sticker

package module
v0.0.0-...-5d0fad7 Latest Latest
Warning

This package is not in the latest version of its module.

 Go to latest Published: May 17, 2018 License: MIT Imports: 12 Imported by: 0

README

Package sticker provides a framework for multi-label classification.

sticker logo

Build Status Report Status

Copyright 2017- Tatsuhiro Aoshima (hiro4bbh@gmail.com).

Introduction

Package sticker provides a framework for multi-label classification.

sticker is written in golang, so everyone can easily modify and compile it on almost every environments. You can see sticker's document on GoDoc.

Installation

First, download golang, and install it. Next, get and install sticker as follows:

go get github.com/hiro4bbh/sticker
go install github.com/hiro4bbh/sticker/sticker-util

Everything has been installed, then you can try sticker's utility command-line tool sticker-util now!

Prepare Datasets

First of all, you should prepare datasets. sticker assumes the following directory structure for a dataset:

+ dataset-root
|-- train.txt: training dataset
|-- text.txt: test dataset
|-- feature_map.txt: feature map (optional)
|-- label_map.txt: label map (optional)

Training and test datasets must be formatted as ReadTextDataset can handle (see GoDoc for data format). Feature and label maps should enumerate the name of each feature and label per line in order of identifier, respectively.

You can check the summary of the dataset at localhost:8080/summary as follows (you can change the port number with option addr):

sticker-util -verbose -debug <dataset-root> @summarize -table=<table-filename-relative-to-root>

If featureMap and labelMap is empty string, then feature and label maps are ignored, respectively.

Implemented Models

LabelNearest: Sparse Weighted Nearest-Neighbor Method

LabelNearest is Sparse Weighted Nearest-Neighbor Method (Aoshima+ 2018) which achieved SOTA performances on several XMLC datasets (Bhatia+ 2016). Recently, the model can process each data entry faster in 15.1 (AmazonCat-13K), 1.14 (Wiki10-31K), 4.88 (Delicious-200K), 15.1 (WikiLSHTC-325K), 4.19 (Amazon-670K), and 15.5 ms (Amazon-3M) on average, under the same settings of the paper (compare to the original result).

For example, you can test this method on Amazon-3M dataset (Bhatia+ 2016) as follows:

sticker-util -verbose -debug ./data/Amazon-3M/ @trainNearest @testNearest -S=75 -alpha=2.0 -beta=1

See the help of @trainNearest and @testNearest for the sub-command options.

LabelNear: A faster implementation of LabelNearest

LabelNear is a faster implementation of LabelNearest which uses the optimal Densified One Permutation Hashing (DOPH) and the reservoir sampling. This method can process every data entry in about 1 ms with little performance degradation. You can see the results on several XMLC datasets (Bhatia+ 2016) at Dropbox.

Almost parameters and options are same with the ones of LabelNearest. See the help of @trainNear and @testNear for details.

Other Models

Implemented in core
  • LabelConst: Multi-label constant model (see GoDoc)
  • LabelOne: One-versus-rest classifier for multi-label ranking (see GoDoc)
Implemented in plugin
  • LabelBoost: Multi-label Boosting model (see GoDoc)
  • LabelForest: Variously-modified FastXML model (see GoDoc)
  • LabelNext: Your next-generation model (you can add your own train and test commands, see plugin/next/init.go)

Implemented Binary Classifiers

  • L1Logistic_PrimalSGD: L1-logistic regression with stochastic gradient descent (SGD) solving the primal problem (see GoDoc)
  • L1SVC_PrimalSGD: L1-Support Vector Classifier with SGD solving the primal problem (see GoDoc)
  • L1SVC_DualCD: L1-Support Vector Classifier with coordinate descent (CD) solving the dual problem (see GoDoc)
  • L2SVC_PrimalCD: L2-Support Vector Classifier with CD solving the primal problem (see GoDoc)

References

  • (Aoshima+ 2018) T. Aoshima, K. Kobayashi, and M. Minami. "Revisiting the Vector Space Model: Sparse Weighted Nearest-Neighbor Method for Extreme Multi-Label Classification." arXiv:1802.03938, 2018.
  • (Bhatia+ 2016) K. Bhatia, H. Jain, Y. Prabhu, and M. Varma. The Extreme Classification Repository. 2016. Retrieved January 4, 2018 from http://manikvarma.org/downloads/XC/XMLRepository.html

Documentation

Overview

Package sticker provides a framework for multi-label classification.

Index

Constants

This section is empty.

Variables

View Source 
var BinaryClassifierTrainers = map[string]BinaryClassifierTrainer{
	"L1Logistic_PrimalSGD": BinaryClassifierTrainer_L1Logistic_PrimalSGD,
	"L1SVC_PrimalSGD":      BinaryClassifierTrainer_L1SVC_PrimalSGD,
}

BinaryClassifierTrainers is the map from the binary classifier trainer name to the corresponding binary classifier trainer.

