Public Member Functions | Static Public Member Functions | Protected Member Functions | Protected Attributes

CSparseFeatures< ST > Class Template Reference


Detailed Description

template<class ST>
class shogun::CSparseFeatures< ST >

Template class SparseFeatures implements sparse matrices.

Features are an array of SGSparseVector, sorted w.r.t. vec_index (increasing) and withing same vec_index w.r.t. feat_index (increasing);

Sparse feature vectors can be accessed via get_sparse_feature_vector() and should be freed (this operation is a NOP in most cases) via free_sparse_feature_vector().

As this is a template class it can directly be used for different data types like sparse matrices of real valued, integer, byte etc type.

(Partly) subset access is supported for this feature type. Simple use the (inherited) set_subset(), remove_subset() functions. If done, all calls that work with features are translated to the subset. See comments to find out whether it is supported for that method

Definition at line 61 of file SparseFeatures.h.

Inheritance diagram for CSparseFeatures< ST >:
Inheritance graph
[legend]

List of all members.

Public Member Functions

 CSparseFeatures (int32_t size=0)
 CSparseFeatures (SGSparseVector< ST > *src, int32_t num_feat, int32_t num_vec, bool copy=false)
 CSparseFeatures (SGSparseMatrix< ST > sparse)
 CSparseFeatures (SGMatrix< ST > dense)
 CSparseFeatures (const CSparseFeatures &orig)
 CSparseFeatures (CFile *loader)
virtual ~CSparseFeatures ()
void free_sparse_feature_matrix ()
void free_sparse_features ()
virtual CFeaturesduplicate () const
ST get_feature (int32_t num, int32_t index)
ST * get_full_feature_vector (int32_t num, int32_t &len)
SGVector< ST > get_full_feature_vector (int32_t num)
virtual int32_t get_nnz_features_for_vector (int32_t num)
SGSparseVector< ST > get_sparse_feature_vector (int32_t num)
ST dense_dot (ST alpha, int32_t num, ST *vec, int32_t dim, ST b)
void add_to_dense_vec (float64_t alpha, int32_t num, float64_t *vec, int32_t dim, bool abs_val=false)
void free_sparse_feature_vector (SGSparseVector< ST > vec, int32_t num)
SGSparseVector< ST > * get_sparse_feature_matrix (int32_t &num_feat, int32_t &num_vec)
SGSparseMatrix< ST > get_sparse_feature_matrix ()
CSparseFeatures< ST > * get_transposed ()
SGSparseVector< ST > * get_transposed (int32_t &num_feat, int32_t &num_vec)
void set_sparse_feature_matrix (SGSparseMatrix< ST > sm)
SGMatrix< ST > get_full_feature_matrix ()
virtual bool set_full_feature_matrix (SGMatrix< ST > full)
virtual bool apply_preprocessor (bool force_preprocessing=false)
virtual int32_t get_size ()
bool obtain_from_simple (CSimpleFeatures< ST > *sf)
virtual int32_t get_num_vectors () const
int32_t get_num_features ()
int32_t set_num_features (int32_t num)
virtual EFeatureClass get_feature_class ()
virtual EFeatureType get_feature_type ()
void free_feature_vector (SGSparseVector< ST > vec, int32_t num)
int64_t get_num_nonzero_entries ()
float64_tcompute_squared (float64_t *sq)
float64_t compute_squared_norm (CSparseFeatures< float64_t > *lhs, float64_t *sq_lhs, int32_t idx_a, CSparseFeatures< float64_t > *rhs, float64_t *sq_rhs, int32_t idx_b)
void load (CFile *loader)
void save (CFile *writer)
CLabelsload_svmlight_file (char *fname, bool do_sort_features=true)
void sort_features ()
bool write_svmlight_file (char *fname, CLabels *label)
virtual int32_t get_dim_feature_space () const
virtual float64_t dot (int32_t vec_idx1, CDotFeatures *df, int32_t vec_idx2)
virtual float64_t dense_dot (int32_t vec_idx1, const float64_t *vec2, int32_t vec2_len)
virtual void * get_feature_iterator (int32_t vector_index)
virtual bool get_next_feature (int32_t &index, float64_t &value, void *iterator)
virtual void free_feature_iterator (void *iterator)
virtual CFeaturescopy_subset (SGVector< index_t > indices)
virtual const char * get_name () const

