public void TestDeserializeModel()
{
var stream = GetType().Assembly.GetManifestResourceStream("SRODecoderEngineTest.config.json");
var engine = new SRODecoderEngine.SRODecoderEngine(stream, null);
Assert.IsTrue(engine.Model.Layers.Count == 2);
// test that they are dense layers
Assert.IsTrue(engine.Model.Layers
.All(layer => layer.ClassName == "Dense"));
// test that first has batch input shape
Assert.IsTrue(engine.Model.Layers
.First().Config.BatchInputShape != null);
Assert.IsTrue(engine.Model.Layers
.Skip(1).All(layer => layer.Config.BatchInputShape == null));
// test that they have different names
Assert.IsTrue(engine.Model.Layers
.Select(layer => layer.Config.Name)
.Distinct().Count()
== engine.Model.Layers.Count());
// test if they are trainable
Assert.IsTrue(engine.Model.Layers.All(layer => layer.Config.Trainable));
// test that the data type is as expected
Assert.IsTrue(engine.Model.Layers.All(layer => layer.Config.DType == DType.Float32));
// test for activation values
Assert.IsTrue(engine.Model.Layers.All(layer => layer.Config.Activation == Activation.Linear));
}
public void TestLoadWeights()
{
var modelStream = GetType().Assembly.GetManifestResourceStream("SRODecoderEngineTest.config.json");
var weightStream = GetType().Assembly.GetManifestResourceStream("SRODecoderEngineTest.weights_posttrain.csv");
var engine = new SRODecoderEngine.SRODecoderEngine(modelStream, weightStream);
// check that each layer has same output as previous input
for (int nLayer = 0; nLayer < engine.Model.Layers.Count; nLayer++)
{
Layer currentLayer = engine.Model.Layers[nLayer];
double[,] kernelTensor = currentLayer.Variables["kernel"];
if (currentLayer.Config.UseBias)
{
Assert.IsTrue(currentLayer.Variables.ContainsKey("bias"));
double[,] biasTensor = currentLayer.Variables["bias"];
Assert.AreEqual(biasTensor.GetLength(0), 1);
Assert.AreEqual(biasTensor.GetLength(1), kernelTensor.GetLength(1));
}
else
{
Assert.IsFalse(currentLayer.Variables.ContainsKey("bias"));
}
Assert.AreEqual(currentLayer.Config.Units,
kernelTensor.GetLength(1));
if (nLayer > 0)
{
Assert.IsTrue(currentLayer.Config.BatchInputShape == null);
Layer previousLayer = engine.Model.Layers[nLayer - 1];
double[,] previousKernelTensor = previousLayer.Variables["kernel"];
Assert.AreEqual(previousKernelTensor.GetLength(1),
kernelTensor.GetLength(0));
}
else
{
Assert.IsTrue(currentLayer.Config.BatchInputShape != null);
Assert.AreEqual(currentLayer.Config.BatchInputShape[1].Value,
kernelTensor.GetLength(0));
}
}
}
public void TestDenseLayer()
{
var modelStream = GetType().Assembly.GetManifestResourceStream("SRODecoderEngineTest.config.json");
var weightStream = GetType().Assembly.GetManifestResourceStream("SRODecoderEngineTest.weights_posttrain.csv");
var engine = new SRODecoderEngine.SRODecoderEngine(modelStream, weightStream);
var predictStream = GetType().Assembly.GetManifestResourceStream("SRODecoderEngineTest.predict_posttrain.csv");
var predictions = SRODecoderEngine.SRODecoderEngine.ReadTensorsCsv(predictStream, 1, values => values[0]);
var predictionInTensor = predictions["in"];
var predictionOutTensor = predictions["out"];
var actualOutput = engine.Predict(predictionInTensor);
Assert.AreEqual(actualOutput.GetLength(0), predictionOutTensor.GetLength(0));
Assert.AreEqual(actualOutput.GetLength(1), predictionOutTensor.GetLength(1));
for (int n = 0; n < actualOutput.GetLength(0); n++)
{
for (int m = 0; m < actualOutput.GetLength(1); m++)
{
Assert.IsTrue(Math.Abs(actualOutput[n, m] - predictionOutTensor[n, m]) < 1e-3);
}
}
}