public void TrainingTwiceStartsWhereFirstCycleEnded()
{
var tests = new[]
{
Tuple.Create(new[] { 1f, 0f, 0f }, new[] { 1f }),
Tuple.Create(new[] { 0f, 0f, 0f }, new[] { 0f }),
Tuple.Create(new[] { 1f, 1f, 0f }, new[] { 1f }),
Tuple.Create(new[] { 1f, 0f, 1f }, new[] { 1f }),
Tuple.Create(new[] { 1f, 1f, 1f }, new[] { 1f }),
Tuple.Create(new[] { 0f, 1f, 0f }, new[] { 0f }),
Tuple.Create(new[] { 0f, 1f, 1f }, new[] { 0f }),
Tuple.Create(new[] { 0, 0f, 1f }, new[] { 0f })
};
var description = SimpleDescriptionBuilder.GetDescription(3, new[] { 3, 3, 1 });
var net = Net.FromDescription(description);
WeightFiller.FillWeights(net, 0.001f);
var source = new CancellationTokenSource();
var token = source.Token;
var trainer = new Trainer(tests, net);
var firstError = trainer.Train(
learnFactor: 0.5f,
inertia: 0.2f,
desiredError: 0.0001f,
maxRuns: 50,
progress: j => { },
cancel: token);
var secondError = float.MaxValue;
trainer.Train(
learnFactor: 0.5f,
inertia: 0.2f,
desiredError: 0.0001f,
maxRuns: 2,
progress: j => { secondError = j.AvgError; },
cancel: token);
Assert.IsTrue(secondError < firstError);
}
public Guid Post(StocksTrainingJobRequest request)
{
// build description
request.HiddenLayerNodeCounts = request.HiddenLayerNodeCounts ?? new List <int>();
request.HiddenLayerNodeCounts.Add(1);
var description = SimpleDescriptionBuilder.GetDescription(4, request.HiddenLayerNodeCounts.ToArray());
foreach (var id in description.Outputs)
{
var outNode = description.Nodes.Single(n => n.NodeId == id);
outNode.Processor = null;
}
// get training events
var tests = _events.TrainingEvents
.Select(evt => Tuple.Create(evt.GetInputArray(), evt.GetOutputArray()));
// get net
var net = Net.FromDescription(description);
// initialize weights
WeightFiller.FillWeights(net, request.WeightVariance);
// create a trainer
var trainer = new Trainer(tests, net);
// add to the repository and return the id
return(_trainingJobRepository
.CreateProcess((progress, token) =>
{
trainer.Train(
learnFactor: request.InitialLearningRate,
inertia: request.InitialMomentum,
desiredError: request.DesiredError,
maxRuns: request.MaxIterations,
progress: progress,
cancel: token);
}));
}
public void TurtleZebraGiraffe()
{
var tests = new[]
{
Tuple.Create(new[] { 1f, 1f, 1f, 1f, 1f, 1f, 1f, 1f, 1f, 1f }, new[] { 0f, 0f, 1f }),
Tuple.Create(new[] { 1f, -1f, -1f, -1f, -1f, -1f, -1f, -1f, -1f, -1f }, new[] { 1f, 0f, 0f }),
Tuple.Create(new[] { 1f, 1f, 1f, 1f, -1f, -1f, -1f, -1f, -1f, -1f }, new[] { 0f, 1f, 0f })
};
var description = SimpleDescriptionBuilder.GetDescription(10, new[] { 5, 5, 5, 3 });
foreach (var node in description.Nodes.Where(node => description.Outputs.Contains(node.NodeId)))
{
node.Processor = "sigmoid";
}
var net = Net.FromDescription(description);
WeightFiller.FillWeights(net, .005f);
var trainer = new SimpleTrainer();
var error = trainer.Train(
net: net,
tests: tests,
desiredError: 0.01f,
maxEpochs: 20000,
learningRate: .5f);
Assert.IsTrue(error < .01f);
var eval = net.GetEvaluationFunction();
foreach (var test in tests)
{
var output = eval(test.Item1);
Console.WriteLine(string.Join(",", output));
}
}
public void TanhNetCanBeTrainedOnXOr()
{
var tests = new[]
{
Tuple.Create(new[] { 1f, -1f }, new[] { 1f }),
Tuple.Create(new[] { 1f, 1f }, new[] { -1f }),