Functions

func Abs32 

func Abs32(x float32) float32

Abs32 is the float32-version of math.Abs.

func AvgTotalVariationAmongSparseVectors 

func AvgTotalVariationAmongSparseVectors(svs SparseVectors) float32

AvgTotalVariationAmongSparseVectors returns the average total-variation distance among the sparse vectors. This function returns 0.0 if there is at most one sparse vector, and returns NaN if some sparse vectors are empty.

func Ceil32 

func Ceil32(x float32) float32

Ceil32 is the float32-version of math.Ceil.

func ClassifyAllToBinaryClass 

func ClassifyAllToBinaryClass(Z []float32) []bool

ClassifyAllToBinaryClass returns the bool slice whose entry indicates positive label if the corresponding z value is positive, otherwise negative label.

func ClassifyToBinaryClass 

func ClassifyToBinaryClass(z float32) bool

ClassifyToBinaryClass returns true indicating positive label if the z value is positive, otherwise false indicating negative label.

func DecodeDataset 

func DecodeDataset(ds *Dataset, r io.Reader) error

DecodeDataset decodes Dataset from r. Directly passing *os.File used by a gob.Decoder to this function causes mysterious errors. Thus, if users use gob.Decoder, then they should call DecodeDatasetWithGobDecoder.

This function returns an error in decoding.

func DecodeDatasetWithGobDecoder 

func DecodeDatasetWithGobDecoder(ds *Dataset, decoder *gob.Decoder) error

DecodeDatasetWithGobDecoder decodes Dataset using decoder.

This function returns an error in decoding.

func DecodeJaccardHashing 

func DecodeJaccardHashing(hashing *JaccardHashing, r io.Reader) error

DecodeJaccardHashing decodes Dataset from r. Directly passing *os.File used by a gob.Decoder to this function causes mysterious errors. Thus, if users use gob.Decoder, then they should call DecodeJaccardHashingWithGobDecoder.

This function returns an error in decoding.

func DecodeJaccardHashingWithGobDecoder 

func DecodeJaccardHashingWithGobDecoder(hashing *JaccardHashing, decoder *gob.Decoder) error

DecodeJaccardHashingWithGobDecoder decodes JaccardHashing using decoder.

This function returns an error in decoding.

func DecodeLabelConst 

func DecodeLabelConst(model *LabelConst, r io.Reader) error

DecodeLabelConst decodes LabelConst from r. Directly passing *os.File used by a gob.Decoder to this function causes mysterious errors. Thus, if users use gob.Decoder, then they should call DecodeLabelConstWithGobDecoder.

This function returns an error in decoding.

func DecodeLabelConstWithGobDecoder 

func DecodeLabelConstWithGobDecoder(model *LabelConst, decoder *gob.Decoder) error

DecodeLabelConstWithGobDecoder decodes LabelConst using decoder.

This function returns an error in decoding.

func DecodeLabelNear 

func DecodeLabelNear(model *LabelNear, r io.Reader) error

DecodeLabelNear decodes LabelNear from r. Directly passing *os.File used by a gob.Decoder to this function causes mysterious errors. Thus, if users use gob.Decoder, then they should call DecodeLabelNearWithGobDecoder.

This function returns an error in decoding.

func DecodeLabelNearWithGobDecoder 

func DecodeLabelNearWithGobDecoder(model *LabelNear, decoder *gob.Decoder) error

DecodeLabelNearWithGobDecoder decodes LabelNear using decoder.

This function returns an error in decoding.

func DecodeLabelNearest 

func DecodeLabelNearest(model *LabelNearest, r io.Reader) error

DecodeLabelNearest decodes LabelNearest from r. Directly passing *os.File used by a gob.Decoder to this function causes mysterious errors. Thus, if users use gob.Decoder, then they should call DecodeLabelNearestWithGobDecoder.

This function returns an error in decoding.

func DecodeLabelNearestWithGobDecoder 

func DecodeLabelNearestWithGobDecoder(model *LabelNearest, decoder *gob.Decoder) error

DecodeLabelNearestWithGobDecoder decodes LabelNearest using decoder.

This function returns an error in decoding.

func DecodeLabelOne 

func DecodeLabelOne(model *LabelOne, r io.Reader) error

DecodeLabelOne decodes LabelOne from r. Directly passing *os.File used by a gob.Decoder to this function causes mysterious errors. Thus, if users use gob.Decoder, then they should call DecodeLabelOneWithGobDecoder.

This function returns an error in decoding.

func DecodeLabelOneWithGobDecoder 

func DecodeLabelOneWithGobDecoder(model *LabelOne, decoder *gob.Decoder) error

DecodeLabelOneWithGobDecoder decodes LabelOne using decoder.

This function returns an error in decoding.

func DotCount 

func DotCount(x, y FeatureVector) (float32, int)

DotCount returns the inner product and the size of the intersect of the supports between x and y.

func EncodeDataset 

func EncodeDataset(ds *Dataset, w io.Writer) error

EncodeDataset encodes Dataset to w. Directly passing *os.File used by a gob.Encoder to this function causes mysterious errors. Thus, if users use gob.Encoder, then they should call EncodeDatasetWithGobEncoder.

