A high-performance neural network library for Rust with CPU and Vulkan GPU support.
- 🚀 Dual Backend Support: Optimized CPU execution and Vulkan compute shaders
 - 🎯 Automatic Hardware Detection: Seamlessly selects between CPU and Vulkan GPU
 - đź§ Advanced Optimizers: Adam, SGD, and other optimization algorithms
 - 🏗️ Flexible Architecture: Dense layers, CNN, batch normalization, dropout, and custom layers
 - đź’ľ Model Persistence: Save/load models with metadata in multiple formats (Binary, JSON, MessagePack)
 - ⚡ Production Ready: SIMD optimizations, parallel processing, and zero-copy operations
 - đź”§ Comprehensive Training: Learning rate scheduling, early stopping, metrics tracking
 - 🎛️ Fine-grained Control: Custom loss functions, weight initialization, and gradient computation
 
Add this to your Cargo.toml:
[dependencies]
nnl = "0.1.0"use nnl::prelude::*;
fn main() -> Result<()> {
    // Select the CPU device
    let device = Device::cpu()?;
    // Create XOR training data
    // Each input and target is a separate Tensor, typically a batch of size 1 for this example.
    let train_inputs = vec![
        Tensor::from_slice_on_device(&[0.0, 0.0], &[1, 2], device.clone())?,
        Tensor::from_slice_on_device(&[0.0, 1.0], &[1, 2], device.clone())?,
        Tensor::from_slice_on_device(&[1.0, 0.0], &[1, 2], device.clone())?,
        Tensor::from_slice_on_device(&[1.0, 1.0], &[1, 2], device.clone())?,
    ];
    let train_targets = vec![
        Tensor::from_slice_on_device(&[0.0], &[1, 1], device.clone())?,
        Tensor::from_slice_on_device(&[1.0], &[1, 1], device.clone())?,
        Tensor::from_slice_on_device(&[1.0], &[1, 1], device.clone())?,
        Tensor::from_slice_on_device(&[0.0], &[1, 1], device.clone())?,
    ];
    // Create a simple neural network
    let mut network = NetworkBuilder::new()
        .add_layer(LayerConfig::Dense {
            input_size: 2,
            output_size: 8, // Hidden layer size
            activation: Activation::ReLU,
            use_bias: true,
            weight_init: WeightInit::Xavier,
        })
        .add_layer(LayerConfig::Dense {
            input_size: 8, // Input to output layer
            output_size: 1,
            activation: Activation::Sigmoid,
            use_bias: true,
            weight_init: WeightInit::Xavier,
        })
        .loss(LossFunction::MeanSquaredError) // Common for regression/binary classification
        .optimizer(OptimizerConfig::Adam {
            learning_rate: 0.01,
            beta1: 0.9,
            beta2: 0.999,
            epsilon: 1e-8,
            weight_decay: None,
            amsgrad: false,
        })
        .device(device.clone()) // Specify the device for the network
        .build()?;
    // Configure training
    let training_config = TrainingConfig {
        epochs: 1000,
        batch_size: 4, // Train on all 4 samples at once
        verbose: false, // Set to true for detailed epoch logs
        ..Default::default() // Use default values for other config fields
    };
    // Train the network
    network.train(&train_inputs, &train_targets, &training_config)?;
    // Make predictions and evaluate
    let test_input_00 = Tensor::from_slice_on_device(&[0.0, 0.0], &[1, 2], device.clone())?;
    let test_input_01 = Tensor::from_slice_on_device(&[0.0, 1.0], &[1, 2], device.clone())?;
    let test_input_10 = Tensor::from_slice_on_device(&[1.0, 0.0], &[1, 2], device.clone())?;
    let test_input_11 = Tensor::from_slice_on_device(&[1.0, 1.0], &[1, 2], device)?;
    let pred_00 = network.forward(&test_input_00)?.to_vec()?[0];
    let pred_01 = network.forward(&test_input_01)?.to_vec()?[0];
    let pred_10 = network.forward(&test_input_10)?.to_vec()?[0];
    let pred_11 = network.forward(&test_input_11)?.to_vec()?[0];
    // Print predictions, converting to binary (0 or 1)
    println!("\n--- XOR Predictions ---");
    println!("XOR(0,0) = {:.4} (class: {:.0})", pred_00, if pred_00 > 0.5 { 1.0 } else { 0.0 });
    println!("XOR(0,1) = {:.4} (class: {:.0})", pred_01, if pred_01 > 0.5 { 1.0 } else { 0.0 });
    println!("XOR(1,0) = {:.4} (class: {:.0})", pred_10, if pred_10 > 0.5 { 1.0 } else { 0.0 });
    println!("XOR(1,1) = {:.4} (class: {:.0})", pred_11, if pred_11 > 0.5 { 1.0 } else { 0.0 });
    println!("-------------------------");
    Ok(())
}[dependencies]
nnl = "0.1.0"[dependencies]
nnl = { version = "0.1.0", features = ["cpu-optimized"] }[dependencies]
nnl = { version = "0.1.0", features = ["intel-mkl"] }- Rust: 1.70 or later (edition 2024)
 - CPU: Any modern x86_64 or ARM64 processor
 - GPU (optional): Any Vulkan 1.2+ compatible GPU (AMD, Intel, NVIDIA)
 - OS: Linux, Windows, macOS
 
