edu.cmu.sphinx.linguist.language.ngram.trie

Class NgramTrieQuant

java.lang.Object

edu.cmu.sphinx.linguist.language.ngram.trie.NgramTrieQuant

public class NgramTrieQuant
extends java.lang.Object

Class for ngram weights quantation. Stores quantation tables for each ngram order for probabilities and backoffs, while ngrams store pointers to this tables. Size of table is specified by quantation parameter, for default QUANT_16 each weight has 65k entries table.

Nested Class Summary

Nested Classes

Modifier and Type	Class and Description
static class	NgramTrieQuant.QuantType

Constructor Summary

Constructors

Constructor and Description
NgramTrieQuant(int order, NgramTrieQuant.QuantType quantType)

Method Summary

Modifier and Type	Method and Description
int	getBackoffTableLen() Getter for length of backoffs quantation table.
int	getProbBoSize() Getter for size of quantaized weights in ngram trie.
int	getProbSize() Getter for size of quantaized weight in ngram trie
int	getProbTableLen() Getter for length of probability quantation table.
float	readBackoff(NgramTrieBitarr bitArr, int memPtr, int bitOffset, int orderMinusTwo) Reads encoded backoff from provided trie bit array and decodes it into actual value for specific ngram
float	readProb(NgramTrieBitarr bitArr, int memPtr, int bitOffset, int orderMinusTwo) Reads encoded probability from provided trie bit array and decodes it into actual value for specific ngram
void	setTable(float[] table, int order, boolean isProb) Setter that is used during quantation reading

Methods inherited from class java.lang.Object

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

Constructor Detail

NgramTrieQuant

public NgramTrieQuant(int order,
                      NgramTrieQuant.QuantType quantType)

Method Detail

setTable

public void setTable(float[] table,
                     int order,
                     boolean isProb)

Setter that is used during quantation reading

Parameters:

table - - array of weights to be used as quantation table

order - - ngrams order which quantation table corresponds to

isProb - - specifies if provided table is for probability (backoffs otherwise)

getProbTableLen

public int getProbTableLen()

Getter for length of probability quantation table.

Returns:

length of quantation table.

getBackoffTableLen

public int getBackoffTableLen()

Getter for length of backoffs quantation table.

Returns:

length of quantation table

getProbBoSize

public int getProbBoSize()

Getter for size of quantaized weights in ngram trie.

Returns:

amount of bits required to store quantized probability and backoff

getProbSize

public int getProbSize()

Getter for size of quantaized weight in ngram trie

Returns:

amount of bits required to store quantized probability

readProb

public float readProb(NgramTrieBitarr bitArr,
                      int memPtr,
                      int bitOffset,
                      int orderMinusTwo)

Reads encoded probability from provided trie bit array and decodes it into actual value for specific ngram

Parameters:

bitArr - - trie bit array

memPtr - - memory pointer for specific ngram order

bitOffset - - offset from memPtr that is calculated according to ngram index

orderMinusTwo - - order of ngram minus two

Returns:

probability of ngram

readBackoff

public float readBackoff(NgramTrieBitarr bitArr,
                         int memPtr,
                         int bitOffset,
                         int orderMinusTwo)

Reads encoded backoff from provided trie bit array and decodes it into actual value for specific ngram

Parameters:

bitArr - - trie bit array

memPtr - - memory pointer for specific ngram order

bitOffset - - offset from memPtr that is calculated according to ngram index

orderMinusTwo - - order of ngram minus two

Returns:

backoffs of ngram