Static Public Member Functions

static ST sparse_dot (ST alpha, SGSparseVectorEntry< ST > *avec, int32_t alen, SGSparseVectorEntry< ST > *bvec, int32_t blen)
static void clean_tsparse (SGSparseVector< ST > *sfm, int32_t num_vec)

Protected Member Functions

virtual SGSparseVectorEntry< ST > * compute_sparse_feature_vector (int32_t num, int32_t &len, SGSparseVectorEntry< ST > *target=NULL)

Protected Attributes

int32_t num_vectors
 total number of vectors
int32_t num_features
 total number of features
SGSparseVector< ST > * sparse_feature_matrix
 array of sparse vectors of size num_vectors
CCache< SGSparseVectorEntry
< ST > > * 
feature_cache

Constructor & Destructor Documentation

CSparseFeatures ( int32_t  size = 0  ) 

constructor

Parameters:
size cache size

Definition at line 68 of file SparseFeatures.h.

CSparseFeatures ( SGSparseVector< ST > *  src,
int32_t  num_feat,
int32_t  num_vec,
bool  copy = false 
)

convenience constructor that creates sparse features from the ones passed as argument

Parameters:
src dense feature matrix
num_feat number of features
num_vec number of vectors
copy true to copy feature matrix

Definition at line 81 of file SparseFeatures.h.

CSparseFeatures ( SGSparseMatrix< ST >  sparse  ) 

convenience constructor that creates sparse features from sparse features

Parameters:
sparse sparse matrix

Definition at line 107 of file SparseFeatures.h.

CSparseFeatures ( SGMatrix< ST >  dense  ) 

convenience constructor that creates sparse features from dense features

Parameters:
dense dense feature matrix

Definition at line 121 of file SparseFeatures.h.

CSparseFeatures ( const CSparseFeatures< ST > &  orig  ) 

copy constructor

Definition at line 131 of file SparseFeatures.h.

CSparseFeatures ( CFile loader  ) 

constructor loading features from file

Parameters:
loader File object to load data from

Definition at line 159 of file SparseFeatures.h.

virtual ~CSparseFeatures (  )  [virtual]

default destructor

Definition at line 169 of file SparseFeatures.h.


Member Function Documentation

void add_to_dense_vec ( float64_t  alpha,
int32_t  num,
float64_t vec,
int32_t  dim,
bool  abs_val = false 
) [virtual]

add a sparse feature vector onto a dense one dense+=alpha*sparse

possible with subset

Parameters:
alpha scalar to multiply with
num index of feature vector
vec dense vector
dim length of the dense vector
abs_val if true, do dense+=alpha*abs(sparse)

Implements CDotFeatures.

Definition at line 495 of file SparseFeatures.h.

virtual bool apply_preprocessor ( bool  force_preprocessing = false  )  [virtual]

apply preprocessor

possible with subset

Parameters:
force_preprocessing if preprocssing shall be forced
Returns:
if applying was successful

Definition at line 834 of file SparseFeatures.h.

static void clean_tsparse ( SGSparseVector< ST > *  sfm,
int32_t  num_vec 
) [static]

clean SGSparseVector

Parameters:
sfm sparse feature matrix
num_vec number of vectors in matrix

Definition at line 586 of file SparseFeatures.h.

virtual SGSparseVectorEntry<ST>* compute_sparse_feature_vector ( int32_t  num,
int32_t &  len,
SGSparseVectorEntry< ST > *  target = NULL 
) [protected, virtual]

compute feature vector for sample num if target is set the vector is written to target len is returned by reference

NOT IMPLEMENTED!

