MyCaffe.common.BeamSearch< T > Class Template Reference

The BeamSearch uses the softmax output from the network and continually runs the net on each output (using the output as input) until the end of token is reached. The beam-search tree is returned. More...

Public Member Functions
	BeamSearch (Net< T > net)
	The constructor. More...
List< Tuple< double, bool, List< Tuple< string, int, double > > > >	Search (PropertySet input, int nK, double dfThreshold=0.01, int nMax=80)
	Perform the beam-search. More...

Detailed Description

The BeamSearch uses the softmax output from the network and continually runs the net on each output (using the output as input) until the end of token is reached. The beam-search tree is returned.

Template Parameters

T

Definition at line 18 of file BeamSearch.cs.

Constructor & Destructor Documentation

BeamSearch()

MyCaffe.common.BeamSearch< T >.BeamSearch

(

Net< T >

net

)

The constructor.

Parameters

net	Specifies the net used for the forward passes.

Definition at line 27 of file BeamSearch.cs.

Member Function Documentation

Search()

List< Tuple< double, bool, List< Tuple< string, int, double > > > > MyCaffe.common.BeamSearch< T >.Search	(	PropertySet	input,
		int	nK,
		double	dfThreshold = 0.01,
		int	nMax = 80
	)

Perform the beam-search.

Parameters

input	Specifies the input data (e.g. the encoder input)
nK	Specifies the beam width for the search.
dfThreshold	Specifies the threshold where detected items with probabilities less than the threshold are ignored (default = 0.01).
nMax	Specifies the maximum length to process (default = 80)

Returns

The list of top sequences is returned.

The beam-search algorithm is inspired by the article

See also

How to Implement a Beam Search Decoder for Natural Language Processing by Jason Brownlee, "Machine Learning Mastery", 2018

Definition at line 56 of file BeamSearch.cs.

---

The documentation for this class was generated from the following file:

• C:/Data/Data/SS_Projects/Intelligence/GitHub/MyCaffe/MyCaffe/common/BeamSearch.cs