public override sealed void Backward(Scalar <Type> delta, Backpropagation bp)
{
_dx = _dx ?? Dx(x, y, this);
_dy = _dy ?? Dy(x, y, this);
bp.PushGradientTo(x, delta * _dx);
bp.PushGradientTo(y, delta * _dy);
}
public static Backpropagation Backward <T>(Scalar <T> expr, Scalar <T> delta)
{
var bp = new Backpropagation();
bp.Backpropagate(expr, delta);
return(bp);
}
protected Tensor <Type> D(int i)
{
bp = bp ?? Backpropagation.Backward(Abstraction, Numeric <Type> .One);
var _x = Vars[i];
var d_x = (Scalar <Type>)bp.ScalarDerivatives[_x];
return(Lift(d_x));
}
public static Backpropagation Backward <T>(Tensor <T> expr, Tensor <T> delta)
{
var bp = new Backpropagation();
delta.AssertOfShape(expr);
bp.Backpropagate(expr, delta);
return(bp);
}
public override void Backward(Tensor <Type> delta, Backpropagation bp)
{
if (delta.IsZero)
{
return;
}
foreach (var a in broadcast)
{
delta = Op.Sum(delta, axis: a, keepDims: true);
}
bp.PushGradientTo(x, delta);
}
public override sealed void Backward(Tensor <Type> delta, Backpropagation bp)
{
delta.AssertOfShape(Shape);
for (int i = 0; i < Inputs.Length; ++i)
{
var x = Inputs[i];
var deltaX = delta * D(i);
foreach (int axis in broadcast[x])
{
deltaX = Op.Sum(deltaX, axis, keepDims: true);
}
bp.PushGradientTo(x, deltaX);
}
// TODO: fix. This may push gradient using variables of the lambda outside of the Apply.
bp.PushGradientTo(Abstraction, Op.Sum(delta));
}
public override void Backward(Scalar <T> delta, Backpropagation bp)
{
if (_derivatives == null && Inputs.Count > 0)
{
throw new Exception($"The customOp {CustomFunctionName} doesn't support derivation.");
}
int n = Inputs.Count;
for (int i = 0; i < n; ++i)
{
if (_derivatives[i] != null)
{
dynamic x = Inputs[i];
dynamic d = _derivatives[i](Inputs, delta);
bp.PushGradientTo(x, d);
}
}
}
public override void Backward(Scalar <Type> delta, Backpropagation bp)
{
}
public static void Backward(this ITuple thiz, int item, object delta, Backpropagation bp)
{
dynamic x = thiz;
_Backward(x, item, delta, bp);
}
private static void _Backward <A, A_, B, B_>(ITuple <A, A_, B, B_> thiz, int item, object delta, Backpropagation bp)
where A : class, IExpr <A_>
where B : class, IExpr <B_>
{
if (item == 1)
{
thiz.Backward1((A)delta, bp);
}
else if (item == 2)
{
thiz.Backward2((B)delta, bp);
}
else
{
throw new ArgumentException(string.Format("There is no item {0} in this tuple-2.", item));
}
}
void Backward1(A delta, Backpropagation bp);
void Backward2(B delta, Backpropagation bp);
public override void Backward(Scalar <A_> delta, Backpropagation bp) => ITupleExtension.Backward(Parent, ItemIndex, delta, bp);
public override void Backward(Tensor <A_> delta, Backpropagation bp) => ITupleExtension.Backward(x, ItemIndex, delta, bp);
public abstract void Backward(Scalar <Type> delta, Backpropagation bp);
public abstract void Backward(Tensor <Type> delta, Backpropagation bp);
public override void Backward(Scalar <Type> delta, Backpropagation bp) => bp.PushGradientTo(x, delta * (_dx = _dx ?? Dx(x, this)));
private void AddDeltaFromBackpropagate(Loop backLoop, Dictionary <IExpr, IFor> others, Dictionary <ISymbol, IFor> forsDic, Backpropagation bp)
{
foreach (var partial_i in bp.TensorDerivatives)
{
if (!(partial_i.Key is ISymbol))
{
continue;
}
if (partial_i.Key is ITensorVar x && Loop.IsVariable(x)) // partial_i.Value = d(expr[j]) / d(partial_i.Key)
{
var dExpr_dx = partial_i.Value;
forsDic[x].Expression = _add(forsDic[x].Expression, dExpr_dx);
forsDic[x].Expression.Comment = $"dL/d{x}";
}
public override void Backward(Tensor <Type> deltas, Backpropagation bp)
{
deltas.AssertOfShape(Shape);
var deltaFromRecursive = OutputInfo != null;
// var in the forward -> for in the backward
var forsDic = new Dictionary <ISymbol, IFor>(); // ITensorSymbol
var backLoop = new Loop("d" + Loop.Name);
backLoop.Length = Loop.Length;
var substitution = new Patch(preserveShape: true);
