Art_Painting, Cartoon, and Sketch. Before the adaptation, we should have the source model. You may train the source model with the script scripts/train_PACS_source.sh as shown below. The pre-trained source models for seed 2022 can be found on Hugging Face.

After obtaining the source models, put them under ${SRC_MODEL_DIR}=./output/PACS/source and run scripts/train_PACS_target.sh to execute the adaptation.

# train source model
bash scripts/train_PACS_source.sh <SOURCE_DOMAIN>
# example: bash scripts/train_PACS_source.sh photo

# train SPM SFDA
bash scripts/train_PACS_target.sh <SOURCE_DOMAIN> <TARGET_DOMAIN> <SRC_MODEL_DIR>
# example: bash scripts/train_PACS_target.sh photo art_painting "output/PACS/source"

This will reproduce Tables 1 and 2 from the main paper:

Tables 1 & 2

For Windows users, the commands can be found in scripts_win/.

Reproduce VisDA-C

Prepare dataset

Please download the VisDA-C dataset, and put it under ${DATA_ROOT}. The .txt files for the image labels are provided under ./datasets/VISDA-C/. The prepared directory would look like this:

${DATA_ROOT}
├── VISDA-C
│   ├── train
│   ├── validation
│   ├── train_list.txt
│   ├── validation_list.txt

Training

VISDA-C experiments are done for train to validation adaptation. Before the adaptation, we should have the source model. You may train the source model with the script scripts/train_VISDA-C_source.sh as shown below. The pre-trained source models for seed 2022 can be found on Hugging Face.

After obtaining the source models, put them under ${SRC_MODEL_DIR}=./output/VISDA-C/source and run scripts/train_VISDA-C_target.sh to execute the adaptation.

# train source model
bash scripts/train_VISDA-C_source.sh

# train SPM SFDA
bash scripts/train_VISDA-C_target.sh <SRC_MODEL_DIR>
# example: bash scripts/train_VISDA-C_target.sh "output/VISDA-C/source"

This will reproduce the Table. 3 from the main paper:

Tables 3

For Windows users, the commands can be found in scripts_win/.

Reproduce DomainNet-126

Prepare dataset

Please download the DomainNet dataset (cleaned version), and put it under ${DATA_ROOT}. Notice that we follow MME to use a subset that contains 126 classes from 4 domains. The .txt files for the image labels are provided under ./datasets/domainnet-126/. The prepared directory would look like this:

${DATA_ROOT}
├── domainnet-126
│   ├── real
│   ├── sketch
│   ├── clipart
│   ├── painting
│   ├── real_list.txt
│   ├── sketch_list.txt
│   ├── clipart_list.txt
│   ├── painting_list.txt

Training

DomainNet-126 experiments are done for 7 domain shifts constructed from combinations of Real, Sketch, Clipart, and Painting. Before the adaptation, we should have the source model. You may train the source model with the script scripts/train_domainnet-126_source.sh as shown below. The pre-trained source models for seed 2022 can be found on Hugging Face.

After obtaining the source models, put them under ${SRC_MODEL_DIR}=./output/domainnet-126/source and run scripts/train_domainnet-126_target.sh to execute the adaptation.

# train source model
bash scripts/train_domainnet-126_source.sh <SOURCE_DOMAIN>
# example: bash scripts/train_domainnet-126_source.sh real

# train SPM SFDA
bash scripts/train_domainnet-126_target.sh <SOURCE_DOMAIN> <TARGET_DOMAIN> <SRC_MODEL_DIR>
# example: bash scripts/train_domainnet-126_target.sh real sketch "output/domainnet-126/source"

This will reproduce Table 4 from the main paper:

Tables 4

For Windows users, the commands can be found in scripts_win/.

Citation

If you find this work valuable for your research, we kindly request that you cite the following paper:

@INPROCEEDINGS{pulakurthi2025shuffle,
  author={Pulakurthi, Prasanna Reddy and Rabbani, Majid and Heard, Jamison and Dianat, Sohail and de Melo, Celso M. and Rao, Raghuveer},
  booktitle={2025 IEEE International Conference on Image Processing (ICIP)}, 
  title={Shuffle PatchMix Augmentation with Confidence-Margin Weighted Pseudo-Labels for Enhanced Source-Free Domain Adaptation}, 
  year={2025},
  volume={},
  number={},
  pages={1702-1707},
  keywords={Training;Image processing;Noise;Benchmark testing;Semisupervised learning;Picture archiving and communication systems;Data models;Reliability;Noise measurement;Overfitting;Source-Free Domain Adaptation;Classification;Contrastive Learning;Pseudo-Labels;Self-Training},
  doi={10.1109/ICIP55913.2025.11084606}}

Acknowledgement

Codebases from AdaContrast and DRA.