eneural_net 2.0.0

AI Library to create efficient Artificial Neural Networks. Pure Dart + SIMD, portable to native, JS/Web, Wasm and Flutter, with optional Metal, CUDA and WebGPU acceleration.

eneural_net

eNeural.net / Dart is an AI Library for efficient Artificial Neural Networks. The library is portable (native, JS/Web, Wasm, Flutter) and the computation is capable to use SIMD (Single Instruction Multiple Data) to improve performance.

Training can also be GPU-accelerated, while keeping the same pure-Dart API: Metal (macOS) and CUDA (Windows/Linux) natively, and WebGPU in the browser.

Usage

import 'package:eneural_net/eneural_net.dart';
import 'package:eneural_net/eneural_net_extensions.dart';

void main() {
  // Type of scale to use to compute the ANN:
  var scale = ScaleDouble.ZERO_TO_ONE;

  // The samples to learn in Float32x4 data type:
  var samples = SampleFloat32x4.toListFromString(
    [
      '0,0=0',
      '1,0=1',
      '0,1=1',
      '1,1=0',
    ],
    scale,
    true, // Already normalized in the scale.
  );

  var samplesSet = SamplesSet(samples, subject: 'xor');

  // The activation function to use in the ANN:
  var activationFunction = ActivationFunctionSigmoid();

  // The ANN using layers that can compute with Float32x4 (SIMD compatible type).
  var ann = ANN(
    scale,
    // Input layer: 2 neurons with linear activation function:
    LayerFloat32x4(2, true, ActivationFunctionLinear()),
    // 1 Hidden layer: 3 neurons with sigmoid activation function:
    [HiddenLayerConfig(3, true, activationFunction)],
    // Output layer: 1 neuron with sigmoid activation function:
    LayerFloat32x4(1, false, activationFunction),
  );

  print(ann);

  // Training algorithm:
  var backpropagation = Backpropagation(ann, samplesSet);

  print(backpropagation);

  print('\n---------------------------------------------------');

  var chronometer = Chronometer('Backpropagation').start();

  // Train the ANN using Backpropagation until global error 0.01,
  // with max epochs per training session of 1000000 and
  // a max retry of 10 when a training session can't reach
  // the target global error:
  var achievedTargetError = backpropagation.trainUntilGlobalError(
          targetGlobalError: 0.01, maxEpochs: 50000, maxRetries: 10);

  chronometer.stop(operations: backpropagation.totalTrainingActivations);

  print('---------------------------------------------------\n');

  // Compute the current global error of the ANN:
  var globalError = ann.computeSamplesGlobalError(samples);

  print('Samples Outputs:');
  for (var i = 0; i < samples.length; ++i) {
    var sample = samples[i];

    var input = sample.input;
    var expected = sample.output;

    // Activate the sample input:
    ann.activate(input);

    // The current output of the ANN (after activation):
    var output = ann.output;

    print('- $i> $input -> $output ($expected) > error: ${output - expected}');
  }

  print('\nglobalError: $globalError');
  print('achievedTargetError: $achievedTargetError\n');

  print(chronometer);
}

Output:

ANN<double, Float32x4, SignalFloat32x4, Scale<double>>{ layers: 2+ -> [3+] -> 1 ; ScaleDouble{0.0 .. 1.0}  ; ActivationFunctionSigmoid }
Backpropagation<double, Float32x4, SignalFloat32x4, Scale<double>, SampleFloat32x4>{name: Backpropagation}

---------------------------------------------------
Backpropagation> [INFO] Started Backpropagation training session "xor". { samples: 4 ; targetGlobalError: 0.01 }
Backpropagation> [INFO] Selected initial ANN from poll of size 100, executing 600 epochs. Lowest error: 0.2451509315860858 (0.2479563313068569)
Backpropagation> [INFO] (OK) Reached target error in 2317 epochs (107 ms). Final error: 0.009992250436771877 <= 0.01
---------------------------------------------------

Samples Outputs:
- 0> [0, 0] -> [0.11514352262020111] ([0]) > error: [0.11514352262020111]
- 1> [1, 0] -> [0.9083549976348877] ([1]) > error: [-0.0916450023651123]
- 2> [0, 1] -> [0.9032943248748779] ([1]) > error: [-0.09670567512512207]
- 3> [1, 1] -> [0.09465821087360382] ([0]) > error: [0.09465821087360382]

globalError: 0.009992250436771877
achievedTargetError: true

Backpropagation{elapsedTime: 111 ms, hertz: 83495.49549549549 Hz, ops: 9268, startTime: 2021-05-26 06:25:34.825383, stopTime: 2021-05-26 06:25:34.936802}

SIMD (Single Instruction Multiple Data)

Dart has support for SIMD when computation is made using Float32x4 and Int32x4. The Activation Functions are implemented using Float32x4, improving performance by 1.5x to 2x, when compared to normal implementation.