This function returns an error in encoding.

func EncodeDatasetWithGobEncoder 

func EncodeDatasetWithGobEncoder(ds *Dataset, encoder *gob.Encoder) error

EncodeDatasetWithGobEncoder decodes Dataset using encoder.

This function returns an error in decoding.

func EncodeJaccardHashing 

func EncodeJaccardHashing(hashing *JaccardHashing, w io.Writer) error

EncodeJaccardHashing encodes JaccardHashing to w. Directly passing *os.File used by a gob.Encoder to this function causes mysterious errors. Thus, if users use gob.Encoder, then they should call EncodeJaccardHashingWithGobEncoder.

This function returns an error in encoding.

func EncodeJaccardHashingWithGobEncoder 

func EncodeJaccardHashingWithGobEncoder(hashing *JaccardHashing, encoder *gob.Encoder) error

EncodeJaccardHashingWithGobEncoder decodes JaccardHashing using encoder.

This function returns an error in decoding.

func EncodeLabelConst 

func EncodeLabelConst(model *LabelConst, w io.Writer) error

EncodeLabelConst encodes LabelConst to w. Directly passing *os.File used by a gob.Encoder to this function causes mysterious errors. Thus, if users use gob.Encoder, then they should call EncodeLabelBoostWithGobEncoder.

This function returns an error in encoding.

func EncodeLabelConstWithGobEncoder 

func EncodeLabelConstWithGobEncoder(model *LabelConst, encoder *gob.Encoder) error

EncodeLabelConstWithGobEncoder decodes LabelConst using encoder.

This function returns an error in decoding.

func EncodeLabelNear 

func EncodeLabelNear(model *LabelNear, w io.Writer) error

EncodeLabelNear encodes LabelNear to w. Directly passing *os.File used by a gob.Encoder to this function causes mysterious errors. Thus, if users use gob.Encoder, then they should call EncodeLabelNearWithGobEncoder.

This function returns an error in encoding.

func EncodeLabelNearWithGobEncoder 

func EncodeLabelNearWithGobEncoder(model *LabelNear, encoder *gob.Encoder) error

EncodeLabelNearWithGobEncoder decodes LabelNear using encoder.

This function returns an error in decoding.

func EncodeLabelNearest 

func EncodeLabelNearest(model *LabelNearest, w io.Writer) error

EncodeLabelNearest encodes LabelNearest to w. Directly passing *os.File used by a gob.Encoder to this function causes mysterious errors. Thus, if users use gob.Encoder, then they should call EncodeLabelNearestWithGobEncoder.

This function returns an error in encoding.

func EncodeLabelNearestWithGobEncoder 

func EncodeLabelNearestWithGobEncoder(model *LabelNearest, encoder *gob.Encoder) error

EncodeLabelNearestWithGobEncoder decodes LabelNearest using encoder.

This function returns an error in decoding.

func EncodeLabelOne 

func EncodeLabelOne(model *LabelOne, w io.Writer) error

EncodeLabelOne encodes LabelOne to w. Directly passing *os.File used by a gob.Encoder to this function causes mysterious errors. Thus, if users use gob.Encoder, then they should call EncodeLabelOneWithGobEncoder.

This function returns an error in encoding.

func EncodeLabelOneWithGobEncoder 

func EncodeLabelOneWithGobEncoder(model *LabelOne, encoder *gob.Encoder) error

EncodeLabelOneWithGobEncoder decodes LabelOne using encoder.

This function returns an error in decoding.

func Exp32 

func Exp32(x float32) float32

Exp32 is the float32-version of math.Exp.

func Floor32 

func Floor32(x float32) float32

Floor32 is the float32-version of math.Floor.

func HashUint32 

func HashUint32(x uint32) uint32

HashUint32 returns the hashed value of the given uint32 x. This comes from MurmurHash3 fmix32 (https://github.com/aappleby/smhasher/blob/master/src/MurmurHash3.cpp).

func IdealDCG 

func IdealDCG(K uint) float32

IdealDCG returns the calculated ideal DCG. Ideal DCG@K is defined as \sum_{k=1}^K 1/log_2(1+k), which is the maximum of possible DCG@K values. These value are cached persistently.

