io.improbable.keanu.vertices

Class Vertex<T>

java.lang.Object

io.improbable.keanu.vertices.Vertex<T>

Direct Known Subclasses:

ContinuousVertex, DiscreteVertex, NonProbabilistic, Probabilistic

public abstract class Vertex<T>
extends java.lang.Object

Field Summary

Fields

Modifier and Type	Field and Description
static java.util.concurrent.atomic.AtomicLong	ID_GENERATOR

Constructor Summary

Constructors

Constructor and Description
Vertex()

Method Summary

Modifier and Type	Method and Description
void	addChild(Vertex<?> v)
void	addParent(Vertex<?> parent)
void	addParents(java.util.Collection<? extends Vertex> parents)
abstract java.util.Map<java.lang.Long,DoubleTensor>	dLogProb(T value) The partial derivatives of the natural log prob.
java.util.Map<java.lang.Long,DoubleTensor>	dLogProbAtValue()
boolean	equals(java.lang.Object o)
T	eval() This causes a backwards propagating calculation of the vertex value.
java.util.Set<Vertex>	getChildren()
java.util.Set<Vertex>	getConnectedGraph()
long	getId()
java.util.Set<Vertex>	getParents()
protected T	getRawValue()
int[]	getShape()
T	getValue()
int	hashCode()
boolean	hasValue()
boolean	isObserved()
abstract boolean	isProbabilistic()
T	lazyEval() This is similar to eval() except it only propagates as far up the graph as required until there are values present to operate on.
abstract double	logProb(T value) This is the natural log of the probability at the supplied value.
double	logProbAtValue()
void	observe(T value) This marks the vertex's value as being observed and unchangeable.
void	observeOwnValue() Cause this vertex to observe its own value, for example when generating test data
T	sample()
abstract T	sample(KeanuRandom random)
void	setAndCascade(T value) This sets the value in this vertex and tells each child vertex about the new change.
void	setParents(java.util.Collection<? extends Vertex> parents)
void	setParents(Vertex<?>... parents)
void	setValue(T value) Sets the value if the vertex isn't already observed.
void	unobserve()
abstract T	updateValue() This causes a non-probabilistic vertex to recalculate it's value based off it's parent's current values.

Methods inherited from class java.lang.Object

clone, finalize, getClass, notify, notifyAll, toString, wait, wait, wait

Field Detail

ID_GENERATOR

public static final java.util.concurrent.atomic.AtomicLong ID_GENERATOR

Constructor Detail

Vertex

public Vertex()

Method Detail

logProb

public abstract double logProb(T value)

This is the natural log of the probability at the supplied value. In the case of continuous vertices, this is actually the log of the density, which will differ from the probability by a constant.

Parameters:

value - The supplied value.

Returns:

The natural log of the probability density at the supplied value

logProbAtValue

public double logProbAtValue()

dLogProb

public abstract java.util.Map<java.lang.Long,DoubleTensor> dLogProb(T value)

The partial derivatives of the natural log prob.

Parameters:

value - at a given value

Returns:

the partial derivatives of the log density

dLogProbAtValue

public final java.util.Map<java.lang.Long,DoubleTensor> dLogProbAtValue()

sample

public abstract T sample(KeanuRandom random)

Parameters:

random - source of randomness

Returns:

a sample from the vertex's distribution. For non-probabilistic vertices, this will always be the same value.

sample

public T sample()

updateValue

public abstract T updateValue()

This causes a non-probabilistic vertex to recalculate it's value based off it's parent's current values.

Returns:

The updated value

lazyEval

public final T lazyEval()

This is similar to eval() except it only propagates as far up the graph as required until there are values present to operate on. On a graph that is completely uninitialized, this would be the same as eval()

Returns:

the value of the vertex based on the already calculated upstream values

eval

public final T eval()

This causes a backwards propagating calculation of the vertex value. This propagation only happens for vertices with values dependent on parent values i.e. non-probabilistic vertices. This will also cause probabilistic vertices that have no value to set their value by calling their sample method.

Returns:

The value at this vertex after recalculating any parent non-probabilistic vertices.

isProbabilistic

public abstract boolean isProbabilistic()

Returns:

True if the vertex is probabilistic, false otherwise. A probabilistic vertex is defined as a vertex whose value is not derived from it's parents. However, the probability of the vertex's value may be dependent on it's parents values.

setValue

public void setValue(T value)

Sets the value if the vertex isn't already observed.

Parameters:

value - the observed value

getValue

public T getValue()

getRawValue

protected T getRawValue()

hasValue

public boolean hasValue()

getShape

public int[] getShape()

setAndCascade

public void setAndCascade(T value)

This sets the value in this vertex and tells each child vertex about the new change. This causes a cascading change of values if any of the children vertices are non-probabilistic vertices (e.g. mathematical operations).

Parameters:

value - The new value at this vertex

observe

public void observe(T value)

This marks the vertex's value as being observed and unchangeable.

Non-probabilistic vertices of continuous types (integer, double) are prohibited from being observed due to it's negative impact on inference algorithms. Non-probabilistic booleans are allowed to be observed as well as user defined types.

Parameters:

value - the value to be observed

observeOwnValue

public void observeOwnValue()

Cause this vertex to observe its own value, for example when generating test data

unobserve

public void unobserve()

isObserved

public boolean isObserved()

getId

public long getId()

getChildren

public java.util.Set<Vertex> getChildren()

addChild

public void addChild(Vertex<?> v)

setParents

public void setParents(java.util.Collection<? extends Vertex> parents)

setParents

public void setParents(Vertex<?>... parents)

addParents

public void addParents(java.util.Collection<? extends Vertex> parents)

addParent

public void addParent(Vertex<?> parent)

getParents

public java.util.Set<Vertex> getParents()

equals

public boolean equals(java.lang.Object o)

Overrides:

equals in class java.lang.Object

hashCode

public int hashCode()

Overrides:

hashCode in class java.lang.Object

getConnectedGraph

public java.util.Set<Vertex> getConnectedGraph()