ai-dan-core 0.1.3

CPU-based neural network library written in C++

ai-dan-core

High-performance neural network library with C++ backend and Python bindings.

Features

• Fast C++ backend with adaptive multi-threading
• Simple Python API for easy integration
• Built-in MNIST support for quick experimentation
• Flexible architecture supporting multiple activation functions and loss functions
• Model persistence - save and load trained models

Installation

pip install ai-dan-core

Quick Start

import ai_dan

# Create a neural network
network = ai_dan.Network([784, 256, 128, 10],
                         ai_dan.LEAKY_RELU,
                         ai_dan.SOFTMAX)

# Configure training
optimizer = ai_dan.Optimizer(learning_rate=0.001, batch_size=32)
loss = ai_dan.CrossEntropy()

# Train on MNIST
trainer = ai_dan.Trainer(network, optimizer, loss)
trainer.train(ai_dan.MNIST())

# Save the model
network.save("model.ai-dan")

# Load and benchmark
loaded = ai_dan.Network.load("model.ai-dan")
benchmarker = ai_dan.Benchmarker(loaded, ai_dan.MNISTBenchmark())
benchmarker.benchmark()

Activation Functions

• ai_dan.SIGMOID
• ai_dan.RELU
• ai_dan.LEAKY_RELU
• ai_dan.TANH
• ai_dan.SOFTMAX

Loss Functions

• ai_dan.MSE (Mean Squared Error)
• ai_dan.CROSS_ENTROPY

Training with Callbacks

Monitor training progress with custom callbacks:

def on_step(step, loss_val, input_vec, label_vec, output_vec):
    if step % 1000 == 0:
        print(f"Step {step}, Loss: {loss_val:.4f}")

trainer = ai_dan.Trainer(network, optimizer, loss, on_step)
trainer.train(mnist_data)

Performance

• Training on MNIST: ~20 seconds for 60,000 samples
• Throughput: ~3,000 samples/second
• Typical accuracy: 96%+ on MNIST test set

Requirements

• Python 3.8+
• C++23 compatible compiler (GCC 13+, Clang 16+)

License

MIT