org.knowceans.topics.cgen
Class MixNetKernelGenerator

java.lang.Object
  org.knowceans.topics.cgen.MixNetKernelGenerator

public class MixNetKernelGenerator
extends java.lang.Object

reimplementation of generator code for support of: multiple sequences, sparse label indexing, parallelisms, fast sampling state, independent samplers and novel C3 and C5 mixnet structures.

Alternative development method: Start (1) generate existing LDA code programmatically, (2) parametrise the variable names in a MixNet structure that is created programmatically, (3) generalise structures, and (4) hook to mixnet script parser. (5) and (6) are parallel and serial accelerations.

This implementation directly makes use of the mixnet structural primitives, simplifying the generic solver into a straight-forward rule-based system.

Author:

gregor

Field Summary

double	alpha standard value for alpha
java.lang.String	corpus standard value for corpus
static boolean	directAssignments use direct assignments of samples in edge sequences z[m][n] etc., otherwise use temporary variables hz etc.
boolean	fastParallel generate parallelisation structures (copied counts)
boolean	fastSerial generate a fast sampling state (bounds-based method)
boolean	formatCode whether to format the code, default = true
boolean	indepSamplers generate each variable in its own sampler, ignoring dependencies
int	iter standard value for iterations
int	iterq standard value for query iterations
int	K standard value for K
boolean	nameSuffix add a suffix to the model name of the generated class, indicating independent, serial and parallel accellerations
int	parallelMode parallelisation model: 0 = active waiting, 1 = semaphore
static boolean	runModel whether to directly run the model
(package private) static java.util.Set<java.lang.String>	uniqueVars

Constructor Summary

MixNetKernelGenerator(MixNet model)

Method Summary

java.lang.String	beginClass()
java.lang.String	beginInit(boolean q)
java.lang.String	beginKernel(boolean q)
java.lang.String	beginMain()
java.lang.String	codeName() get the name as it should be used in the code
java.lang.String	computeWeights(MixSequence seq, boolean q, java.util.List<MixEdge> eeSamp) compute the weights for sampling
java.lang.String	constructor() create a standard constructor
java.lang.String	declareSelectors(boolean theta, boolean alpha) declare all selectors of the associated nodes
java.lang.String	endClass()
java.lang.String	endInit(boolean q)
java.lang.String	endKernel(boolean q)
java.lang.String	endMain()
java.lang.String	fields() generate the fields of a given MN, both for training and querying
java.lang.String	generateClassCode(MixNet lda)
java.lang.String	genVars() creates generic variables
java.lang.String	header()
java.lang.String	incTopics(MixSequence seq, boolean q, boolean inc, java.util.List<MixEdge> eeSamp) decrement topics of this sequence, basically decrementing nodes belonging to the edges ee
java.lang.String	init(boolean q) create an init routine
java.lang.String	kernel(boolean q) generate a calculation kernel
static void	main(java.lang.String[] args)
java.lang.String	makeMain() generate main method with standard calls to include corpus and run training and queries
java.lang.String	makePpx()
java.lang.String	makeToString()
java.lang.String	printWeights(MixSequence seq, boolean q, java.util.List<MixEdge> eeSamp, boolean mathml) compute the weights for sampling
java.lang.String	sampleTopics(java.util.List<MixEdge> eeSamp, boolean q) sample topics of all or active edges
java.lang.String	weightsArray() creates the array for sampling topics

Methods inherited from class java.lang.Object

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

Field Detail

uniqueVars

static java.util.Set<java.lang.String> uniqueVars

indepSamplers

public boolean indepSamplers

generate each variable in its own sampler, ignoring dependencies

formatCode

public boolean formatCode

whether to format the code, default = true

fastSerial

public boolean fastSerial

generate a fast sampling state (bounds-based method)

fastParallel

public boolean fastParallel

generate parallelisation structures (copied counts)

parallelMode

public int parallelMode

parallelisation model: 0 = active waiting, 1 = semaphore

TODO: change this to parallelisation modes in thesis

directAssignments

public static boolean directAssignments

use direct assignments of samples in edge sequences z[m][n] etc., otherwise use temporary variables hz etc. default = true

nameSuffix

public boolean nameSuffix

add a suffix to the model name of the generated class, indicating independent, serial and parallel accellerations

runModel

public static boolean runModel

whether to directly run the model

K

public int K

standard value for K

alpha

public double alpha

standard value for alpha

corpus

public java.lang.String corpus

standard value for corpus

iter

public int iter

standard value for iterations

iterq

public int iterq

standard value for query iterations

Constructor Detail

MixNetKernelGenerator

public MixNetKernelGenerator(MixNet model)

Method Detail

main

public static void main(java.lang.String[] args)
                 throws java.lang.Exception

Throws:

java.lang.Exception

generateClassCode

public java.lang.String generateClassCode(MixNet lda)

beginClass

public java.lang.String beginClass()

makeMain

public java.lang.String makeMain()

generate main method with standard calls to include corpus and run training and queries

Returns:

codeName

public java.lang.String codeName()

get the name as it should be used in the code

Returns:

beginMain

public java.lang.String beginMain()

endMain

public java.lang.String endMain()

endClass

public java.lang.String endClass()

header

public java.lang.String header()

fields

public java.lang.String fields()

generate the fields of a given MN, both for training and querying

Returns:

genVars

public java.lang.String genVars()

creates generic variables

Parameters:

u -

Returns:

weightsArray

public java.lang.String weightsArray()

creates the array for sampling topics

Parameters:

u -

Returns:

declareSelectors

public java.lang.String declareSelectors(boolean theta,
                                         boolean alpha)

declare all selectors of the associated nodes

Returns:

constructor

public java.lang.String constructor()

create a standard constructor

Returns:

init

public java.lang.String init(boolean q)

create an init routine

Parameters:

q - for querying

Returns:

beginInit

public java.lang.String beginInit(boolean q)

Parameters:

u -

q - for querying

Returns:

endInit

public java.lang.String endInit(boolean q)

kernel

public java.lang.String kernel(boolean q)

generate a calculation kernel

Returns:

beginKernel

public java.lang.String beginKernel(boolean q)

endKernel

public java.lang.String endKernel(boolean q)

computeWeights

public java.lang.String computeWeights(MixSequence seq,
                                       boolean q,
                                       java.util.List<MixEdge> eeSamp)

compute the weights for sampling

Parameters:

sequence -

u -

q -

eeSamp - if non-null, the edges to sample, otherwise all edges

Returns:

sampleTopics

public java.lang.String sampleTopics(java.util.List<MixEdge> eeSamp,
                                     boolean q)

sample topics of all or active edges

Parameters:

u -

eeSamp - active edges

Returns:

incTopics

public java.lang.String incTopics(MixSequence seq,
                                  boolean q,
                                  boolean inc,
                                  java.util.List<MixEdge> eeSamp)

decrement topics of this sequence, basically decrementing nodes belonging to the edges ee

Parameters:

u -

q -

inc - false for dec

ee - list of active edges

Returns:

makePpx

public java.lang.String makePpx()

makeToString

public java.lang.String makeToString()

printWeights

public java.lang.String printWeights(MixSequence seq,
                                     boolean q,
                                     java.util.List<MixEdge> eeSamp,
                                     boolean mathml)

compute the weights for sampling

Parameters:

sequence -

u -

q -

eeSamp - if non-null, the edges to sample, otherwise all edges

mathml - or latex (if false)

Returns: