public static FunctionBinder.ParamsFunction <X> FiniteDifference_ <X>(IList <IVar> inputs, Scalar <X> output, Tensor <X> .Shared x, IDictionary givens = null)
{
var epsilon = Scalar <X>("epsilon");
var i_ = Scalar <int>("i");
var j_ = Scalar <int>("j");
var k_ = Scalar <int>("k");
var indexes = new Scalar <int> .Var[0];
if (x.NDim == 1)
{
indexes = new[] { i_ }
}
;
if (x.NDim == 2)
{
indexes = new[] { i_, j_ }
}
;
if (x.NDim == 3)
{
indexes = new[] { i_, j_, k_ }
}
;
var inputSet = new List <IVar>();
foreach (var i in inputs)
{
inputSet.Add(i);
}
foreach (var i in indexes)
{
inputSet.Add(i);
}
inputSet.Add(epsilon);
var eps = Op.OneHot(x.Shape, indexes, epsilon);
var out_m_eps = (Scalar <X>)output.Patch(new Patch {
[x] = x - eps
});
var out_p_eps = (Scalar <X>)output.Patch(new Patch {
[x] = x + eps
});
var delta = (out_p_eps - out_m_eps) / (Numeric <X> .Two * epsilon);
return(Function(inputSet, delta, givens: givens));
}
/// <summary>
/// Creates a function that help checking the gradient backpropagated to a shared.
/// The function created will expect one argument for each given input and one float for "epsilon".
/// The function returns a gradient computed by finite difference and by the <paramref name="computed"/> expression.
/// </summary>
/// <typeparam name="X">float</typeparam>
/// <param name="inputs">The inputs of the graph. The created function will expect a value for each of these inputs</param>
/// <param name="output">The function to derive</param>
/// <param name="x">The gradient of x will be checked</param>
/// <param name="computed">The gradient to be checked. By default will use the "Grad" operator.</param>
/// <param name="givens"></param>
/// <returns>The test function</returns>
public static FunctionBinder.ParamsFunction <X, X> RandomGradientCheck <X>(IList <IVar> inputs, Scalar <X> output, Scalar <X> x, Scalar <X> computed = null, IDictionary givens = null)
{
var eps = Scalar <X>("epsilon");
var inputSet = new List <IVar>();
foreach (var i in inputs)
{
inputSet.Add(i);
}
inputSet.Add(eps);
var x_m_eps = x - eps;
var x_p_eps = x + eps;
if (x.Name != null)
{
x_m_eps.Name = x.Name + "_m_eps";
x_p_eps.Name = x.Name + "_p_eps";
}
var out_m_eps = (Scalar <X>)output.Patch(new Patch {
[x] = x_m_eps
});
var out_p_eps = (Scalar <X>)output.Patch(new Patch {
[x] = x_p_eps
});
if (output.Name != null && out_m_eps != output)
{
out_m_eps.Name = output.Name + "_m_eps";
out_p_eps.Name = output.Name + "_p_eps";
}
var finite = (out_p_eps - out_m_eps) / (Numeric <X> .Two * eps);
if (finite.IsZero)
{
Trace.WriteLine($"The given output {output} doesn't depend on {x}", "Warning");
}
finite.Name = nameof(finite);
computed = computed ?? Grad(output, x);
return(Function(inputSet, output: (finite, computed), givens: givens));
}
/// <summary>
/// Creates a function that help checking the gradient backpropagated to a shared.
/// The function created will expect one argument for each given input and one float for "epsilon".
/// The function returns a gradient computed by finite difference and by the <paramref name="computed"/> expression.
/// </summary>
/// <remarks>
/// "epsilon" is the length of the step used during finite difference.
/// The gradient is checked every time in a different direction.
/// </remarks>
/// <typeparam name="X">float</typeparam>
/// <param name="inputs">The inputs of the graph. The created function will expect a value for each of these inputs</param>
/// <param name="output">The function to derive</param>
/// <param name="x">The gradient of x will be checked</param>
/// <param name="computed">The gradient to be checked. By default will use the "Grad" operator.</param>
/// <param name="givens"></param>
/// <returns>The test function</returns>
public static FunctionBinder.ParamsFunction <X, X> RandomGradientCheck <X>(IList <IVar> inputs, Scalar <X> output, Tensor <X> x, Tensor <X> computed = null, IDictionary givens = null)
{
var epsilon = Scalar <X>("epsilon");
var inputSet = new List <IVar>();
foreach (var i in inputs)
{
inputSet.Add(i);
}
inputSet.Add(epsilon);
var eps = Random.Uniform(-epsilon, epsilon, x.Shape); eps.Name = nameof(eps);
var x_m_eps = x - eps;
var x_p_eps = x + eps;
if (x.Name != null)
{
x_m_eps.Name = x.Name + "_m_eps";
x_p_eps.Name = x.Name + "_p_eps";
}
var out_m_eps = (Scalar <X>)output.Patch(new Patch {
[x] = x_m_eps
});
out_m_eps.Name = output.Name + "_m_eps";
var out_p_eps = (Scalar <X>)output.Patch(new Patch {
[x] = x_p_eps
});
out_p_eps.Name = output.Name + "_p_eps";
var finite = (out_p_eps - out_m_eps); finite.Name = nameof(finite);
computed = computed ?? Grad(output, x);
var backpropagated = Numeric <X> .Two * Sum(eps * computed);
return(Function(inputSet, output: (finite, backpropagated), givens: givens));
}
public override NAry Clone(IReadOnlyList <IExpr> inputs)
{
var newVars = Vars.Select(v => new Scalar <Type> .Var(v.Name)).ToArray();
var patch = new Patch();
for (int i = 0; i < Vars.Length; ++i)
{
patch[Vars[i]] = newVars[i];
}
var abstraction = (Scalar <Type>)Abstraction.Patch(patch);
return((NAry)Create(inputs.Cast <Tensor <Type> >().ToArray(), newVars, abstraction));
}
