public void Backwardがただしい()
{
var x = new Variable(Matrix <float> .Build.DenseOfArray(new float[, ] {
{ 1, 2, 3 }
}));
var y = new Variable(Matrix <float> .Build.DenseOfArray(new float[, ] {
{ 4, -1, 0.1f }
}));
var result = Multiply.ForwardStatic(
x, y
);
var target = new Variable(Matrix <float> .Build.DenseOfArray(new float[, ] {
{ 0.5f, 0.5f, 0.5f }
}));
var loss = MeanSquaredError.ForwardStatic(
result,
target
);
UnityEngine.Debug.Log(loss.Value);
loss.Backward();
UnityEngine.Debug.Log(result.Grad);
Helper.AssertMatrixAlmostEqual(
x.Grad,
Matrix <float> .Build.DenseOfArray(new float[, ]
{
{ 9.33333397f, 1.66666675f, -0.01333333f }
}) // chainer-python
);
}
public void 簡単なoptimizeしてみる()
{
var x = new Variable(Matrix <float> .Build.DenseOfArray(new float[, ] {
{ 1, 1, 1 }
}));
var constant = new Variable(Matrix <float> .Build.DenseOfArray(new float[, ] {
{ 1, 1, 1 }
}));
var target = new Variable(Matrix <float> .Build.DenseOfArray(new float[, ] {
{ 1, 2, 3 }
}));
const float lr = 0.1f;
var convergence = false;
for (var i = 0; i < 100; i++)
{
var loss = MeanSquaredError.ForwardStatic(
Add.ForwardStatic(x, constant),
target
);
x.Grad = null;
loss.Backward();
x.Value -= x.Grad * lr;
if (loss.Value[0, 0] < 0.1f)
{
convergence = true;
break;
}
}
Assert.IsTrue(convergence);
}
public void Backwardの式が正しい()
{
var x = new Variable(Matrix <float> .Build.DenseOfArray(new float[, ] {
{ 1, 2, 3 }
}));
var W = new Variable(Matrix <float> .Build.DenseOfArray(new float[, ]
{
{ 3, 4, 3 },
{ 1, 2, 3 },
{ 5, 4, 1 },
}));
var b = new Variable(Matrix <float> .Build.DenseOfArray(new float[, ] {
{ 4, 2, 1 }
}));
var target = new Variable(Matrix <float> .Build.DenseOfArray(new float[, ] {
{ 0.5f, 0.5f, 0.5f }
}));
var loss = MeanSquaredError.ForwardStatic(
Linear.ForwardStatic(x, W, b),
target
);
loss.Backward();
chainer.Helper.AssertMatrixAlmostEqual(
x.Grad,
Matrix <float> .Build.DenseOfArray(new float[, ] {
{ 112.33333588f, 127.33333588f, 89f }
})
);
}
public void ForwardBackwardで形が変わらない()
{
var func = new MeanSquaredError();
var x0 = new Variable(Matrix <float> .Build.DenseOfArray(new float[, ] {
{ 1, 1, 1 }
}));
var loss = func.Forward(
x0,
new Variable(Matrix <float> .Build.DenseOfArray(new float[, ] {
{ 1, 2, 3 }
}))
);
loss.CurrentGrad = loss.Value;
var gxs = func.Backward(loss).ToList();
Assert.AreEqual(gxs[0].ColumnCount, x0.Value.ColumnCount);
Assert.AreEqual(gxs[0].RowCount, x0.Value.RowCount);
}
public void Forwardできる()
{
var func = new MeanSquaredError();
var loss = func.Forward(
new Variable(Matrix <float> .Build.DenseOfArray(new float[, ] {
{ 1, 1, 1 }
})),
new Variable(Matrix <float> .Build.DenseOfArray(new float[, ] {
{ 1, 2, 3 }
}))
);
Helper.AssertMatrixAlmostEqual(
loss.Value,
Matrix <float> .Build.DenseOfArray(new float[, ] {
{ 1.6666666269302368f }
}),
delta: 0.00001f
);
}
public void Backardの式が正しい()
{
var testInput = new Variable(DenseMatrix.OfArray(new float[, ] {
{ 1, 2, 3 }
}));
var target = new Variable(DenseMatrix.OfArray(new float[, ] {
{ 0.5f, 0.5f, 0.5f }
}));
var loss = MeanSquaredError.ForwardStatic(
ReLU.ForwardStatic(testInput),
target
);
loss.Backward();
chainer.Helper.AssertMatrixAlmostEqual(
DenseMatrix.OfArray(new float[, ] {
{ 0.33333334f, 1f, 1.66666675f }
}), // by chainer.py
testInput.Grad
);
}
public void Backardの式が正しい()
{
var testInput = new Variable(DenseMatrix.OfArray(new float[, ] {
{ 1, 2, 3 }
}));
var target = new Variable(DenseMatrix.OfArray(new float[, ] {
{ 0.5f, 0.5f, 0.5f }
}));
var loss = MeanSquaredError.ForwardStatic(
Sigmoid.ForwardStatic(testInput),
target
);
loss.Backward();
chainer.Helper.AssertMatrixAlmostEqual(
testInput.Grad,
DenseMatrix.OfArray(new float[, ] {
{ 0.03028592f, 0.02665417f, 0.01363052f }
}) // by chainer.py
);
}
