Syntropy

Syntropy is a lightweight research toolkit for experimenting with hybrid convolutional and attention-based neural architectures. The initial release packages both TensorFlow and PyTorch building blocks for Efficient Axial Networks (EffAxNet) with aligned APIs, making it easy to compare implementations across frameworks and prototype new ideas quickly.

Installation

Syntropy targets Python 3.8+ and ships optional extras for framework-specific dependencies:

pip install syntropy
pip install syntropy[tf]
pip install syntropy[torch]

The base install depends only on numpy. TensorFlow and PyTorch packages are delegated to extras to keep the default footprint small.

Package Layout

syntropy/
├── core          # Framework-agnostic utilities and registries
├── tf            # TensorFlow layers, models, and training loops
├── torch         # PyTorch mirrors of the TensorFlow components
└── examples      # Jupyter notebooks and end-to-end demos

Quick Start

from syntropy.tf.models import effaxnet_2d

model = effaxnet_2d.build_model(input_shape=(128, 128, 3), num_classes=10)
model.summary()

Development

1. Install development requirements:

   pip install -e .[tf,torch]
   

2. Run the unit tests:

   pytest
   

License

This project is licensed under the MIT License. See LICENSE for details.