NNL uses Vulkan compute shaders for GPU acceleration, which works on:
- AMD GPUs: Radeon RX 400 series and newer
 - NVIDIA GPUs: GTX 900 series and newer
 - Intel GPUs: Arc series and modern integrated graphics
 
Run the included examples to see the library in action:
# Basic XOR problem (CPU)
cargo run --example xor
# XOR with GPU acceleration (if Vulkan GPU available)
cargo run --example xor_gpu
# MNIST digit classification
cargo run --example mnist
# MNIST with GPU
cargo run --example mnist_gpu
# Convolutional Neural Network
cargo run --example simple_cnn
# CNN with GPU support
cargo run --example simple_cnn_gpu
# Small MNIST examples for testing
cargo run --example mnist_small
cargo run --example mnist_small_gpuxor.rs- Solve XOR problem with a simple neural network (CPU)xor_gpu.rs- XOR with Vulkan GPU accelerationmnist.rs- MNIST handwritten digit classification (CPU)mnist_gpu.rs- MNIST with GPU accelerationmnist_small.rs- Smaller MNIST dataset for testing (CPU)mnist_small_gpu.rs- Small MNIST with GPUsimple_cnn.rs- Convolutional neural network (CPU)simple_cnn_gpu.rs- CNN with GPU acceleration
// Automatic device selection (prefers GPU if available, falls back to CPU)
let device = Device::auto_select()?;
// Specific device types
let cpu_device = Device::cpu()?;
let vulkan_device = Device::vulkan()?;  // May fail if no Vulkan GPU available
// Check device capabilities
println!("Device: {}", device.device_type());
println!("Memory: {:?}", device.info().memory_size);// Create tensors (uses auto-selected device)
let zeros = Tensor::zeros(&[3, 4])?;
let ones = Tensor::ones(&[2, 2])?;
let from_data = Tensor::from_slice(&[1.0, 2.0, 3.0], &[3])?;
// Create tensors on specific device
let device = Device::vulkan()?;
let gpu_tensor = Tensor::from_slice_on_device(&[1.0, 2.0, 3.0], &[3], device)?;
// Tensor operations
let a = Tensor::randn(&[2, 3])?;
let b = Tensor::randn(&[2, 3])?;
let result = a.add(&b)?;  // Element-wise addition
let matmul = a.matmul(&b.transpose(&[1, 0])?)?;  // Matrix multiplicationlet network = NetworkBuilder::new()
    .add_layer(LayerConfig::Dense {
        input_size: 784,
        output_size: 128,
        activation: Activation::ReLU,
        use_bias: true,
        weight_init: WeightInit::Xavier,
    })
    .add_layer(LayerConfig::Dropout { dropout_rate: 0.2 })
    .add_layer(LayerConfig::Dense {
        input_size: 128,
        output_size: 10,
        activation: Activation::Softmax,
        use_bias: true,
        weight_init: WeightInit::Xavier,
    })
    .loss(LossFunction::CategoricalCrossEntropy)
    .optimizer(OptimizerConfig::Adam {
        learning_rate: 0.001,
        beta1: 0.9,
        beta2: 0.999,
        epsilon: 1e-8,
        weight_decay: Some(1e-4),
        amsgrad: false,
    })
    .device(Device::auto_select()?)  // Automatically choose best device
    .build()?;let config = TrainingConfig {
    epochs: 100,
    batch_size: 32,
    verbose: true,
    early_stopping_patience: Some(10),
    early_stopping_threshold: 1e-4,
    validation_split: 0.2,
    shuffle: true,
    random_seed: Some(42),
    ..Default::default()
};
let history = network.train(&train_data, &train_labels, &config)?;
println!("Final loss: {:.4}", history.final_loss());use nnl::io::{save_model, load_model, ModelFormat};
// Save model
save_model(&network, "my_model.bin", ModelFormat::Binary, None)?;
// Load model
let loaded_network = load_model("my_model.bin")?;
// Save with metadata
let metadata = ModelMetadata {
    name: "MNIST Classifier".to_string(),
    description: "CNN for digit classification".to_string(),
    ..Default::default()
};
save_model(&network, "model_with_meta.json", ModelFormat::Json, Some(&metadata))?;Performance comparison on common tasks (Intel i7-10700K, RTX 3060 via Vulkan):
| Task | CPU (8 threads) | Vulkan GPU | Speedup | 
|---|---|---|---|
| Dense 1000x1000 MatMul | 12.5ms | 3.2ms | 3.9x | 
| Conv2D 224x224x64 | 145ms | 28ms | 5.2x | 
| MNIST Training (60k samples) | 45s | 18s | 2.5x | 
Note: Performance varies significantly based on GPU model and driver quality. Vulkan performance on NVIDIA may be lower than native CUDA.
- Use appropriate batch sizes: 32-128 for GPU, 8-32 for CPU
 - Enable CPU optimizations: Use 
features = ["cpu-optimized"]for OpenBLAS - Intel CPUs: Use 
features = ["intel-mkl"]for maximum CPU performance - Memory management: Call 
network.zero_grad()regularly to free unused memory - Data loading: Use parallel data loading for large datasets
 - GPU memory: Monitor GPU memory usage, reduce batch size if running out
 