Ideal DCG@0 is undefined, so this function returns NaN.

func Inf32 

func Inf32(sign int) float32

Inf32 is the float32-version of math.Inf.

func InvertRanks 

func InvertRanks(labelRanks LabelVector) map[uint32]int

InvertRanks returns the inverted ranking list.

func IsInf32 

func IsInf32(f float32, sign int) bool

IsInf32 is the float32-version of math.Inf.

func IsNaN32 

func IsNaN32(f float32) (is bool)

IsNaN32 is the float32-version of math.IsNaN.

func Log32 

func Log32(x float32) float32

Log32 is the float32-version of math.Log.

func LogBinary32 

func LogBinary32(x float32) float32

LogBinary32 is the float32-version of math.Log2.

func Modf32 

func Modf32(x float32) (i, f float32)

Modf32 is the float32-version of math.Modf.

func NaN32 

func NaN32() float32

NaN32 is the float32-version of math.NaN.

func Pow32 

func Pow32(x, y float32) float32

Pow32 is the float32-version of math.Pow.

func ReportMaxPrecision 

func ReportMaxPrecision(Y LabelVectors, K uint) []float32

ReportMaxPrecision reports the maximum Precision@K value of each label vector in Y.

func ReportNDCG 

func ReportNDCG(Y LabelVectors, K uint, Yhat LabelVectors) []float32

ReportNDCG reports the nDCG@K (normalized DCG@K) value of each label vector in Y.

nDCG@0 is undefined, so this function returns a slice filled with NaN.

NOTICE: The maximum nDCG@K is always 1.0, because nDCG@K is normalized.

func ReportPrecision 

func ReportPrecision(Y LabelVectors, K uint, Yhat LabelVectors) []float32

ReportPrecision reports the Precision@K value of each label vector in Y.

func Sqrt32 

func Sqrt32(x float32) float32

Sqrt32 is the float32-version of math.Sqrt32.

func SummarizeFloat32Slice 

func SummarizeFloat32Slice(x []float32) (min, q25, med, q75, max, avg float32)

SummarizeFloat32Slice returns the 5-summary(minimum, 1st quantile, median, 3rd quantile and maximum) and the average of the given float32 slice.

Types

type BinaryClassifier 

type BinaryClassifier struct {
	// Bias is the bias parameter.
	Bias float32
	// Weight is the weight parameter.
	Weight SparseVector
	// The following members are not required.
	//
	// Beta is used by some solvers (using dual problems) as the optimization target.
	// Weight can be expressed as the sum of y_ix_i weighted with the corresponding elements of Beta.
	Beta []float32
}

BinaryClassifier is the data structure having information about binary classifiers.

BinaryClassifier classifies the given entry x as positive if dot(Weight, x) + Bias > 0, otherwise as negative.

func BinaryClassifierTrainer_L1Logistic_PrimalSGD 

func BinaryClassifierTrainer_L1Logistic_PrimalSGD(X FeatureVectors, Y []bool, C, epsilon float32, debug *log.Logger) (*BinaryClassifier, error)

BinaryClassifierTrainer_L1Logistic_PrimalSGD returns an trained BinaryClassifier with FTRL-Proximal (McMahan+ 2013) method for L1-penalized logistic regression. This can be used for estimating the probability which the given data point belongs to the positive class, and this algorithm would produce the smaller model.

This function returns no error currently.

References:

(McMahan+ 2013) H. B. McMahan, et al. "Ad Click Prediction: a View from the Trenches." Proceedings of the 19th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 2013.

func BinaryClassifierTrainer_L1SVC_PrimalSGD 

func BinaryClassifierTrainer_L1SVC_PrimalSGD(X FeatureVectors, Y []bool, C, epsilon float32, debug *log.Logger) (*BinaryClassifier, error)

BinaryClassifierTrainer_L1SVC_PrimalSGD trains a L1-Support Vector Classifier with primal stochastic gradient descent. This is registered to BinaryClassifierTrainers.

The used update procedure is the one used by Online Passive-Aggressive Algorithm (Crammer+ 2006) with the dynamic penalty parameter depending on the round number t. This update is proven to be safe, that is, this leads to sane results even when the learning rate is large (Karampatziakis+ 2011, SubSection 4.2). Thus, although we fix the eta0 as 1.0 and the learning rate as eta0 / t, this algorithm is enough fast and accurate.

This function returns no error currently.

Reference:

(Crammer+ 2006) K.Crammer, O. Dekel, J. Keshet, S. Shalev-Shwarts, and Y. Singer. "Online Passive-Aggressive Algorithms." Journal of Machine Learning Research, vol. 7, pp. 551-585, 2006.

(Karampatziakis+ 2011) N. Karampatziakis, and J. Langford, "Online Importance Weight Aware Updates." Association for Uncertainty in Artificial Intelligence, 2011.

func (*BinaryClassifier) Predict 

func (bc *BinaryClassifier) Predict(x FeatureVector) float32

Predict returns the predicted value dot(Weight, x) + Bias.

func (*BinaryClassifier) PredictAll 

func (bc *BinaryClassifier) PredictAll(X FeatureVectors) []float32

PredictAll returns the predicted values dot(Weight, x) + Bias for each feature vector.

func (*BinaryClassifier) PredictAndCount 

func (bc *BinaryClassifier) PredictAndCount(x FeatureVector) (float32, uint32)

PredictAndCount returns the predicted value dot(Weight, x) + Bias and the splitter count (the number of times the splitter hits).

func (*BinaryClassifier) PredictAndCountAll 

func (bc *BinaryClassifier) PredictAndCountAll(X FeatureVectors) ([]float32, []uint32)

