HelixNet 0.7.0

A powerful simple Neural Network building blocks based on MyGrad

A small, extensible, and lightweight framework for building powerful neural networks in Python, built on mygrad.

HelixNet is designed to be a transparent and easy-to-understand tool for learning and experimentation. It is powerful enough to run complex models like CNNs and LSTMs, but simple enough to run efficiently on modest hardware.

---

Key Features

• Lightweight & Simple: No complex compilation or heavy dependencies. Just plug and play.
• Extensible by Design: A clean, object-oriented structure (Layer, Optimiser) makes it easy to create your own custom layers and optimizers.
• Modern Architecture: Includes common layers like Dense, Conv2D, MaxPooling2D, LSTM, and Embedding.
• Powerful Optimizers: Comes with robust implementations of SGD (with momentum), RMSProp andAdam.
• Full Documentation: Comprehensive documentation available here.

Installation

# For the latest stable version
pip install HelixNet

# For development (editable) mode
pip install -e .

Quickstart: Training a Model

Here's a quick example of how to build and train a model on the classic "spiral" dataset.

import numpy as np
import mygrad as mg
import helixnet.layers as layers
import helixnet.optimisers as optimisers
import helixnet.activations as activations
import helixnet.models as models
from nnfs.datasets import spiral_data

# Create dataset
X, y = spiral_data(samples=100, classes=3)
X = mg.tensor(X)
y = mg.tensor(y, dtype=int)

# Build model
model = models.Sequental([
    layers.Dense(2, 64, activation=activations.ReLU),
    layers.Dense(64, 3, activation=(lambda x: x)) # Logits
])

optim = optimisers.Adam(learning_rate=0.05, decay=5e-7)

# Train the model
for epoch in range(10001):
    logits = model.forward(X)
    loss = mg.nnet.losses.softmax_crossentropy(logits, y)

    loss.backward()
    optim.optimise(model)
    model.null_grads()

    if epoch % 100 == 0:
        accuracy = np.mean(np.argmax(logits.data, axis=1) == y.data)
        print(f'Epoch: {epoch}, Loss: {loss.data:.4f}, Accuracy: {accuracy:.4f}')

Documentation

For a full guide to all layers, optimizers, and functionalities, please see the Full HelixNet Documentation.

Contributing

Contributions are welcome! If you find a bug or have a feature request, please open an issue. If you'd like to contribute code, please see the notes on our own contributions to mygrad (#445) for an example of our development philosophy.