Welcome to HeteroSymNN’s documentation!

HeteroSymNN is a symbolic JIT-Compiled Deep Learning framework for Heterogeneous Neural Networks.

Unlike standard frameworks that optimize for homogeneous layers, HeteroSymNN uses a Symbolic JIT Compiler to generate fused kernels at runtime. This allows every single neuron in a layer to have a distinct, custom mathematical activation function (e.g., sin(x), tanh(x), alpha * x + beta) with zero computational overhead.

Note

This project is tailored for Neuroevolution (NEAT), Control Systems, and Scientific Machine Learning.

How It Works

HeteroSymNN acts as a Differentiable Compiler:

  1. Parse: Accepts mathematical strings (e.g., "alpha * sin(x)") and parses them using SymPy.

  2. Derive: Automatically calculates the symbolic derivative for backpropagation.

  3. Compile: Generates C++ or CUDA code at runtime, creating a switch statement for distinct neuron instructions.

  4. Fuse: Fuses memory access into a single kernel launch.

Installation

Standard installation (CPU Python/JIT):

pip install heterosymnn

To enable high-performance GPU acceleration (requires NVIDIA Drivers & CuPy):

pip install heterosymnn[gpu]

For more details, see the installation guide

Quickstart: The “Cocktail” Layer

Here is a 5-line example of creating a network that mixes periodic (Sin) and linear features in the same layer.

from heterosymnn.core import FlexibleNN
from heterosymnn.api import Wraper

# Define a layer with 25 Sine neurons and 25 Linear neurons
model = FlexibleNN(
    input_size=10,
    layer_specs=[([(25, "sin(num)"), (25, "num")])]
)

# Train using Scikit-Learn style API
agent = Wraper(model)
agent.fit(X_train, y_train, epochs=100)

For more examples, see the quickstart guide

Documentation

User Guide:

API Reference:

Indices and tables