The basic principle with SIMD is to execute math operations simultaneously in 4 numbers.

Float32x4 is a lane of 4 double (32 bits single precision floating points). Example of multiplication:

  var fs1 = Float32x4( 1.1 , 2.2 , 3.3  , 4.4  );
  var fs2 = Float32x4( 10  , 100 , 1000 , 1000 );
  
  var fs3 = fs1 * fs2 ;
  
  print(fs3);
  // Output:
  // [11.000000, 220.000000, 3300.000000, 4400.000000]

See "dart:typed_data library" and "Using SIMD in Dart".

Native Acceleration (macOS: CPU + Metal · Windows/Linux: CUDA)

The training loop can be offloaded to a native backend: on macOS via Apple Accelerate (BLAS/vDSP, CPU) or Metal (GPU); on Windows/Linux via CUDA (NVIDIA GPU). The network, its weights, the optimizer state and the full training sample set are uploaded once and each epoch runs entirely in native code (forward + backprop + weight update), reproducing the pure-Dart iRProp+/Backpropagation numerics within float32 tolerance.

Native acceleration is optional. The package stays pure-Dart (it still publishes to pub.dev and runs on the web); if the native library can't be built or loaded, the trainers transparently fall back to the pure-Dart SIMD path.

Building the native libraries

Nothing to do — the libraries are built for you.

hook/build.dart is a Dart build hook: the SDK runs it automatically on dart run, dart test and dart build, compiles the backend for the platform you are building for, and publishes each library as a code asset that dart:ffi resolves at runtime. The build is cached, so only the first run pays for it.

Platform	Built by the hook	Requires
macOS	libeneural_cpu_<arch>.dylib, libeneural_metal_<arch>.dylib	swiftc (Xcode command line tools)
Linux	libeneural_cuda_<arch>.so	NVIDIA CUDA Toolkit (nvcc + cuBLAS) on PATH
Windows	libeneural_cuda_<arch>.dll	NVIDIA CUDA Toolkit (nvcc + cuBLAS) on PATH, in a VS Developer Prompt

When a toolchain is missing the hook builds nothing and the pure-Dart path is used; it never fails your build. CUDA is only attempted when nvcc is on PATH.

dart compile exe is not supported for apps that depend on this package. The Dart SDK refuses dart compile exe for any project with build hooks in its dependency graph. Use dart build cli instead. Likewise, dart test --compiler exe runs without the native assets — use dart test --compiler cli.

Building by hand (debugging, or outside a Dart command)

The scripts under native/ mirror what the hook does. Their output is picked up by the loader's DynamicLibrary.open fallback, so a library built this way is used when the hook did not supply an asset:

# macOS (CPU + Metal):
bash native/macos/build.sh
# produces (git-ignored):
#   native/macos/cpu/build/libeneural_cpu_<arch>.dylib
#   native/macos/metal/build/libeneural_metal_<arch>.dylib

# Windows/Linux (CUDA — requires the NVIDIA CUDA Toolkit: nvcc + cuBLAS):
native\cuda\build.bat        # Windows
bash native/cuda/build.sh    # Linux
# produces (git-ignored):
#   native/cuda/build/libeneural_cuda_<arch>.{dll,so}

Note that this fallback only searches the working directory, so it applies to development inside a checkout, not to consuming the package from the pub cache.

Use an accelerated trainer

NativeRProp and NativeBackpropagation are drop-in replacements for RProp and Backpropagation on Float32x4 networks:

import 'package:eneural_net/eneural_net.dart';

var trainer = NativeRProp(
  ann,
  SamplesSet(samples, subject: 'xor'),
  backend: NativeBackend.auto, // auto | cpu | metal | cuda | none
);

print(trainer.activeBackend); // NativeBackend.cpu (or .metal / .cuda / .none)

trainer.trainUntilGlobalError(targetGlobalError: 1e-6);

Backend selection:

• auto (default) — picks the best backend available for the host: on macOS the CPU backend for small networks and Metal (GPU) for large ones (Metal has a fixed per-epoch overhead, so it only overtakes the CPU backend above ~16K weights); on Windows/Linux the CUDA (GPU) backend when available; otherwise pure Dart.
• cpu (macOS) — Apple Accelerate (BLAS/vDSP). ~2.4x faster than pure Dart on the 64 -> 256 -> 16 benchmark.
• metal (macOS) — Apple GPU. The trainer is batched: all samples are processed at once, so each epoch is a handful of MetalPerformanceShaders GEMMs plus elementwise kernels (dispatch count is independent of the sample count). On the 64 -> 256 -> 16 benchmark it trains ~3.6x faster than pure Dart, and the lead grows with network size (~2x over the CPU backend at ~80K weights).
• cuda (Windows/Linux) — NVIDIA GPU. Same batched whole-epoch design as Metal, with the three per-layer GEMMs on cuBLAS and the elementwise activation/delta/update steps as CUDA kernels (see native/cuda). Requires the CUDA Toolkit at build time (the build hook skips CUDA when nvcc is not on PATH) and an NVIDIA GPU at run time.
• none — forces the pure-Dart path.