PredictAndCountAll returns the predicted values dot(Weight, x) + Bias and the splitter count for each feature vector.

func (*BinaryClassifier) ReportPerformance 

func (bc *BinaryClassifier) ReportPerformance(X FeatureVectors, Y []bool) (tn, fn, fp, tp uint, predVals []float32, Yhat []bool)

ReportPerformance returns the true-negative/false-negative/false-positive/true-positive and the predicted values on X.

type BinaryClassifierTrainer 

type BinaryClassifierTrainer func(X FeatureVectors, Y []bool, C, epsilon float32, debug *log.Logger) (*BinaryClassifier, error)

BinaryClassifierTrainer is the type of binary classifier trainers. A trainer returns a new BinaryClassifier on X and Y. C is the inverse of the penalty parameter. epsilon is the tolerance parameter for checking the convergence. debug is used for debug logs.

type Dataset 

type Dataset struct {
	X FeatureVectors
	Y LabelVectors
}

Dataset is a collection of the pair of one feature vector and one label vector.

func ReadTextDataset 

func ReadTextDataset(reader io.Reader) (*Dataset, error)

ReadTextDataset returns a new Dataset from reader.

The data from reader is assumed to be formatted like used in LIBLINEAR and LIBSVM. It is a plain text whose cells are separated by single space. The first line has the number of entries, features, and labels. The remaining lines have entries of the dataset. Each entry is encoded in one line like (comma-separated label) (feature:value)*.

This function returns an error in reading the dataset.

func (*Dataset) FeatureSubSet 

func (ds *Dataset) FeatureSubSet(features map[uint32]struct{}) *Dataset

FeatureSubSet returns the sub-set of the dataset whose entry has only features in the given set of features. For efficiency, the label vectors of sub-dataset has references to the one of the dataset.

func (*Dataset) GobEncode 

func (ds *Dataset) GobEncode() ([]byte, error)

GobEncode returns the error always, because users should encode large Dataset objects with EncodeDataset.

func (*Dataset) Size 

func (ds *Dataset) Size() int

Size returns the size of the dataset.

func (*Dataset) SubSet 

func (ds *Dataset) SubSet(indices []int) *Dataset

SubSet returns the sub-set of the dataset selected by the given index slice. For efficiency, the all vectors of sub-dataset has references to the ones of the dataset.

This function panics if the index is out of range.

func (*Dataset) WriteTextDataset 

func (ds *Dataset) WriteTextDataset(w io.Writer) error

WriteTextDataset writes the dataset to the given writer as ReadTextDataset supports.

This function returns an error in writing.

type FeatureVector 

type FeatureVector = KeyValues32OrderedByKey

FeatureVector is the sparse and static feature vector. The elements should be ordered by feature ID (key).

type FeatureVectors 

type FeatureVectors []FeatureVector

FeatureVectors is the FeatureVector slice.

func (FeatureVectors) Dim 

func (X FeatureVectors) Dim() (d int)

Dim returns the calculated dimension of FeatureVectors. This is the maximum feature key plus 1.

type Float32Slice 

type Float32Slice []float32

Float32Slice implements the interface sort.Interface.

func (Float32Slice) Len 

func (x Float32Slice) Len() int

func (Float32Slice) Less 

func (x Float32Slice) Less(i, j int) bool

func (Float32Slice) Swap 

func (x Float32Slice) Swap(i, j int)

type JaccardHashing 

type JaccardHashing struct {
	// contains filtered or unexported fields
}

JaccardHashing is the optimal Densified One Permutation Hashing (DOPH) for estimating Jaccard similarity (Wang+ 2017).

References:

(Wang+ 2017) Y. Wang, A. Shrivastava, and J. Ryu. "FLASH: Randomized Algorithms Accelerated over CPU-GPU for Ultra-High Dimensional Similarity Search." arXiv preprint arXiv:1709.01190, 2017.

func NewJaccardHashing 

func NewJaccardHashing(K, L, R uint) *JaccardHashing

NewJaccardHashing returns an new JaccardHashing. Recommended K, L and R are 64, 16 and 64, respectively.

func (*JaccardHashing) Add 

func (hashing *JaccardHashing) Add(vec FeatureVector, i uint32)

Add adds the given feature vector as i-th index to the hash tables.

func (*JaccardHashing) FindNears 

func (hashing *JaccardHashing) FindNears(vec FeatureVector) KeyCountMap32

FindNears returns the histogram of the neighbors from the given feature vector.

func (*JaccardHashing) GobEncode 

func (hashing *JaccardHashing) GobEncode() ([]byte, error)

GobEncode returns the error always, because users should encode large JaccardHashing objects with EncodeJaccardHashing.

func (*JaccardHashing) Hash 

func (hashing *JaccardHashing) Hash(vec FeatureVector) []uint32

Hash returns the K hashed values of the given feature vector.

func (*JaccardHashing) K 

func (hashing *JaccardHashing) K() uint

K returns the number of the hash tables.

func (*JaccardHashing) L 

func (hashing *JaccardHashing) L() uint

L returns the bit-width for backet indices in each hash table.

