|
|
| MixtureLineage (Allocator_ const &allocator) |
| | An empty record, with no mark, whose shares come from allocator.
|
|
| MixtureLineage (MixtureLineage const &other) |
| | Copies the record into storage from other's allocator (see Allocation).
|
|
| MixtureLineage (MixtureLineage const &other, Allocator_ const &allocator) |
| | Copies other into storage from allocator.
|
|
| MixtureLineage (MixtureLineage &&other)=default |
| | Takes other's record.
|
|
| MixtureLineage (MixtureLineage &&other, Allocator_ const &allocator) |
| | Moves other into storage from allocator, taking its storage when the allocators are equal.
|
|
MixtureLineage & | operator= (MixtureLineage const &other)=default |
| | Copies other's record into this one's storage.
|
|
MixtureLineage & | operator= (MixtureLineage &&other)=default |
| | Takes other's record, as std::vector's move assignment does.
|
| allocator_type | get_allocator () const |
| Index | count () const |
| Index | parentCount () const |
| Index | size () const |
| std::span< Scalar const > | logShares (Index i) const |
| Scalar | logShare (Index i, Index parent) const |
| template<typename Parents, typename Children> |
| void | carry (Eigen::MatrixBase< Parents > const &parents, Eigen::MatrixBase< Children > &children, Scalar orphan=0) const |
| | Carries a value per parent to the components, as a statistic kept per hypothesis follows them through updates, merges, and prunes: children(i), one per component, becomes the mean of parents over component i's parents, weighted by its shares from them, so a component with one parent takes that parent's value and a merge takes the share-weighted mean of two; a component with no parent gets orphan.
|
template<RealScalar Scalar_, typename Allocator_ = DefaultAllocator>
class Emrys::MixtureLineage< Scalar_, Allocator_ >
How each component of a Mixture descends from the components it had at its last startLineage(): each component's share from each of those parents.
Row i holds component i's log-shares, one per parent, which sum to one: the fraction of its weight that came through each parent, \( P(\text{parent } n \mid \text{component } i) \). MixtureParams::startLineage() gives every component itself as its one parent, and from then on the mixture keeps the rows as its components change: a reweighted, moved, or replaced component keeps its row; a merge takes the weighted sum of the two rows; split() gives each child its parent's row; removal drops rows; and add() and append() give a new component a row of zeros, no parent at all. A filter that splits a component into several updated branches gives each its prior's row. This is the f-graph of Rahmathullah, Svensson, and Svensson (2014): the branches merged into each node and their weights, kept as the normalized share per parent rather than one weight per merged branch, which is what the backward pass of GaussianMixtureHistory reads.
count() counts the mixture's startLineage() calls, so a record of the lineage can say how many marks fell between two records. A copy of a mixture is a new mixture and has no lineage, so a filter's state swapped for a copy shows as a count that went back; a move keeps it. Until the first startLineage() there are no rows, and the mixture does no work to keep them.
The smoother is one reader. carry() is for another: a value kept per component outside the filter, such as each hypothesis's running NIS, which the rows take from the parents to the components through the updates, merges, and prunes between two marks (usage.md, "Checking an estimator").
The rows come from Allocator_, one shared buffer of parentCount() log-shares per component.
- Template Parameters
-
| Scalar_ | scalar type of the shares |
| Allocator_ | allocator for the shares (see Allocation) |
- See also
- MixtureParams::lineage(), GaussianMixtureHistory
template<RealScalar Scalar_, typename Allocator_ = DefaultAllocator>
template<typename Parents, typename Children>
| void Emrys::MixtureLineage< Scalar_, Allocator_ >::carry |
( |
Eigen::MatrixBase< Parents > const & | parents, |
|
|
Eigen::MatrixBase< Children > & | children, |
|
|
Scalar | orphan = 0 ) const |
|
inline |
Carries a value per parent to the components, as a statistic kept per hypothesis follows them through updates, merges, and prunes: children(i), one per component, becomes the mean of parents over component i's parents, weighted by its shares from them, so a component with one parent takes that parent's value and a merge takes the share-weighted mean of two; a component with no parent gets orphan.
parents has parentCount() values and children size(), both sized by the caller.
1.18.0