quantified.uncertainty/squiggle
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Fixed some indexing errors in convolution code

Browse Source
This commit is contained in:
Sebastian Kosch 2020-06-26 22:37:24 -07:00
parent dc1ec1bb86
commit d2e7e5f928
5 changed files with 156 additions and 129 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

9
src/components/DistBuilder.re
View File

@ -171,7 +171,8 @@ let make = () => {
~schema,
~onSubmit=({state}) => {None},
~initialState={
guesstimatorString: "mm(normal(-10, 2), uniform(18, 25), lognormal({mean: 10, stdev: 8}), triangular(31,40,50))",
//guesstimatorString: "mm(normal(-10, 2), uniform(18, 25), lognormal({mean: 10, stdev: 8}), triangular(31,40,50))",
guesstimatorString: "uniform(0, 1) + normal(1, 2)",
domainType: "Complete",
xPoint: "50.0",
xPoint2: "60.0",
@ -180,9 +181,9 @@ let make = () => {
unitType: "UnspecifiedDistribution",
zero: MomentRe.momentNow(),
unit: "days",
sampleCount: "30000",
outputXYPoints: "10000",
downsampleTo: "1000",
sampleCount: "3000",
outputXYPoints: "100",
downsampleTo: "100",
kernelWidth: "5",
},
(),