func (*JaccardHashing) R 

func (hashing *JaccardHashing) R() uint

R returns the size of a reservoir of each backet.

func (*JaccardHashing) ResetRng 

func (hashing *JaccardHashing) ResetRng()

ResetRng resets the internal random number generator.

func (*JaccardHashing) Summary 

func (hashing *JaccardHashing) Summary() (backetUsage []int, backetHist []int)

Summary returns the slice of bucket usage and the bucket size (size of each reservoir) histogram.

type KeyCount32 

type KeyCount32 struct {
	Key, Count uint32
}

KeyCount32 is the pair of uint32 feature key and its uint32 value.

type KeyCountMap32 

type KeyCountMap32 KeyCounts32

KeyCountMap32 is the faster map of KeyCount32s. This cannot has entries with value 0.

Currently, this does not support expansion at insertion. Users can iterate the entries with raw access to the internal, so this is for expert use in order to achieve faster counting.

func NewKeyCountMap32 

func NewKeyCountMap32(capacity uint) KeyCountMap32

NewKeyCountMap32 returns a new KeyCountMap32.

func (KeyCountMap32) Get 

func (m KeyCountMap32) Get(key uint32) KeyCount32

Get returns the entry with the given key.

func (KeyCountMap32) Inc 

func (m KeyCountMap32) Inc(key uint32) KeyCount32

Inc increments the entry's value with the given key, and returns the entry.

func (KeyCountMap32) Map 

func (m KeyCountMap32) Map() map[uint32]uint32

Map returns the map version of self.

type KeyCounts32 

type KeyCounts32 []KeyCount32

KeyCounts32 is the slice of KeyCount32.

func (KeyCounts32) ExtractLargestCountsByInsert 

func (kcs KeyCounts32) ExtractLargestCountsByInsert(K uint) KeyCounts32

ExtractLargestCountsByInsert returns the only K largest entries.

func (KeyCounts32) SortLargestCountsWithHeap 

func (kcs KeyCounts32) SortLargestCountsWithHeap(K uint) KeyCounts32

SortLargestCountsWithHeap sorts the only K largest entries at the first as maintaining the heap, and returns the shrunk slice to the self.

type KeyValue32 

type KeyValue32 struct {
	Key   uint32
	Value float32
}

KeyValue32 is the pair of uint32 feature key and its float32 value.

type KeyValues32 

type KeyValues32 []KeyValue32

KeyValues32 is the KeyValue32 slice.

type KeyValues32OrderedByKey 

type KeyValues32OrderedByKey KeyValues32

KeyValues32OrderedByKey is KeyValues32 implementing sort.Interface for sorting in the key order. If the keys are same, then these key-values are sorted in increasing value order.

func (KeyValues32OrderedByKey) Len 

func (kvs KeyValues32OrderedByKey) Len() int

func (KeyValues32OrderedByKey) Less 

func (kvs KeyValues32OrderedByKey) Less(i, j int) bool

func (KeyValues32OrderedByKey) Swap 

func (kvs KeyValues32OrderedByKey) Swap(i, j int)

type KeyValues32OrderedByValue 

type KeyValues32OrderedByValue KeyValues32

KeyValues32OrderedByValue is KeyValues32 implementing sort.Interface for sorting in the value order. If the values are same, then these key-values are sorted in decreasing key order, because this is intended for sorting in decreasing key/value order in reverse mode.

func (KeyValues32OrderedByValue) Len 

func (kvs KeyValues32OrderedByValue) Len() int

func (KeyValues32OrderedByValue) Less 

func (kvs KeyValues32OrderedByValue) Less(i, j int) bool

func (KeyValues32OrderedByValue) Swap 

func (kvs KeyValues32OrderedByValue) Swap(i, j int)

type LabelConst 

type LabelConst struct {
	// LabelList and LabelFreqList are the label and its frequency list in descending order in the training set occurrences.
	LabelList     LabelVector
	LabelFreqList []float32
}

LabelConst is the multi-label constant model.

func TrainLabelConst 

func TrainLabelConst(ds *Dataset, debug *log.Logger) (*LabelConst, error)

TrainLabelConst returns an trained LabelConst on the given dataset ds.

func (*LabelConst) GobEncode 

func (model *LabelConst) GobEncode() ([]byte, error)

GobEncode returns the error always, because users should encode large LabelConst objects with EncodeLabelConst.

func (*LabelConst) PredictAll 

func (model *LabelConst) PredictAll(X FeatureVectors, K uint) LabelVectors

PredictAll returns the top-K labels for each data entry in X.

type LabelNear 

type LabelNear struct {
	// Dataset is the training dataset.
	Dataset *Dataset
	// Hashing is the Jaccard hashing.
	Hashing *JaccardHashing
}

LabelNear is a faster implementation of LabelNearest which uses the optimal Densified One Permutation Hashing (DOPH) and the reservoir sampling (Wang+ 2017).

