Namespace Microsoft.ML.Probabilistic.Models.Attributes

Classes

Algorithm

Attribute which associates a specified algorithm to a targetted variable or statement. This is used for hybrid inference where different algorithms are used for different parts of the model

For expert use only! When sharing a variable between multiple models (e.g. using SharedVariable) you can add this attribute to have the variable be treated as a derived variable, even if it is not derived in the submodel where it appears.

DistributedCommunication

Attached to an index array.

DistributedCommunicationExpression

Attached to an index array.

DistributedSchedule

When attached to a Sequential Range, specifies which indices should be processed by each thread

DivideMessages

Attached to Variable objects to specify if outgoing messages should be computed by division

DoNotInfer

When attached to a variable, indicates that the variable will not be inferred, producing more efficient generated code.

FactorAlgorithm

Attribute which associates a specified algorithm to all factors that define a variable. This is used for hybrid inference where different algorithms are used for different parts of the model

GivePriorityTo

Attached to Variable or MethodInvoke to give priority in the operator search path

GroupMember

Group member attribute - attached to MSL variables based on inference engine groups

InitialiseBackward

When attached to a Variable, indicates that the backward messages to factors with NoInit attributes should be treated as initialised by the scheduler, even though they will be initialised to uniform

ListenToMessages

Attribute to cause message update events to be generated for the messages associated with the target variable

MarginalPrototype

Specifies a prototype marginal distribution for a variable. This attribute can be used to explicitly specify the marginal distribution type for a variable in cases where it cannot be deduced by the model compiler.

ParallelSchedule

When attached to a Sequential Range, specifies which indices should be processed by each thread

Partitioned

Attached to Ranges to specify that only one element should be in memory at a time (per thread)

PointEstimate

Attached to Variable objects to indicate that their uncertainty should be ignored during inference.
The inferred marginal will always be a point mass.

Sequential

Changes the order of inference updates in Expectation Propagation. When attached to a Range, indicates that the elements of VariableArrays indexed by the range should be updated sequentially rather than in parallel. This can sometimes accelerate convergence. Not supported for all models.

Interfaces

IAlgorithm

Interface for inference algorithms