156
src/distPlus/distribution/Distributions.re
View File

@ -154,6 +154,7 @@ module Continuous = {
// This is essentially like integrateWithTriangles, without the accumulation.
let toDiscretePointMasses = (t: t): DistTypes.discreteShape => {
let tl = t |> getShape |> XYShape.T.length;
let pointMassesX: array(float) = Belt.Array.make(tl - 1, 0.0);
let pointMassesY: array(float) = Belt.Array.make(tl - 1, 0.0);
let {xs, ys}: XYShape.T.t = t |> getShape;
for (x in 0 to E.A.length(xs) - 2) {
@ -163,75 +164,24 @@ module Continuous = {
x,
(xs[x + 1] -. xs[x]) *. ((ys[x] +. ys[x + 1]) /. 2.),
); // = dx * (1/2) * (avgY)
let _ =
Belt.Array.set(
pointMassesX,
x,
(xs[x] +. xs[x + 1]) /. 2.,
); // midpoints
();
};
{
xyShape: {
xs,
xs: pointMassesX,
ys: pointMassesY,
},
knownIntegralSum: t.knownIntegralSum,
};
};
/* Performs a discrete convolution between two continuous distributions A and B.
* It is an extremely good idea to downsample the distributions beforehand,
* because the number of samples in the convolution can be up to length(A) * length(B).
*
* Conventional convolution uses fn = (+.), but we also allow other operations to combine the xs.
*
* In practice, the convolution works by multiplying the ys for each possible combo of points of
* the two shapes. This creates a new shape for each point of A. These new shapes are then combined
* linearly. This may not always be the most efficient way, but it is probably the most robust for now.
*
* In the future, it may be possible to use a non-uniform fast Fourier transform instead (although only for addition).
*/
let convolveWithDiscrete = (fn, t1: t, t2: DistTypes.discreteShape) => {
let t1s = t1 |> getShape;
let t2s = t2.xyShape; // would like to use Discrete.getShape here, but current file structure doesn't allow for that
let t1n = t1s |> XYShape.T.length;
let t2n = t2s |> XYShape.T.length;
let outXYShapes: array(array((float, float))) =
Belt.Array.makeUninitializedUnsafe(t1n);
for (i in 0 to t1n - 1) {
// create a new distribution
let dxyShape: array((float, float)) =
Belt.Array.makeUninitializedUnsafe(t2n);
for (j in 0 to t2n - 1) {
let _ =
Belt.Array.set(
dxyShape,
j,
(fn(t1s.xs[i], t2s.xs[j]), t1s.ys[i] *. t2s.ys[j]),
);
();
};
let _ = Belt.Array.set(outXYShapes, i, dxyShape);
();
};
let combinedIntegralSum =
switch (t1.knownIntegralSum, t2.knownIntegralSum) {
| (None, _)
| (_, None) => None
| (Some(s1), Some(s2)) => Some(s1 *. s2)
};
outXYShapes
|> E.A.fmap(s => {
let xyShape = XYShape.T.fromZippedArray(s);
make(`Linear, xyShape, None);
})
|> reduce((+.))
|> updateKnownIntegralSum(combinedIntegralSum);
};
let convolve = (fn, t1: t, t2: t) =>
convolveWithDiscrete(fn, t1, toDiscretePointMasses(t2));
let mapY = (~knownIntegralSumFn=previousKnownIntegralSum => None, fn, t: t) => {
let u = E.O.bind(_, knownIntegralSumFn);
let yMapFn = shapeMap(XYShape.T.mapY(fn));
@ -350,6 +300,69 @@ module Continuous = {
XYShape.Analysis.getMeanOfSquaresContinuousShape,
);
});
/* Performs a discrete convolution between two continuous distributions A and B.
* It is an extremely good idea to downsample the distributions beforehand,
* because the number of samples in the convolution can be up to length(A) * length(B).
*
* Conventional convolution uses fn = (+.), but we also allow other operations to combine the xs.
*
* In practice, the convolution works by multiplying the ys for each possible combo of points of
* the two shapes. This creates a new shape for each point of A. These new shapes are then combined
* linearly. This may not always be the most efficient way, but it is probably the most robust for now.
*
* In the future, it may be possible to use a non-uniform fast Fourier transform instead (although only for addition).
*/
let convolveWithDiscrete = (~downsample=false, fn, t1: t, t2: DistTypes.discreteShape) => {
let t1s = t1 |> getShape;
let t2s = t2.xyShape; // would like to use Discrete.getShape here, but current file structure doesn't allow for that
let t1n = t1s |> XYShape.T.length;
let t2n = t2s |> XYShape.T.length;
let outXYShapes: array(array((float, float))) =
Belt.Array.makeUninitializedUnsafe(t1n);
for (i in 0 to t1n - 1) {
// create a new distribution
let dxyShape: array((float, float)) =
Belt.Array.makeUninitializedUnsafe(t2n);
for (j in 0 to t2n - 1) {
let _ =
Belt.Array.set(
dxyShape,
j,
(fn(t1s.xs[i], t2s.xs[j]), t1s.ys[i] *. t2s.ys[j]),
);
();
};
let _ = Belt.Array.set(outXYShapes, i, dxyShape);
();
};
let combinedIntegralSum = Common.combineIntegralSums((a, b) => Some(a *. b), t1.knownIntegralSum, t2.knownIntegralSum);
outXYShapes
|> E.A.fmap(s => {
let xyShape = XYShape.T.fromZippedArray(s);
make(`Linear, xyShape, None);
})
|> reduce((+.))
|> updateKnownIntegralSum(combinedIntegralSum);
};
let convolve = (~downsample=false, fn, t1: t, t2: t) => {
let downsampleIfTooLarge = (t: t) => {
let sqtl = sqrt(float_of_int(t |> getShape |> XYShape.T.length));
sqtl > 10. && downsample ? T.downsample(int_of_float(sqtl), t) : t;
};
let t1d = downsampleIfTooLarge(t1);
let t2d = downsampleIfTooLarge(t2);
convolveWithDiscrete(~downsample=false, fn, t1, toDiscretePointMasses(t2));
};
};
module Discrete = {
@ -489,6 +502,9 @@ module Discrete = {
let downsample = (~cache=None, i, t: t): t => {
// It's not clear how to downsample a set of discrete points in a meaningful way.
// The best we can do is to clip off the smallest values.
let currentLength = t |> getShape |> XYShape.T.length;
if (i < currentLength) {
let clippedShape =
t
|> getShape
@ -500,6 +516,9 @@ module Discrete = {
|> XYShape.T.fromZippedArray;
make(clippedShape, None); // if someone needs the sum, they'll have to recompute it
} else {
t;
}
};
let truncate =
@ -663,6 +682,8 @@ module Mixed = {
Continuous.T.Integral.sum(~cache=None, continuous);
let totalIntegralSum = discreteIntegralSum +. continuousIntegralSum;
// TODO: figure out what to do when the totalIntegralSum is zero.
let downsampledDiscrete =
Discrete.T.downsample(
int_of_float(
@ -811,17 +832,16 @@ module Mixed = {
};
});
let convolve = (fn: (float, float) => float, t1: t, t2: t): t => {
let convolve = (~downsample=false, fn: (float, float) => float, t1: t, t2: t): t => {
// Discrete convolution can cause a huge increase in the number of samples,
// so we'll first downsample.
// An alternative (to be explored in the future) may be to first perform the full convolution and then to downsample the result;
// to use non-uniform fast Fourier transforms (for addition only), add web workers or gpu.js, etc. ...
// TODO: make this optional or customizable
let downsampleIfTooLarge = (t: t) => {
let sqtl = sqrt(float_of_int(totalLength(t)));
sqtl > 10. ? T.downsample(int_of_float(sqtl), t) : t;
sqtl > 10. && downsample ? T.downsample(int_of_float(sqtl), t) : t;
};
let t1d = downsampleIfTooLarge(t1);
@ -830,11 +850,11 @@ module Mixed = {
// continuous (*) continuous => continuous, but also
// discrete (*) continuous => continuous (and vice versa). We have to take care of all combos and then combine them:
let ccConvResult =
Continuous.convolve(fn, t1d.continuous, t2d.continuous);
Continuous.convolve(~downsample=false, fn, t1d.continuous, t2d.continuous);
let dcConvResult =
Continuous.convolveWithDiscrete(fn, t2d.continuous, t1d.discrete);
Continuous.convolveWithDiscrete(~downsample=false, fn, t2d.continuous, t1d.discrete);
let cdConvResult =
Continuous.convolveWithDiscrete(fn, t1d.continuous, t2d.discrete);
Continuous.convolveWithDiscrete(~downsample=false, fn, t1d.continuous, t2d.discrete);
let continuousConvResult =
Continuous.reduce((+.), [|ccConvResult, dcConvResult, cdConvResult|]);
@ -870,7 +890,13 @@ module Shape = {
));
let convolve = (fn, t1: t, t2: t): t => {
Mixed(Mixed.convolve(fn, toMixed(t1), toMixed(t2)));
switch ((t1, t2)) {
| (Continuous(m1), Continuous(m2)) => DistTypes.Continuous(Continuous.convolve(~downsample=true, fn, m1, m2))
| (Discrete(m1), Discrete(m2)) => DistTypes.Discrete(Discrete.convolve(fn, m1, m2))
| (m1, m2) => {
DistTypes.Mixed(Mixed.convolve(~downsample=true, fn, toMixed(m1), toMixed(m2)))
}
};
};
let combine = (~knownIntegralSumsFn=(_, _) => None, fn, t1: t, t2: t) =>

