Skip to content


Repository files navigation

Build status ConvNetSharp NuGet package ConvNetSharp NuGet package


Started initially as C# port of ConvNetJS. You can use ConvNetSharp to train and evaluate convolutional neural networks (CNN).

Thank you very much to the original author of ConvNetJS (Andrej Karpathy) and to all the contributors!
ConvNetSharp relies on ManagedCuda library to acces NVidia's CUDA

3 ways to create neural network

Core.Layers Flow.Layers Computation graph
No computation graph Layers that create a computation graph behind the scene 'Pure flow'
Network organised by stacking layers Network organised by stacking layers 'Ops' connected to each others. Can implement more complex networks
Layers Layers Layers
E.g. MnistDemo E.g. MnistFlowGPUDemo or Flow version of Classify2DDemo E.g. ExampleCpuSingle

Example Code

Here's a minimum example of defining a 2-layer neural network and training it on a single data point:

using System;
using ConvNetSharp.Core;
using ConvNetSharp.Core.Layers.Double;
using ConvNetSharp.Core.Training.Double;
using ConvNetSharp.Volume;
using ConvNetSharp.Volume.Double;

namespace MinimalExample
    internal class Program
        private static void Main()
            // specifies a 2-layer neural network with one hidden layer of 20 neurons
            var net = new Net<double>();

            // input layer declares size of input. here: 2-D data
            // ConvNetJS works on 3-Dimensional volumes (width, height, depth), but if you're not dealing with images
            // then the first two dimensions (width, height) will always be kept at size 1
            net.AddLayer(new InputLayer(1, 1, 2));

            // declare 20 neurons
            net.AddLayer(new FullyConnLayer(20));

            // declare a ReLU (rectified linear unit non-linearity)
            net.AddLayer(new ReluLayer());

            // declare a fully connected layer that will be used by the softmax layer
            net.AddLayer(new FullyConnLayer(10));

            // declare the linear classifier on top of the previous hidden layer
            net.AddLayer(new SoftmaxLayer(10));

            // forward a random data point through the network
            var x =  BuilderInstance.Volume.From(new[] { 0.3, -0.5 }, new Shape(2));

            var prob = net.Forward(x);

            // prob is a Volume. Volumes have a property Weights that stores the raw data, and WeightGradients that stores gradients
            Console.WriteLine("probability that x is class 0: " + prob.Get(0)); // prints e.g. 0.50101

            var trainer = new SgdTrainer(net) { LearningRate = 0.01, L2Decay = 0.001 };
            trainer.Train(x, BuilderInstance.Volume.From(new[] { 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0 }, new Shape(1, 1, 10, 1))); // train the network, specifying that x is class zero

            var prob2 = net.Forward(x);
            Console.WriteLine("probability that x is class 0: " + prob2.Get(0));
            // now prints 0.50374, slightly higher than previous 0.50101: the networks
            // weights have been adjusted by the Trainer to give a higher probability to
            // the class we trained the network with (zero)

Fluent API (see FluentMnistDemo)

var net = FluentNet<double>.Create(24, 24, 1)
                   .Conv(5, 5, 8).Stride(1).Pad(2)
                   .Pool(2, 2).Stride(2)
                   .Conv(5, 5, 16).Stride(1).Pad(2)
                   .Pool(3, 3).Stride(3)


To switch to GPU mode:

  • add 'GPU' in the namespace: using ConvNetSharp.Volume.GPU.Single; or using ConvNetSharp.Volume.GPU.Double;
  • add BuilderInstance<float>.Volume = new ConvNetSharp.Volume.GPU.Single.VolumeBuilder(); or BuilderInstance<double>.Volume = new ConvNetSharp.Volume.GPU.Double.VolumeBuilder(); at the beggining of your code

You must have CUDA version 10.0 and cuDNN v7.6.4 (September 27, 2019), for CUDA 10.0 installed. cuDNN bin path should be referenced in the PATH environment variable.

Mnist GPU demo here

Save and Load Network

Serialization in ConvNetSharp.Core

using ConvNetSharp.Core.Serialization;


// Serialize to json 
var json = net.ToJson();

// Deserialize from json
Net deserialized = SerializationExtensions.FromJson<double>(json);

Serialization in ConvNetSharp.Flow

using ConvNetSharp.Flow.Serialization;


// Serialize to two files: MyNetwork.graphml (graph structure) / MyNetwork.json (volume data)

// Deserialize from files
var deserialized = SerializationExtensions.Load<double>("MyNetwork", false)[0];  // first element is the network (second element is the cost if it was saved along)