| Feature | Description | Dependencies | 
|---|---|---|
default | 
CPU-optimized + examples | ["cpu-optimized"] | 
cpu-optimized | 
OpenBLAS acceleration | openblas-src | 
intel-mkl | 
Intel MKL acceleration | intel-mkl-src | 
Note: Vulkan support is always enabled and does not require a feature flag.
Vulkan not available
# Install Vulkan drivers and loader
# Ubuntu/Debian:
sudo apt install vulkan-tools vulkan-utils mesa-vulkan-drivers
# Verify Vulkan works:
vulkaninfo
# For NVIDIA GPUs, ensure latest drivers are installed
# For AMD GPUs on Linux, ensure AMDGPU driver is loadedSlow CPU performance
# Enable OpenBLAS optimizations
nnl = { version = "0.1.0", features = ["cpu-optimized"] }
# Or for Intel CPUs, use MKL:
nnl = { version = "0.1.0", features = ["intel-mkl"] }Out of memory on GPU
- Reduce batch size in 
TrainingConfig - Use smaller model architectures
 - Monitor GPU memory usage with 
nvidia-smior similar tools 
Compilation errors with MKL
# Ensure Intel MKL is properly installed
# Or switch to OpenBLAS:
nnl = { version = "0.1.0", features = ["cpu-optimized"] }Poor GPU performance
- Ensure you're using 
Device::vulkan()orDevice::auto_select() - Check that Vulkan drivers are up to date
 - Some operations may not be optimized for GPU yet
 - Consider using CPU with optimizations for small models
 
For detailed API documentation, see docs.rs/nnl.
Key modules:
tensor- Tensor operations and data structuresnetwork- Neural network building and traininglayers- Layer implementations and configurationsoptimizers- Optimization algorithmsdevice- Device management and backend selectionio- Model saving and loading
We welcome contributions! Please:
- Fork the repository
 - Create a feature branch (
git checkout -b feature/amazing-feature) - Make your changes with tests
 - Run 
cargo testandcargo clippy - Submit a pull request
 
For major changes, please open an issue first to discuss the proposed changes.
git clone https://github.com/hotplugindev/NNL.git
cd NNL
cargo build
cargo test
cargo run --example xor
# Test GPU functionality (requires Vulkan)
cargo run --example xor_gpu- CUDA Support: Native NVIDIA CUDA backend for better performance
 - ROCm Support: AMD ROCm backend for compute-focused workloads
 - Distributed Training: Multi-GPU support
 - Mobile Deployment: ARM optimization and model quantization
 - Web Assembly: Browser-based inference
 - Model Zoo: Pre-trained models for common tasks
 - Auto-ML: Neural architecture search
 - Graph Optimization: Operator fusion and memory optimization
 
- CUDA: Not yet supported (Vulkan used for NVIDIA GPUs)
 - ROCm: Not yet supported (Vulkan used for AMD GPUs)
 - Distributed Training: Single device only
 - Model Formats: Limited compared to PyTorch/TensorFlow
 - Layer Types: Growing but not comprehensive
 - Performance: Vulkan overhead may impact small models
 
This project is licensed under the MIT License - see the LICENSE file for details.
- Built on excellent Rust ecosystem crates: 
ndarray,rayon,vulkano - Inspired by PyTorch and TensorFlow APIs
 - Thanks to the Rust ML community and all contributors
 
Questions? Open an issue on GitHub.