2
src/distPlus/distribution/XYShape.re
View File

@ -9,7 +9,7 @@ let interpolate =
};
// TODO: Make sure that shapes cannot be empty.
let extImp = E.O.toExt("Should not be possible");
let extImp = E.O.toExt("Tried to perform an operation on an empty XYShape.");
module T = {
type t = xyShape;

4
src/distPlus/symbolic/SymbolicDist.re
View File

@ -283,11 +283,11 @@ module GenericDistFunctions = {
(~xSelection: [ | `Linear | `ByWeight]=`Linear, dist: dist, n) => {
switch (xSelection, dist) {
| (`Linear, _) => E.A.Floats.range(min(dist), max(dist), n)
| (`ByWeight, `Uniform(n)) =>
/* | (`ByWeight, `Uniform(n)) =>
// In `ByWeight mode, uniform distributions get special treatment because we need two x's
// on either side for proper rendering (just left and right of the discontinuities).
let dx = 0.00001 *. (n.high -. n.low);
[|n.low -. dx, n.low +. dx, n.high -. dx, n.high +. dx|];
[|n.low -. dx, n.low +. dx, n.high -. dx, n.high +. dx|]; */
| (`ByWeight, _) =>
let ys = E.A.Floats.range(minCdfValue, maxCdfValue, n);
ys |> E.A.fmap(y => inv(y, dist));