Requesting a backend not available on the host (e.g. cuda on macOS, or metal on Windows) resolves to the pure-Dart path.

Unsupported networks (integer Int32x4 signals, or an activation function other than Linear/Sigmoid/SigmoidFast/SigmoidBoundedFast) also fall back to pure Dart automatically.

WebGPU (browser GPU)

In the browser, training can run on the GPU via WebGPU (WGSL compute shaders). Like the Metal backend it is a batched whole-epoch trainer, but WebGPU is asynchronous (GPU readback is a Future), so it is exposed through separate trainers with Future-returning methods rather than the synchronous train():

var trainer = WebGpuRProp(ann, samplesSet); // or WebGpuBackpropagation

if (await trainer.isWebGpuAccelerated) {
  print('training on the GPU');
}

await trainer.trainUntilGlobalErrorAsync(targetGlobalError: 1e-4);
// also: await trainer.trainAsync(epochs, targetError);

WebGpuRProp / WebGpuBackpropagation extend RProp / Backpropagation, so they keep the synchronous pure-Dart API too. When WebGPU is unavailable (not a browser, no WebGPU support, or an unsupported network) the async methods transparently fall back to the synchronous pure-Dart trainer — so the same code runs on the Dart VM and the web.

What runs on the device: the whole epoch. The weights, the optimizer state (iRProp+ steps/gradients) and the full sample set stay resident in GPU buffers; each epoch is a chain of WGSL compute passes (forward → output delta → backprop → gradients → weight update) over every sample at once, with a single readback for the epoch error. activateWebGpu(input) runs a single-sample forward pass on the GPU (the pure-Dart ann.activate() keeps working as usual).

Requirements: a browser with WebGPU (Chrome/Edge 113+, Safari 26+, Firefox 141+ on supported platforms), an ANN on Float32x4 signals, and activation functions among Linear/Sigmoid/SigmoidFast/SigmoidBoundedFast — anything else falls back to pure Dart.

See example/eneural_net_webgpu_example.dart.

WebGPU tests

test/eneural_net_webgpu_test.dart runs everywhere (it covers the async API and the pure-Dart fallback), while test/eneural_net_webgpu_integration_test.dart is browser-only and exercises the real GPU path: forward-pass parity with the pure-Dart activate() over several topologies/activations, exact weight upload/download round-trips, and iRProp+/Backpropagation epochs compared against the pure-Dart trainers.

The default chrome platform launches Chrome with --disable-gpu, which removes navigator.gpu, so dart_test.yaml defines two WebGPU-enabled platforms:

# Real GPU (desktop with a GPU):
dart test test/eneural_net_webgpu_integration_test.dart -p chrome_webgpu

# Software adapter (machines/CI runners without a GPU):
dart test test/eneural_net_webgpu_integration_test.dart -p chrome_webgpu_swiftshader

# Require a WebGPU device: no device -> failure instead of skip (used by CI):
dart test test/eneural_net_webgpu_integration_test.dart -p chrome_webgpu \
  --dart2js-args=-DWEBGPU_REQUIRED=true

Without a WebGPU device the integration tests self-skip, so they are safe to run anywhere. The WebGPU CI workflow runs them on Linux (SwiftShader) and macOS (real GPU), always with -DWEBGPU_REQUIRED=true so a job cannot pass through the fallback path, plus a job that checks the Dart VM fallback. Chrome on the hosted Windows runners exposes no WebGPU adapter, so there is no Windows job.

