private static void DoTest(NNInitParameters initPars, bool parallelize)
{
Console.WriteLine("Testing: {0}", initPars.GetType().Name + (parallelize ? " parallel" : " no parallel"));
int count = 10000;
var sw = new Stopwatch();
var arch = CreateArch(initPars);
using (var nn = arch.CreateNetwork())
using (var vectorBuff = nn.CreateVectorBuffer())
{
int cpus = Environment.ProcessorCount * 2;
var input = new float[10];
var output = Enumerable.Repeat(1.1f, 20).ToArray();
if (parallelize)
{
var contexts = new NeuralComputationContext[cpus];
for (int i = 0; i < cpus; i++) contexts[i] = nn.CreateContext();
sw.Start();
for (int i = 0; i < count; i+= cpus)
{
int ccount = cpus;
if (i + ccount > count) ccount = (i + ccount) - count;
System.Threading.Tasks.Parallel.For(0, ccount, x =>
{
var ctx = contexts[x];
nn.WriteInput(ctx, vectorBuff.GetOrCreate(x, 0, () => input));
nn.Reset(ctx);
nn.Iteration(ctx);
nn.Backpropagate(ctx, BackprogrationMode.FeedForward);
nn.ReadOutput(ctx, output);
});
}
sw.Stop();
}
else
{
using (var ctx = nn.CreateContext())
{
sw.Start();
for (int i = 0; i < count; i++)
{
nn.WriteInput(ctx, vectorBuff.GetOrCreate(i % cpus, 0, () => input));
nn.Reset(ctx);
nn.Iteration(ctx);
nn.Backpropagate(ctx, BackprogrationMode.FeedForward);
nn.ReadOutput(ctx, output);
}
sw.Stop();
}
}
}
Console.WriteLine(sw.Elapsed.TotalMilliseconds);
}
private static NeuralNetworkArchitecture CreateArch(NNInitParameters initPars)
{
var arch = new StandardMultilayerArchitecture
{
InitParameters = initPars,
InputSize = 10,
HiddenLayers = new[]
{
new ActivationLayer(200, new SigmoidActivationFunction(), new GradientDescentRule()),
new ActivationLayer(100, new SigmoidActivationFunction(), new GradientDescentRule())
},
OutputLayer = new ActivationLayer(20, new LinearActivationFunction(), new GradientDescentRule())
};
return arch;
}