public void TrainOneMnist(double learningRate)
{
InputOutputPair pair = new InputOutputPair("E:\\Programs\\MNIST-JPG-master\\output\\training\\0\\1.jpg",
"E:\\Programs\\MNIST-JPG-master\\output\\training\\0");
model.TrainMnist(pair, learningRate);
}
public void SingleIterationTest()
{
// Initialize network
var layerSizes = new List <int>()
{
1, 1
};
var network = new NeuralNetwork(layerSizes);
network.InitializeWeights(0.5);
network.InitializeBiases(0.5);
// Initialize trainer
var trainer = new BackpropNetworkTrainer(network, 0.20f);
var example0 = new InputOutputPair(new float[] { 0.0f }, new float[] { -0.3f });
var example1 = new InputOutputPair(new float[] { 1.0f }, new float[] { 0.6f });
var examples = new List <InputOutputPair>()
{
example0, example1
};
const float trueWeight = 1.002666784763057f;
const float trueBias = -0.30951960420311175f;
// Track progress of weight and bias
float weight0 = network.Weights[0][0, 0];
float bias0 = network.Biases[0][0];
const int iterations = 50;
float[] weights = new float[iterations];
float[] biases = new float[iterations];
for (int i = 0; i < iterations; i++)
{
trainer.SingleIteration(examples);
weights[i] = network.Weights[0][0, 0];
biases[i] = network.Biases[0][0];
Console.WriteLine("weight: " + weights[i] + " bias: " + biases[i]);
// Verify eventual convergence of network parameters
if (i > iterations / 2)
{
Assert.IsTrue(Math.Abs(weights[i] - trueWeight) < Math.Abs(weights[i - 1] - trueWeight));
Assert.IsTrue(Math.Abs(biases[i] - trueBias) < Math.Abs(biases[i - 1] - trueBias));
}
}
}