Parameters:
num num
len len
target target

Definition at line 1543 of file SparseFeatures.h.

float64_t* compute_squared ( float64_t sq  ) 

compute a^2 on all feature vectors

possible with subset

Parameters:
sq the square for each vector is stored in here
Returns:
the square for each vector

Definition at line 963 of file SparseFeatures.h.

float64_t compute_squared_norm ( CSparseFeatures< float64_t > *  lhs,
float64_t sq_lhs,
int32_t  idx_a,
CSparseFeatures< float64_t > *  rhs,
float64_t sq_rhs,
int32_t  idx_b 
)

compute (a-b)^2 (== a^2+b^2-2ab) usually called by kernels'/distances' compute functions works on two feature vectors, although it is a member of a single feature: can either be called by lhs or rhs.

possible wiht subsets on lhs or rhs

Parameters:
lhs left-hand side features
sq_lhs squared values of left-hand side
idx_a index of left-hand side's vector to compute
rhs right-hand side features
sq_rhs squared values of right-hand side
idx_b index of right-hand side's vector to compute

Definition at line 996 of file SparseFeatures.h.

virtual CFeatures* copy_subset ( SGVector< index_t indices  )  [virtual]

Creates a new CFeatures instance containing copies of the elements which are specified by the provided indices.

Parameters:
indices indices of feature elements to copy
Returns:
new CFeatures instance with copies of feature data

Reimplemented from CFeatures.

Definition at line 1504 of file SparseFeatures.h.

virtual float64_t dense_dot ( int32_t  vec_idx1,
const float64_t vec2,
int32_t  vec2_len 
) [virtual]

compute dot product between vector1 and a dense vector

possible with subset

Parameters:
vec_idx1 index of first vector
vec2 pointer to real valued vector
vec2_len length of real valued vector

Implements CDotFeatures.

Definition at line 1388 of file SparseFeatures.h.

ST dense_dot ( ST  alpha,
int32_t  num,
ST *  vec,
int32_t  dim,
ST  b 
)

compute the dot product between dense weights and a sparse feature vector alpha * sparse^T * w + b

possible with subset

Parameters:
alpha scalar to multiply with
num index of feature vector
vec dense vector to compute dot product with
dim length of the dense vector
b bias
Returns:
dot product between dense weights and a sparse feature vector

Definition at line 463 of file SparseFeatures.h.

virtual float64_t dot ( int32_t  vec_idx1,
CDotFeatures df,
int32_t  vec_idx2 
) [virtual]

compute dot product between vector1 and vector2, appointed by their indices

possible with subset of this instance and of DotFeatures

Parameters:
vec_idx1 index of first vector
df DotFeatures (of same kind) to compute dot product with
vec_idx2 index of second vector

Implements CDotFeatures.

Definition at line 1361 of file SparseFeatures.h.

virtual CFeatures* duplicate (  )  const [virtual]

duplicate feature object

Returns:
feature object

Implements CFeatures.

Definition at line 202 of file SparseFeatures.h.

virtual void free_feature_iterator ( void *  iterator  )  [virtual]

clean up iterator call this function with the iterator returned by get_first_feature

Parameters:
iterator as returned by get_first_feature

Implements CDotFeatures.

Definition at line 1488 of file SparseFeatures.h.

void free_feature_vector ( SGSparseVector< ST >  vec,
int32_t  num 
)

free feature vector

possible with subset

Parameters:
vec feature vector to free
num index of vector in cache

Definition at line 934 of file SparseFeatures.h.

void free_sparse_feature_matrix (  ) 

free sparse feature matrix

any subset is removed

Definition at line 178 of file SparseFeatures.h.

void free_sparse_feature_vector ( SGSparseVector< ST >  vec,
int32_t  num 
)

free sparse feature vector

possible with subset

Parameters:
vec feature vector to free
num index of this vector in the cache

Definition at line 536 of file SparseFeatures.h.

void free_sparse_features (  ) 

free sparse feature matrix and cache

any subset is removed

Definition at line 191 of file SparseFeatures.h.

virtual int32_t get_dim_feature_space (  )  const [virtual]

obtain the dimensionality of the feature space

(not mix this up with the dimensionality of the input space, usually obtained via get_num_features())

Returns:
dimensionality

Implements CDotFeatures.

Definition at line 1347 of file SparseFeatures.h.

