Sampler< Random > Class Template Reference

Create random samples from a sequence of indices. More...

#include <vigra/sampling.hxx>

Public Types
typedef ArrayVectorView < IndexType >	IndexArrayViewType
typedef Int32	IndexType

Public Member Functions
IndexArrayViewType	oobIndices () const
IndexType	operator[] (int k) const
void	sample ()
IndexArrayViewType	sampledIndices () const
	Sampler (UInt32 totalCount, SamplerOptions const &opt=SamplerOptions(), Random const *rnd=0)
template<class Iterator >
	Sampler (Iterator strataBegin, Iterator strataEnd, SamplerOptions const &opt=SamplerOptions(), Random const *rnd=0)
int	sampleSize () const
int	size () const
int	strataCount () const
bool	stratifiedSampling () const
int	totalCount () const
bool	withReplacement () const

Detailed Description

template<class Random = MersenneTwister>
class vigra::Sampler< Random >

Create random samples from a sequence of indices.

Selecting data items at random is a basic task of machine learning, for example in boostrapping, RandomForest training, and cross validation. This class implements various ways to select random samples via their indices. Indices are assumed to be consecutive in the range 0 <= index < total_sample_count.

The class always contains a current sample which can be accessed by the index operator or by the function sampledIndices(). The indices that are not in the current sample (out-of-bag indices) can be accessed via the function oobIndices().

The sampling method (with/without replacement, stratified or not) and the number of samples to draw are determined by the option object SamplerOptions.

Usage:

#include <vigra/sampling.hxx>
Namespace: vigra

Create a Sampler with default options, i.e. sample as many indices as there are data elements, with replacement. On average, the sample will contain 0.63*totalCount distinct indices.

int totalCount = 10000;   // total number of data elements
int numberOfSamples = 20; // repeat experiment 20 times
Sampler<> sampler(totalCount);
for(int k=0; k<numberOfSamples; ++k)
{
    // process current sample
    for(int i=0; i<sampler.sampleSize(); ++i)
    {
        int currentIndex = sampler[i];
        processData(data[currentIndex]);
    }
    // create next sample
    sampler.sample();
}

Create a Sampler for stratified sampling, without replacement.

// prepare the strata (i.e. specify which stratum each element belongs to)
int stratumSize1 = 2000, stratumSize2 = 8000,
    totalCount = stratumSize1 + stratumSize2;
ArrayVerctor<int> strata(totalCount);
for(int i=0; i<stratumSize1; ++i)
    strata[i] = 1;
for(int i=stratumSize1; i<stratumSize2; ++i)
    strata[i] = 2;

int sampleSize = 200; // i.e. sample 100 elements from each of the two strata
int numberOfSamples = 20; // repeat experiment 20 times
Sampler<> stratifiedSampler(strata.begin(), strata.end(),
                 SamplerOptions().withoutReplacement().stratified().sampleSize(sampleSize));
// create first sample
sampler.sample();

for(int k=0; k<numberOfSamples; ++k)
{
    // process current sample
    for(int i=0; i<sampler.sampleSize(); ++i)
    {
        int currentIndex = sampler[i];
        processData(data[currentIndex]);
    }
    // create next sample
    sampler.sample();
}

Member Typedef Documentation

typedef Int32 IndexType

Internal type of the indices. Currently, 64-bit indices are not supported because this requires extension of the random number generator classes.

typedef ArrayVectorView<IndexType> IndexArrayViewType

Type of the array view object that is returned by sampledIndices() and oobIndices().

Constructor & Destructor Documentation

Sampler	(	UInt32	totalCount,
		SamplerOptions const &	opt = SamplerOptions(),
		Random const *	rnd = 0
	)

Create a sampler for totalCount data objects.

In each invocation of sample() below, it will sample indices according to the options passed. If no options are given, totalCount indices will be drawn with replacement.

Sampler	(	Iterator	strataBegin,
		Iterator	strataEnd,
		SamplerOptions const &	opt = SamplerOptions(),
		Random const *	rnd = 0
	)

Create a sampler for stratified sampling.

strataBegin and strataEnd must refer to a sequence which specifies for each sample the stratum it belongs to. The total number of data objects will be set to strataEnd - strataBegin. Equally many samples (subject to rounding) will be drawn from each stratum, unless the option object explicitly requests unstratified sampling, in which case the strata are ignored.

Member Function Documentation

IndexType operator[]

(

int

k

)

const

Return the k-th index in the current sample.

void sample

(

)

Create a new sample.

int totalCount

(

)

const

The total number of data elements.

int sampleSize

(

)

const

The number of data elements that have been sampled.

int size

(

)

const

Same as sampleSize().

int strataCount

(

)

const

The number of strata to be used. Will be 1 if no strata are given. Will be ignored when stratifiedSampling() is false.

bool stratifiedSampling

(

)

const

Whether to use stratified sampling. (If this is false, strata will be ignored even if present.)

bool withReplacement

(

)

const

Whether sampling should be performed with replacement.

IndexArrayViewType sampledIndices

(

)

const

Return an array view containing the indices in the current sample.

IndexArrayViewType oobIndices

(

)

const

Return an array view containing the out-of-bag indices. (i.e. the indices that are not in the current sample)

---

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

• vigra/sampling.hxx