Tuple.Create(new[] { -1f, 1f }, new[] { 1f }),
Tuple.Create(new[] { -1f, -1f }, new[] { -1f })
};
var initialDescription = new NetDescription
{
Nodes = new[]
{
new NodeDescription
{
NodeId = 0, Aggregator = "sum", Processor = "tanh",
Weight = .21f, Inputs = new []
{
new NodeInputDescription
{
FromInputVector = true, InputId = 0, Weight = -.07f
},
new NodeInputDescription
{
FromInputVector = true, InputId = 1, Weight = -.28f
}
}
},
new NodeDescription
{
NodeId = 1, Aggregator = "sum", Processor = "tanh",
Weight = -.29f, Inputs = new []
{
new NodeInputDescription
{
FromInputVector = true, InputId = 0, Weight = .41f
},
new NodeInputDescription
{
FromInputVector = true, InputId = 1, Weight = -.05f
}
}
},
new NodeDescription
{
NodeId = 2, Aggregator = "sum", Processor = "tanh",
Weight = .11f, Inputs = new []
{
new NodeInputDescription
{
FromInputVector = false, InputId = 0, Weight = -.1f
},
new NodeInputDescription
{
FromInputVector = false, InputId = 1, Weight = -.21f
}
}
}
},
Outputs = new[] { 2 }
};
var net = Net.FromDescription(initialDescription);
WeightFiller.FillWeights(net, .05f);
var trainer = new SimpleTrainer();
var error = trainer.Train(
net: net,
tests: tests,
desiredError: .001f,
maxEpochs: 100000,
learningRate: 5f);
Console.WriteLine(error);
Assert.IsTrue(error < 0.1f);
}
public void FullTrainToDescAndBack()
{
var tests = new[]
{
Tuple.Create(new[] { 1f, 0f, 0f }, new[] { 1f }),
Tuple.Create(new[] { 0f, 0f, 0f }, new[] { 0f }),
Tuple.Create(new[] { 1f, 1f, 0f }, new[] { 1f }),
Tuple.Create(new[] { 1f, 0f, 1f }, new[] { 1f }),
Tuple.Create(new[] { 1f, 1f, 1f }, new[] { 1f }),
Tuple.Create(new[] { 0f, 1f, 0f }, new[] { 0f }),
Tuple.Create(new[] { 0f, 1f, 1f }, new[] { 0f }),
Tuple.Create(new[] { 0, 0f, 1f }, new[] { 0f })
};
var description = SimpleDescriptionBuilder.GetDescription(3, new[] { 3, 1 });
var net = Net.FromDescription(description);
WeightFiller.FillWeights(net, 0.001f);
var source = new CancellationTokenSource();
var token = source.Token;
var trainer = new Trainer(tests, net);
var firstError = trainer.Train(
learnFactor: 0.5f,
inertia: 0.2f,
desiredError: 0.0001f,
maxRuns: 50,
progress: j => { },
cancel: token);
var desc2 = net.Description;
var net2 = Net.FromDescription(desc2);
var eval1 = net.GetEvaluationFunction();
var eval2 = net2.GetEvaluationFunction();
var avgError1 = 0d;
var avgError2 = 0d;
foreach (var test in tests)
{
var result1 = eval1(test.Item1);
var result2 = eval2(test.Item1);
avgError1 += Math.Pow(result1[0] - test.Item2[0], 2);
avgError2 += Math.Pow(result2[0] - test.Item2[0], 2);
}
avgError2 /= tests.Length;
Assert.IsTrue(avgError2 <= firstError);
Func <float[], float[]> a = ((float[] inputs) =>
{
var in0 = inputs[0];
var in1 = inputs[1];
var in2 = inputs[2];
var agg0 = (in0 * -2.458647) + (in1 * -0.07651551) + (in2 * -0.0518001) + -0.3729805;
var out0 = Math.Log(1 + Math.Exp(agg0));
var agg1 = (in0 * -2.523692) + (in1 * -0.05195593) + (in2 * -0.03320933) + -0.3351826;
var out1 = Math.Log(1 + Math.Exp(agg1));
var agg2 = (in0 * -2.458876) + (in1 * -0.07796045) + (in2 * -0.05187172) + -0.3720873;
var out2 = Math.Log(1 + Math.Exp(agg2));
var agg3 = (out0 * -0.8511373) + (out1 * -0.9345574) + (out2 * -0.8509701) + 4.971978;
var out3 = Math.Log(1 + Math.Exp(agg3));
return(new float[] { (float)out3 });
});
}