References:

(Wang+ 2017) Y. Wang, A. Shrivastava, and J. Ryu. "FLASH: Randomized Algorithms Accelerated over CPU-GPU for Ultra-High Dimensional Similarity Search." arXiv preprint arXiv:1709.01190, 2017.

func TrainLabelNear 

func TrainLabelNear(ds *Dataset, params *LabelNearParameters, debug *log.Logger) (*LabelNear, error)

TrainLabelNear returns an trained LabelNear on the given training dataset ds.

Currently, this function returns no error.

func (*LabelNear) FindNears 

func (model *LabelNear) FindNears(x FeatureVector, c, S uint, beta float32) KeyValues32

FindNears returns the S near entries with each similarity for the given entry. The quantity c*S is used for sieving the candidates by hashing. See Predict for hyper-parameter details.

func (*LabelNear) GobEncode 

func (model *LabelNear) GobEncode() ([]byte, error)

GobEncode returns the error always, because users should encode large LabelNear objects with EncodeLabelNear.

func (*LabelNear) Predict 

func (model *LabelNear) Predict(x FeatureVector, K, c, S uint, alpha, beta float32) (LabelVector, map[uint32]float32, KeyValues32)

Predict returns the results for the given data entry x with the sparse S-near neighborhood. The returned results are the top-K labels, the label histogram, and the slice of the data entry index and its similarity.

alpha is the smoothing parameter for weighting the votes by each neighbor. beta is the smoothing parameter for balancing the Jaccard similarity and the cosine similarity.

func (*LabelNear) PredictAll 

func (model *LabelNear) PredictAll(X FeatureVectors, K, c, S uint, alpha, beta float32) LabelVectors

PredictAll returns the top-K labels for each data entry in X with the sparse S-near neighborhood. See Predict for hyper-parameter details.

type LabelNearParameters 

type LabelNearParameters struct {
	// K is the number of the hash tables.
	K uint
	// L is the bit-width of bucket indices in each hash table.
	L uint
	// R is the size of a reservoir of each bucket.
	R uint
}

LabelNearParameters is the parameters for LabelNear.

func NewLabelNearParameters 

func NewLabelNearParameters() *LabelNearParameters

NewLabelNearParameters returns an new default LabelNearParameters.

type LabelNearest 

type LabelNearest struct {
	// NfeaturesList is the slice of the number of features contained in each training data entry.
	NfeaturesList []uint32
	// FeatureIndexList is the map from the feature to the feature index.
	// The feature index contains the list of the pair of the entry index and the corresponding feature value.
	// The feature vector for the each entry is normalized for computing cos similarity effectively in the inference.
	FeatureIndexList map[uint32]KeyValues32
	// LabelVectors is the label vectors in the training dataset.
	LabelVectors LabelVectors
}

LabelNearest is the sparse weighted nearest neighborhood. Sparse means in the following 3 reasons: 

(i) The similarity used in constructing the nearest neighborhood defined by the inner-product on the features activated in the given entry.
(ii) The positive labels assigned to the entries in the training dataset are only used for averaging.
(iii) The entries in the training dataset whose inner-product is not positive are not used.

LabelNearest is only the optimized data structure of the training dataset for searching nearest neighborhood.

func TrainLabelNearest 

func TrainLabelNearest(ds *Dataset, debug *log.Logger) (*LabelNearest, error)

TrainLabelNearest returns an trained LabelNearest on the given training dataset ds.

Currently, this function returns no error.

func (*LabelNearest) FindNearests 

func (model *LabelNearest) FindNearests(x FeatureVector, S uint, beta float32) KeyValues32

FindNearests returns the S nearest entries with each similarity for the given entry. See Predict for hyper-parameter details.

func (*LabelNearest) FindNearestsWithContext 

func (model *LabelNearest) FindNearestsWithContext(x FeatureVector, S uint, beta float32, ctx LabelNearestContext) KeyValues32

FindNearestsWithContext is FindNearests with the specified LabelNearestContext.

func (*LabelNearest) GobEncode 

func (model *LabelNearest) GobEncode() ([]byte, error)

GobEncode returns the error always, because users should encode large LabelNearest objects with EncodeLabelNearest.

func (*LabelNearest) NewContext 

func (model *LabelNearest) NewContext() LabelNearestContext

NewContext returns a new context for some inference memory.

func (*LabelNearest) Predict 

func (model *LabelNearest) Predict(x FeatureVector, K, S uint, alpha, beta float32) (LabelVector, map[uint32]float32, KeyValues32)

Predict returns the results for the given data entry x with the sparse S-nearest neighborhood. The returned results are the top-K labels, the label histogram, and the slice of the data entry index and its similarity.

alpha is the smoothing parameter for weighting the votes by each neighbor. beta is the smoothing parameter for balancing the Jaccard similarity and the cosine similarity.

func (*LabelNearest) PredictAll 

func (model *LabelNearest) PredictAll(X FeatureVectors, K, S uint, alpha, beta float32) LabelVectors

