TCSLBCNN (PyTorch)

A PyTorch implementation of TCS-LBCNN, a lightweight convolutional neural network that uses fixed sparse local binary pattern–inspired filters combined with learnable 1×1 convolutions.

This repository provides:

• A clean, installable Python package
• Ready-to-use Docker images
• Reproducible training and evaluation workflows

---

What is TCS-LBCNN?

TCS-LBCNN replaces standard 3×3 convolution kernels with fixed, sparse, directional contrast filters (two non-zero weights per kernel) followed by learnable channel mixing.

Benefits

• Fewer trainable parameters
• Strong inductive bias for local texture and edge patterns
• Suitable for lightweight or constrained environments

---

Supported datasets

• MNIST
• CIFAR-10 (optionally resized to 28×28 as configured)

---

Quick start (Docker – recommended)

The easiest way to use this project is via Docker. No Python installation required.

Pull the latest image

docker pull ghcr.io/ebrahimalwajih/tcslbcnn-pytorch-implementation:latest

Verify installation

docker run --rm ghcr.io/ebrahimalwajih/tcslbcnn-pytorch-implementation:latest \
  python -c "import tcslbcnn; print('tcslbcnn ready')"

Train (MNIST example)

docker run --rm \
  ghcr.io/ebrahimalwajih/tcslbcnn-pytorch-implementation:latest \
  python -c "from tcslbcnn.training import train; train(dataset='mnist', n_epochs=1, batch_size=64, use_compile=False)"

---

Running locally (Python)

Requirements

• Python ≥ 3.10
• pip

Install

pip install -r requirements.txt
pip install -e .

Train

python -c "from tcslbcnn.training import train; train(dataset='mnist', n_epochs=1, batch_size=64, use_compile=False)"

---

Output artifacts

During training, the following directories are created:

• artifacts/models/ — model checkpoints
• artifacts/runs/ — TensorBoard logs

To view training metrics:

tensorboard --logdir artifacts/runs

---

Project structure

src/tcslbcnn/      Core Python package
tests/             Unit tests
artifacts/          Training outputs (not committed)

---

Reproducibility

• Fixed initialization for TCS-LBP filters
• Deterministic training options
• Docker images ensure environment consistency

---

License

MIT License.

---

Citation

If you use this implementation in academic or research work, please cite the following paper:

E. Al-wajih and R. Ghazali, Threshold center-symmetric local binary convolutional neural networks for bilingual handwritten digit recognition, Knowledge-Based Systems, vol. 259, Article 110079, 2023. https://doi.org/10.1016/j.knosys.2022.110079

BibTeX

@article{ALWAJIH2023110079,
  title   = {Threshold center-symmetric local binary convolutional neural networks for bilingual handwritten digit recognition},
  author  = {Al-wajih, Ebrahim and Ghazali, Rozaida},
  journal = {Knowledge-Based Systems},
  volume  = {259},
  pages   = {110079},
  year    = {2023},
  issn    = {0950-7051},
  doi     = {10.1016/j.knosys.2022.110079},
  url     = {https://www.sciencedirect.com/science/article/pii/S0950705122011753}
}