ST get_feature ( int32_t  num,
int32_t  index 
)

get a single feature

possible with subset

Parameters:
num number of feature vector to retrieve
index index of feature in this vector
Returns:
sum of features that match dimension index and 0 if none is found

Definition at line 216 of file SparseFeatures.h.

virtual EFeatureClass get_feature_class (  )  [virtual]

get feature class

Returns:
feature class SPARSE

Implements CFeatures.

Definition at line 919 of file SparseFeatures.h.

virtual void* get_feature_iterator ( int32_t  vector_index  )  [virtual]

iterate over the non-zero features

call get_feature_iterator first, followed by get_next_feature and free_feature_iterator to cleanup

possible with subset

Parameters:
vector_index the index of the vector over whose components to iterate over
Returns:
feature iterator (to be passed to get_next_feature)

Implements CDotFeatures.

Definition at line 1441 of file SparseFeatures.h.

virtual EFeatureType get_feature_type (  )  [virtual]

get feature type

Returns:
templated feature type

Implements CFeatures.

SGMatrix<ST> get_full_feature_matrix (  ) 

gets a copy of a full feature matrix

possible with subset

Returns:
full dense feature matrix

Definition at line 699 of file SparseFeatures.h.

SGVector<ST> get_full_feature_vector ( int32_t  num  ) 

get the fully expanded dense feature vector num

Returns:
dense feature vector
Parameters:
num index of feature vector

Definition at line 275 of file SparseFeatures.h.

ST* get_full_feature_vector ( int32_t  num,
int32_t &  len 
)

converts a sparse feature vector into a dense one preprocessed compute_feature_vector caller cleans up

Parameters:
num index of feature vector
len length is returned by reference
Returns:
dense feature vector

Definition at line 246 of file SparseFeatures.h.

virtual const char* get_name ( void   )  const [virtual]
Returns:
object name

Implements CSGObject.

Definition at line 1530 of file SparseFeatures.h.

virtual bool get_next_feature ( int32_t &  index,
float64_t value,
void *  iterator 
) [virtual]

iterate over the non-zero features

call this function with the iterator returned by get_first_feature and call free_feature_iterator to cleanup

Parameters:
index is returned by reference (-1 when not available)
value is returned by reference
iterator as returned by get_first_feature
Returns:
true if a new non-zero feature got returned

Implements CDotFeatures.

Definition at line 1469 of file SparseFeatures.h.

virtual int32_t get_nnz_features_for_vector ( int32_t  num  )  [virtual]

get number of non-zero features in vector

Parameters:
num which vector
Returns:
number of non-zero features in vector

Implements CDotFeatures.

Definition at line 311 of file SparseFeatures.h.

int32_t get_num_features (  ) 

get number of features

Returns:
number of features

Definition at line 894 of file SparseFeatures.h.

int64_t get_num_nonzero_entries (  ) 

get number of non-zero entries in sparse feature matrix

Returns:
number of non-zero entries in sparse feature matrix

Definition at line 946 of file SparseFeatures.h.

virtual int32_t get_num_vectors (  )  const [virtual]

get number of feature vectors, possibly of subset

Returns:
number of feature vectors

Implements CFeatures.

Definition at line 885 of file SparseFeatures.h.

virtual int32_t get_size (  )  [virtual]

get memory footprint of one feature

Returns:
memory footprint of one feature

Implements CFeatures.

Definition at line 864 of file SparseFeatures.h.

SGSparseVector<ST>* get_sparse_feature_matrix ( int32_t &  num_feat,
int32_t &  num_vec 
)

get the pointer to the sparse feature matrix num_feat,num_vectors are returned by reference

not possible with subset

Parameters:
num_feat number of features in matrix
num_vec number of vectors in matrix
Returns:
feature matrix

Definition at line 553 of file SparseFeatures.h.

SGSparseMatrix<ST> get_sparse_feature_matrix (  ) 

get the sparse feature matrix

not possible with subset

Returns:
sparse matrix

Definition at line 571 of file SparseFeatures.h.

SGSparseVector<ST> get_sparse_feature_vector ( int32_t  num  ) 

get sparse feature vector for sample num from the matrix as it is if matrix is initialized, else return preprocessed compute_feature_vector

possible with subset

Parameters:
num index of feature vector
Returns:
sparse feature vector

Definition at line 328 of file SparseFeatures.h.

