public void TestAddTensorTensorBP()
{
TensorAllocator.InitDevices(ProcessorTypeEnums.CPU, new int[] { 0 });
var graph = new ComputeGraphTensor(new WeightTensorFactory(), 0, true);
var tensorA = new WeightTensor(new long[2] {
2, 2
}, 1, 0, name: "tensorA", isTrainable: true);
var tensorB = new WeightTensor(new long[2] {
2, 2
}, 2, 0, name: "tensorB", isTrainable: true);
var tensorSum = graph.Add(tensorA, tensorB);
tensorSum.CopyWeightsToGradients(tensorSum);
graph.Backward();
float gA = tensorA.GetGradientAt(new long[] { 1, 1 });
float gB = tensorB.GetGradientAt(new long[] { 1, 1, });
Assert.IsTrue(gA == 3.0f);
Assert.IsTrue(gB == 3.0f);
}
public void TestAddTensorTensor()
{
TensorAllocator.InitDevices(ProcessorTypeEnums.CPU, new int[] { 0 });
var graph = new ComputeGraphTensor(new WeightTensorFactory(), 0, true);
var tensorA = new WeightTensor(new long[2] {
2, 2
}, 1, 0, name: "tensorA", isTrainable: true);
var tensorB = new WeightTensor(new long[2] {
2, 2
}, 2, 0, name: "tensorB", isTrainable: true);
var tensorSum = graph.Add(tensorA, tensorB);
float v = tensorSum.GetWeightAt(new long[] { 1, 1 });
Assert.IsTrue(v == 3.0f);
}
public void TestAddSubGradients()
{
int batchSize = 5;
int vocabSize = 20;
TensorAllocator.InitDevices(ProcessorTypeEnums.CPU, new int[] { 0 });
var graph = new ComputeGraphTensor(new WeightTensorFactory(), 0, true);
var tensorA = new WeightTensor(new long[2] {
batchSize, vocabSize
}, 1, 0, name: "tensorA", isTrainable: true);
var tensorB = new WeightTensor(new long[2] {
batchSize, vocabSize
}, 1, 0, name: "tensorB", isTrainable: true);
var tensorIdx = BuildRandomLabelTensor(batchSize, vocabSize, "tensorIdx");
var tensorANeg = graph.Mul(tensorA, -1.0f);
var tensorANegSum = graph.Add(tensorANeg, 100.0f);
var tensorSub = graph.Sub(100.0f, tensorB);
float v1 = tensorANegSum.GetWeightAt(new long[] { 1, 1 });
float v2 = tensorSub.GetWeightAt(new long[] { 1, 1 });
Assert.IsTrue(v1 == v2);
var softmax1 = graph.Softmax(tensorANegSum);
var softmax2 = graph.Softmax(tensorSub);
graph.CrossEntropyLoss(softmax1, tensorIdx);
graph.CrossEntropyLoss(softmax2, tensorIdx);
graph.Backward();
float gA = tensorA.GetGradientAt(new long[] { 1, 1 });
float gB = tensorB.GetGradientAt(new long[] { 1, 1, });
Assert.IsTrue(gA == gB);
}