94
src/distPlus/symbolic/TreeNode.re
View File

@ -55,6 +55,52 @@ module TreeNode = {
type t = treeNode;
type simplifier = treeNode => result(treeNode, string);
let rec toString = (t: t): string => {
let stringFromStandardOperation =
fun
| `Add => " + "
| `Subtract => " - "
| `Multiply => " * "
| `Divide => " / "
| `Exponentiate => "^";
let stringFromPointwiseOperation =
fun
| `Add => " .+ "
| `Multiply => " .* ";
let stringFromFloatFromDistOperation =
fun
| `Pdf(f) => "pdf(x=$f, "
| `Inv(f) => "inv(c=$f, "
| `Sample => "sample("
| `Mean => "mean(";
switch (t) {
| `DistData(`Symbolic(d)) =>
SymbolicDist.GenericDistFunctions.toString(d)
| `DistData(`RenderedShape(s)) => "[shape]"
| `Operation(`StandardOperation(op, t1, t2)) =>
toString(t1) ++ stringFromStandardOperation(op) ++ toString(t2)
| `Operation(`PointwiseOperation(op, t1, t2)) =>
toString(t1) ++ stringFromPointwiseOperation(op) ++ toString(t2)
| `Operation(`ScaleOperation(_scaleOp, t, scaleBy)) =>
toString(t) ++ " @ " ++ toString(scaleBy)
| `Operation(`Normalize(t)) => "normalize(" ++ toString(t) ++ ")"
| `Operation(`FloatFromDist(floatFromDistOp, t)) => stringFromFloatFromDistOperation(floatFromDistOp) ++ toString(t) ++ ")"
| `Operation(`Truncate(lc, rc, t)) =>
"truncate("
++ toString(t)
++ ", "
++ E.O.dimap(Js.Float.toString, () => "-inf", lc)
++ ", "
++ E.O.dimap(Js.Float.toString, () => "inf", rc)
++ ")"
| `Operation(`Render(t)) => toString(t)
};
};
/* The following modules encapsulate everything we can do with
* different kinds of operations. */
@ -482,52 +528,6 @@ module TreeNode = {
| `Operation(op) => operationToDistData(sampleCount, op)
};
};
let rec toString = (t: t): string => {
let stringFromStandardOperation =
fun
| `Add => " + "
| `Subtract => " - "
| `Multiply => " * "
| `Divide => " / "
| `Exponentiate => "^";
let stringFromPointwiseOperation =
fun
| `Add => " .+ "
| `Multiply => " .* ";
let stringFromFloatFromDistOperation =
fun
| `Pdf(f) => "pdf(x=$f, "
| `Inv(f) => "inv(c=$f, "
| `Sample => "sample("
| `Mean => "mean(";
switch (t) {
| `DistData(`Symbolic(d)) =>
SymbolicDist.GenericDistFunctions.toString(d)
| `DistData(`RenderedShape(s)) => "[shape]"
| `Operation(`StandardOperation(op, t1, t2)) =>
toString(t1) ++ stringFromStandardOperation(op) ++ toString(t2)
| `Operation(`PointwiseOperation(op, t1, t2)) =>
toString(t1) ++ stringFromPointwiseOperation(op) ++ toString(t2)
| `Operation(`ScaleOperation(_scaleOp, t, scaleBy)) =>
toString(t) ++ " @ " ++ toString(scaleBy)
| `Operation(`Normalize(t)) => "normalize(" ++ toString(t) ++ ")"
| `Operation(`FloatFromDist(floatFromDistOp, t)) => stringFromFloatFromDistOperation(floatFromDistOp) ++ toString(t) ++ ")"
| `Operation(`Truncate(lc, rc, t)) =>
"truncate("
++ toString(t)
++ ", "
++ E.O.dimap(Js.Float.toString, () => "-inf", lc)
++ ", "
++ E.O.dimap(Js.Float.toString, () => "inf", rc)
++ ")"
| `Operation(`Render(t)) => toString(t)
};
};
};
let toShape = (sampleCount: int, treeNode: treeNode) => {
@ -539,7 +539,7 @@ let toShape = (sampleCount: int, treeNode: treeNode) => {
let continuous = Distributions.Shape.T.toContinuous(rs);
let discrete = Distributions.Shape.T.toDiscrete(rs);
let shape = MixedShapeBuilder.buildSimple(~continuous, ~discrete);
shape |> E.O.toExt("");
shape |> E.O.toExt("Could not build final shape.");
| Ok(_) => E.O.toExn("Rendering failed.", None)
| Error(message) => E.O.toExn("No shape found, error: " ++ message, None)
};