SGSparseVector<ST>* get_transposed ( int32_t &  num_feat,
int32_t &  num_vec 
)

compute and return the transpose of the sparse feature matrix which will be prepocessed. num_feat, num_vectors are returned by reference caller has to clean up

possible with subset

Parameters:
num_feat number of features in matrix
num_vec number of vectors in matrix
Returns:
transposed sparse feature matrix

Definition at line 622 of file SparseFeatures.h.

CSparseFeatures<ST>* get_transposed (  ) 

get a transposed copy of the features

possible with subset

Returns:
transposed copy

Definition at line 603 of file SparseFeatures.h.

void load ( CFile loader  )  [virtual]

load features from file

any subset is removed before

Parameters:
loader File object to load data from

Reimplemented from CFeatures.

CLabels* load_svmlight_file ( char *  fname,
bool  do_sort_features = true 
)

load features from file

any subset is removed before

Parameters:
fname filename to load from
do_sort_features if true features will be sorted to ensure they are in ascending order
Returns:
label object with corresponding labels

Definition at line 1079 of file SparseFeatures.h.

bool obtain_from_simple ( CSimpleFeatures< ST > *  sf  ) 

obtain sparse features from simple features

subset on input is ignored, subset of this instance is removed

Parameters:
sf simple features
Returns:
if obtaining was successful

Definition at line 873 of file SparseFeatures.h.

void save ( CFile writer  )  [virtual]

save features to file

not possible with subset

Parameters:
writer File object to write data to

Reimplemented from CFeatures.

virtual bool set_full_feature_matrix ( SGMatrix< ST >  full  )  [virtual]

creates a sparse feature matrix from a full dense feature matrix necessary to set feature_matrix, num_features and num_vectors where num_features is the column offset, and columns are linear in memory see above for definition of sparse_feature_matrix

any subset is removed before

Parameters:
full full feature matrix

Definition at line 737 of file SparseFeatures.h.

int32_t set_num_features ( int32_t  num  ) 

set number of features

Sometimes when loading sparse features not all possible dimensions are used. This may pose a problem to classifiers when being applied to higher dimensional test-data. This function allows to artificially explode the feature space

Parameters:
num the number of features, must be larger than the current number of features
Returns:
previous number of features

Definition at line 907 of file SparseFeatures.h.

void set_sparse_feature_matrix ( SGSparseMatrix< ST >  sm  ) 

set sparse feature matrix

not possible with subset

Parameters:
sm sparse feature matrix

Definition at line 679 of file SparseFeatures.h.

void sort_features (  ) 

ensure that features occur in ascending order, only call when no preprocessors are attached

not possiblwe with subset

Definition at line 1251 of file SparseFeatures.h.

static ST sparse_dot ( ST  alpha,
SGSparseVectorEntry< ST > *  avec,
int32_t  alen,
SGSparseVectorEntry< ST > *  bvec,
int32_t  blen 
) [static]

compute the dot product between two sparse feature vectors alpha * vec^T * vec

Parameters:
alpha scalar to multiply with
avec first sparse feature vector
alen avec's length
bvec second sparse feature vector
blen bvec's length
Returns:
dot product between the two sparse feature vectors

Definition at line 403 of file SparseFeatures.h.

bool write_svmlight_file ( char *  fname,
CLabels label 
)

write features to file using svm light format

not possible with subset

Parameters:
fname filename to write to
label Label object (number of labels must correspond to number of features)
Returns:
true if successful

Definition at line 1304 of file SparseFeatures.h.


Member Data Documentation

feature cache

Definition at line 1577 of file SparseFeatures.h.

int32_t num_features [protected]

total number of features

Definition at line 1571 of file SparseFeatures.h.

int32_t num_vectors [protected]

total number of vectors

Definition at line 1568 of file SparseFeatures.h.

array of sparse vectors of size num_vectors

Definition at line 1574 of file SparseFeatures.h.


The documentation for this class was generated from the following file:
 All Classes Namespaces Files Functions Variables Typedefs Enumerations Enumerator Friends Defines

SHOGUN Machine Learning Toolbox - Documentation