PredictAll returns the top-K labels for each data entry in X with the sparse S-nearest neighborhood. See Predict for hyper-parameter details.

func (*LabelNearest) PredictWithContext 

func (model *LabelNearest) PredictWithContext(x FeatureVector, K, S uint, alpha, beta float32, ctx LabelNearestContext) (LabelVector, map[uint32]float32, KeyValues32)

PredictWithContext is Predict with the specified LabelNearestContext.

type LabelNearestContext 

type LabelNearestContext []SimCountPair

LabelNearestContext is a context used in inference. This is not protected by any mutex, so this should not be accessed by multiple goroutines.

type LabelOne 

type LabelOne struct {
	// Params is the used LabelOneParameters.
	Params *LabelOneParameters
	// Biases is the bias slice used by splitters on each classifier.
	Biases []float32
	// Weights is the weight sparse matrix used by each classifier.
	// This is the map from the feature key to the (roundID, the weight on the feature of #roundID splitter) slice.
	// This data structure reduces the number of times that the classifier accesses the golang's map a lot.
	WeightLists map[uint32]KeyValues32
	// Labels is the label slice used in each classifier.
	// The t-th label is the target label of the t-th classifier.
	Labels LabelVector
	// The following members are not required.
	//
	// Summaries is the summary object slice for each boosting round.
	// The entries in this summary is considered to provide compact and useful information in best-effort, so this specification would be loose and rapidly changing.
	Summaries []map[string]interface{}
}

LabelOne is the One-versus-Rest classifier for multi-label ranking. The t-th classifier (t = 1, ..., T) is the classifier for the top-t frequently occurring label.

func TrainLabelOne 

func TrainLabelOne(ds *Dataset, params *LabelOneParameters, debug *log.Logger) (*LabelOne, error)

TrainLabelOne returns an trained LabelOne on the given dataset ds.

func (*LabelOne) GobEncode 

func (model *LabelOne) GobEncode() ([]byte, error)

GobEncode returns the error always, because users should encode large LabelOne objects with EncodeLabelOne.

func (*LabelOne) Nrounds 

func (model *LabelOne) Nrounds() uint

Nrounds return the number of the rounds.

func (*LabelOne) Predict 

func (model *LabelOne) Predict(x FeatureVector, K uint, T uint) LabelVector

Predict returns the top-K predicted labels for the given data entry x with the first T rounds.

func (*LabelOne) PredictAll 

func (model *LabelOne) PredictAll(X FeatureVectors, K uint, T uint) LabelVectors

PredictAll returns the slice of the top-K predicted labels for each data entry in X with the first T rounds.

func (*LabelOne) Prune 

func (model *LabelOne) Prune(T uint) *LabelOne

Prune returns the pruned LabelOne which has at most T rounds.

type LabelOneParameters 

type LabelOneParameters struct {
	// ClassifierTrainerName is the used BinaryClassifierTrainer name.
	ClassifierTrainerName string
	// C is the penalty parameter for BinaryClassifierTrainer.
	C float32
	// Epsilon is the tolerance parameter for BinaryClassifierTrainer.
	Epsilon float32
	// T is the maximum number of the rounds, which is equal to the maximum number of the target labels.
	T uint
}

LabelOneParameters is the parameters for LabelOne.

func NewLabelOneParameters 

func NewLabelOneParameters() *LabelOneParameters

NewLabelOneParameters returns an LabelOneParameters initialized with the default values.

type LabelVector 

type LabelVector []uint32

LabelVector is the sparse label vector which is the slice of label key.

func RankTopK 

func RankTopK(labelDist SparseVector, K uint) LabelVector

RankTopK returns the top-K labels.

func (LabelVector) Len 

func (labels LabelVector) Len() int

func (LabelVector) Less 

func (labels LabelVector) Less(i, j int) bool

func (LabelVector) Swap 

func (labels LabelVector) Swap(i, j int)

type LabelVectors 

type LabelVectors []LabelVector

LabelVectors is the LabelVector slice.

func (LabelVectors) Dim 

func (Y LabelVectors) Dim() (d int)

Dim returns the calculated dimension of label vectors. This is the maximum label ID plus 1.

type SimCountPair 

type SimCountPair struct {
	Sim   float32
	Count uint32
}

SimCountPair is the data structure for float32 similarity and uint32 count.

type SparseVector 

type SparseVector map[uint32]float32

SparseVector is the map from uint32 key to float32 value.

func SparsifyVector 

func SparsifyVector(v []float32) SparseVector

SparsifyVector returns a SparseVector converted from v.

type SparseVectors 

type SparseVectors []SparseVector

SparseVectors is the slice of SparseVector.

Source Files

View all Source files

Directories

Path Synopsis
Package plugin provides plugin functions for sticker.
 Package plugin provides plugin functions for sticker.
next
common

Jump to

Keyboard shortcuts

? : This menu
/ : Search site
f or F : Jump to
y or Y : Canonical URL
go.dev uses cookies from Google to deliver and enhance the quality of its services and to analyze traffic. Learn more.