ZeroNeural.Core 1.2.0

ZeroNeural

ZeroNeural is a pure C# deep learning framework and automatic differentiation (Autograd) library for .NET with zero external dependencies. Providing a clean, PyTorch-like programming interface, it enables dynamic computational graph construction, reverse-mode automatic differentiation, neural network layer composition, and gradient optimization directly inside standard .NET applications without Python or native Torch binaries.

🌟 Key Capabilities

• Reverse-Mode Autograd Engine: Dynamic tape-based computational DAG recording operations and backpropagating analytical vector-Jacobian products (Backward()).
• PyTorch-Style Neural Layers (ZeroNeural.Core.nn):
  • Linear (Dense fully-connected layer with Xavier/Kaiming initialization)
  • Sequential (Composed container for multi-stage pipelines)
  • Conv2D (Spatial convolution with backpropagation)
  • Dropout (Inverted dropout for regularized training)
  • BatchNorm (Batch normalization with running mean/variance)
  • Activations: ReLU, Sigmoid, Tanh, Softmax, GELU
• Standard Loss Functions:
  • MseLoss (Mean Squared Error for regression)
  • CrossEntropyLoss (Multi-class classification with numerical log-sum-exp stability)
  • BceWithLogitsLoss (Binary classification with integrated sigmoid)
• First-Class Optimizers (ZeroNeural.Core.Optim):
  • Adam and AdamW (Adaptive Moment Estimation with decoupled weight decay)
  • SGD (Stochastic Gradient Descent with Nesterov momentum)
  • RMSProp
• Zero External Dependencies: Standard .NET runtime only.

📦 Installation

Install via the .NET CLI:

dotnet add package ZeroNeural.Core

🚀 Quick Start

1. Reverse-Mode Autograd Gradient Computation

using ZeroNeural.Core.Autograd;
using ZeroTensor.Core;

// Create leaf variable with gradient tracking
var x = new Variable(Tensor.Create(new[] { 1 }, new float[] { 3.0f }), requiresGrad: true);

// Compute y = x^2 + 2x + 1
var y = x * x + x * 2.0f + 1.0f;

// Backpropagate gradients: dy/dx = 2x + 2 = 8
y.Backward();

Console.WriteLine($"dy/dx at x=3: {x.Grad[0]} (Expected: 8.0)");

2. Training a Multi-Layer Perceptron (MLP)

using ZeroNeural.Core.nn;
using ZeroNeural.Core.Optim;
using ZeroNeural.Core.Loss;
using ZeroTensor.Core;

// Define model
var model = new Sequential(
    new Linear(inFeatures: 4, outFeatures: 16),
    new ReLU(),
    new Linear(inFeatures: 16, outFeatures: 2)
);

var optimizer = new AdamW(model.Parameters(), lr: 0.01f);
var criterion = new CrossEntropyLoss();

// Training step
var xTrain = Tensor.RandomUniform(32, 4);
var yTrain = Tensor.Create(new[] { 32 }, new int[] { /* target classes */ });

optimizer.ZeroGrad();
var pred = model.Forward(xTrain);
var loss = criterion.Compute(pred, yTrain);
loss.Backward();
optimizer.Step();

Console.WriteLine($"Epoch Loss: {loss.Value:F4}");

📊 Benchmark & Performance

Tested on Intel Core i7-13700K (Release x64):

📄 License

MIT License © 2026 Phong Võ. Part of the ZeroPlatform project.