// add the sequences used by the forward
int fwdSeqCount = Loop.Sequences.Count;
for (int i = 0; i < fwdSeqCount; i++)
{
var seq = Loop.Sequences[i];
var variable = Loop.Variable(seq);
var alias = Loop.Sequences[i].Match(
(Tensor <float> s) =>
backLoop.AddSeq(s[Step_m1], variable.Name + "_", Loop.SequenceAxes[i]),
(Tensor <int> s) =>
backLoop.AddSeq(s[Step_m1], variable.Name + "_", Loop.SequenceAxes[i]),
(Func <ITensor>)null
);
substitution.Add_(variable, alias);
}
// add the sequences computed by the forward
foreach (var @for in Loop.Fors)
{
if (@for.IsRecursive)
{
var variable = @for.RecursiveVariable;
var alias = @for.Match(
(Tensor <float> .For f) =>
backLoop.AddSeq(new Insert <float>(f, 0, f.OutputInfo, 0)[From_m2_Step_m1], variable.Name + "_", axis: 0),
(Tensor <int> .For f) =>
backLoop.AddSeq(new Insert <int>(f, 0, f.OutputInfo, 0)[From_m2_Step_m1], variable.Name + "_", axis: 0),
(Func <ITensor>)null
);
substitution.Add_(variable, alias);
}
else
{
var alias = @for.Match(
(Tensor <float> .For f) =>
backLoop.AddSeq(f[Step_m1], @for.Name + "_"),
(Tensor <int> .For f) =>
backLoop.AddSeq(f[Step_m1], @for.Name + "_"),
(Func <ITensor>)null
);
substitution.Add_(@for.Expression, alias);
}
}
// add the retropropagated delta
var deltaOut = backLoop.AddSeq(deltas[Step_m1], $"delta_{RecursiveVariable?.ToString() ?? "f" + Index}_", axis: 0);
// d_ avoid duplicated variables with the same name.
var d_ = new Dictionary <IVar, IVar>();
// add deltas of sequences (inputs of and computed by the forward), initialized to zero
var recVariables = Loop.RecursiveFors.Select(f => Loop.Variable(f));
foreach (var varFwd in Loop.Variables)
{
var zeros = varFwd.Match(
(Tensor <float> .Var x) => Op.ZerosLike(x),
(Tensor <int> .Var x) => Op.ZerosLike(x),
(Func <ITensor>)null
);
var @for = backLoop.AddRecursive_(zeros, zeros, $"d{varFwd.Name}_");
@for.Comment = $"dL/d{varFwd}";
d_[varFwd] = @for.RecursiveVariable;
forsDic[varFwd] = @for;
}
// `others` collect gradients pushed to expressions of the loop that aren't sequences or variables.
var others = new Dictionary <IExpr, IFor>();
AddDeltaFromBackpropagate(backLoop, others, forsDic, Backpropagation.Backward(Expression, deltaFromRecursive ? deltaOut + (Var)d_[RecursiveVariable] : deltaOut));
foreach (var @for in Loop.RecursiveFors)
{
var variable = @for.RecursiveVariable;
if (!deltaFromRecursive || @for != this)
{
var gradExpr = @for.Match(
(Tensor <float> .For f) => Backpropagation.Backward(f.Expression, (Tensor <float>)d_[f.RecursiveVariable]),
(Tensor <int> .For f) => Backpropagation.Backward(f.Expression, (Tensor <int>)d_[f.RecursiveVariable]),
null
);
AddDeltaFromBackpropagate(backLoop, others, forsDic, gradExpr);
}
// else: we already added the delta prior to the loop
// reuse results computed during the forward inside the backward
var alias_tp1 = backLoop.AddRecursive_(variable, @for[-1], variable.Name + "_tp1").RecursiveVariable;
substitution.Add_(@for.Expression, alias_tp1);
}
// Substitute variable in fors
foreach (var @for in backLoop.Fors)
{
var comment = @for.Expression.Comment;
@for.Expression = (ITensor)@for.Expression.Patch(substitution);
@for.Expression.Comment = comment;
}
// deltas of sequences
for (int i = 0; i < Loop.Sequences.Count; ++i)
{
if (Loop.Sequences[i] is Tensor <float> )
{
bp.PushGradientTo((Tensor <float>)Loop.Sequences[i], ((Tensor <float>)backLoop.Fors[i])[Step_m1]);
}
else
{
throw new NotImplementedException();
}
}
// deltas of seed
foreach (var @for in Loop.RecursiveFors)
{
if (@for is Tensor <float> )
{
bp.PushGradientTo((Tensor <float>)@for.OutputInfo, ((Tensor <float>)forsDic[@for.RecursiveVariable])[-1]);
}
else
{
throw new NotImplementedException();
}
}
// other deltas
foreach (var W_dW in others)
{
var W = W_dW.Key; var dW = W_dW.Value;
if (W is Tensor <float> )
{
bp.PushGradientTo((Tensor <float>)W, Op.Sum((Tensor <float>)dW, axis: 0));
}
else
{
throw new NotImplementedException();
}
}
}
