PyCatML

Composable Machine Learning Pipelines via Category-Theoretic Abstractions

PyCatML is a Python framework implementing the theoretical foundations proposed in the paper:
"Composable Machine Learning Pipelines via Category-Theoretic Abstractions: A Functorial Approach to Modular Data Processing"

It provides a mathematical and computational infrastructure for modular, interpretable, and reusable machine learning pipelines using Category Theory.

Features

• Category Abstractions (Category, ImageCategory, TextCategory, etc.)
• Functor Classes representing end-to-end pipelines
• Natural Transformations for adaptive and interpretable computation
• Multimodal Fusion Pipelines using product categories
• Gradient Category and Optimizers with pullbacks and adjoint transformations
• Pipeline Composition, Data Augmentation, and Explanation

Project Structure

pycatml/
├── categories/
│   ├── base.py
│   ├── image.py
│   ├── text.py
│   ├── multimodal.py
│   └── rl.py
├── functors/
│   ├── base.py
│   ├── augmentation.py
│   └── fusion.py
├── transformations/
│   ├── natural_transformation.py
│   └── explanation.py
├── pipelines/
│   ├── mnist_pipeline.py
│   └── multimodal_pipeline.py
├── optimizer/
│   ├── gradient_category.py
│   └── categorical_optimizer.py
├── examples/
│   ├── run_mnist.py
│   └── run_multimodal.py
├── tests/
│   └── test_pipeline.py
├── utils/
│   └── validation.py
├── setup.py
└── README.md

Getting Started

git clone https://github.com/pouyashaeri/pycatml.git
cd pycatml
pip install -e .

Examples

MNIST Preprocessing Functor

from pycatml.pipelines.mnist_pipeline import MNISTPipeline

pipeline = MNISTPipeline()
image, label = pipeline.run()

Multimodal Fusion

from pycatml.pipelines.multimodal_pipeline import MultimodalPipeline

pipeline = MultimodalPipeline(vision_encoder, text_encoder, fusion_module)
fused_output = pipeline.run()

Evaluation

We provide benchmarks for:

• Accuracy
• Compositional Overhead
• Adaptivity Effectiveness
• Interpretability Gains

See examples/ and tests/ for runnable demos.