Training Algorithms

Beyond Backpropagation and RProp (iRProp+), the library ships a broad set of pure-Dart trainers (SIMD Float32x4):

• Gradient optimizers (GradientOptimizer seam, per-weight state, full SIMD): SGD (+ classic/Nesterov momentum), Adam (+ AdamW, Nadam, AMSGrad flags), RMSProp, AdaGrad, AdaDelta, Quickprop, Lion, and ResilientPropagation (RProp+/RProp−/iRProp+/iRProp−). All support mini-batch / online training (batchSize), L2 weight decay, gradient clipping, and pluggable learning-rate schedules (StepDecayStrategy, ExponentialDecayStrategy, CosineAnnealingStrategy, WarmupStrategy).
• Second-order (finite-difference based, ideal for small nets): ConjugateGradient, LBFGS, LevenbergMarquardt.
• Population / gradient-free: EvolutionStrategy, SeparableCMAES, GeneticAlgorithm, ParticleSwarm, DifferentialEvolution, SimulatedAnnealing.
• Regularization: dropout (per hidden layer via HiddenLayerConfig), L2 / weight decay, gradient clipping.

var adam = Adam(ann, samplesSet, learningRate: 0.001, weightDecay: 0.01);
adam.trainUntilGlobalError(targetGlobalError: 1e-4);

Selecting by name and checkpointing

Trainers can be built by name and resumed from a JSON checkpoint:

var trainer = trainingByName('adam', ann, samplesSet, params: {'learningRate': 0.01});

// Save / restore (JSON-serializable) — ANN weights + optimizer state:
var checkpoint = saveTrainingCheckpoint(trainer);
restoreTrainingCheckpoint(trainer, checkpoint); // into a same-config trainer

print(registeredTrainings()); // adam, adamw, nadam, rmsprop, lion, lbfgs, ...

See example/eneural_net_optimizers_example.dart.

Signal

The class Signal represents the collection of numbers (including its related operations) that will flow through the ANN, representing the actual signal that an Artificial Neural Network should compute.

The main implementation is SignalFloat32x4 and represents an ANN Signal based in Float32x4. All the operations prioritizes the use of SIMD.

The framework of Signal allows the implementation of any kind of data to represent the numbers and operations of an eNeural.net ANN. SignalInt32x4 is an experimental implementation to exercise an ANN based in integers.

Activation Functions

ActivationFunction is the base class for ANN neurons activation functions:

• ActivationFunctionSigmoid:

  The classic Sigmoid function (return for x a value between 0.0 and 1.0):

  activation(double x) {
    return 1 / (1 + exp(-x)) ;
  }
  

• ActivationFunctionSigmoidFast:

  Fast approximation version of Sigmoid function, that is not based in exp(x):

  activation(double x) {
    x *= 3 ;
    return 0.5 + ((x) / (2.5 + x.abs()) / 2) ;
  }
  

  Function author: Graciliano M. Passos: gmpassos@GitHub.

• ActivationFunctionSigmoidBoundedFast:

  Fast approximation version of Sigmoid function, that is not based in exp(x), bounded to a lower and upper limit for [x].

  activation(double x) {
    if (x < lowerLimit) {
      return 0.0 ;
    } else if (x > upperLimit) {
      return 1.0 ;
    }
    x = x / scale ;
    return 0.5 + (x / (1 + (x * x))) ;
  }
  

  Function author: Graciliano M. Passos: gmpassos@GitHub.

exp(x)

exp is the function of the natural exponent, e, to the power x.

This is an important ANN function, since is used by the popular Sigmoid function, and usually a high precision version is slow and approximation versions can be used for most ANN models and training algorithms.

Fast Math

An internal Fast Math library is present and can be used for platforms that are not efficient to compute exp (Exponential function).

You can import this library and use it to create a specialized ActivationFunction implementation or use it in any kind of project:

import 'package:eneural_net/eneural_net_fast_math.dart' as fast_math ;

void main() {
  // Fast Exponential function:
  var o = fast_math.exp(2);

  // Fast Exponential function with high precision:
  var highPrecision = <double>[0.0 , 0.0];
  var oHighPrecision = fast_math.expHighPrecision(2, 0.0, highPrecision);
  
  // Fast Exponential function with SIMD acceleration:
  var o32x4 = fast_math.expFloat32x4( Float32x4(2,3,4,5) );
}

The implementation is based in the Dart package Complex:

• https://github.com/rwl/complex/blob/master/lib/src/fastmath.dart

The fast_math.expFloat32x4 function was created by Graciliano M. Passos (gmpassos@GitHub).

eNeural.net

You can find more at: eneural.net

Source

The official source code is hosted @ GitHub:

• https://github.com/eneural-net/eneural_net_dart

Features and bugs

Please file feature requests and bugs at the issue tracker.

Contribution

Any help from the open-source community is always welcome and needed:

• Found an issue?
  • Please fill a bug report with details.
• Wish a feature?
  • Open a feature request with use cases.
• Are you using and liking the project?
  • Promote the project: create an article, do a post or make a donation.
• Are you a developer?
  • Fix a bug and send a pull request.
  • Implement a new feature, like other training algorithms and activation functions.
  • Improve the Unit Tests.
• Have you already helped in any way?
  • Many thanks from me, the contributors and everybody that uses this project!

Author

Graciliano M. Passos: gmpassos@GitHub.

License

Apache License - Version 2.0