Multiclass ecoc #546

…e to messy problems. Parameter registration removed

get rid of SGVector<void/char>

СMap for mallocs tracing

Correct the equations in comment.

So one can actually do a cross-validation with multiclass linear machines

Add vanilla store_model_features for LinearMulticlassMachine

- later SGMatrix/SGStringList/SGNDArray/SGSparseMatrix will all use this

fixes a memory error in one example

* 'master' of git://github.com/shogun-toolbox/shogun: added debug messages code whitespace fix a memory error retun *this on self assignment derive SGVector from SGReferencedData to avoid code duplication Add vanilla store_model_features for LinearMulticlassMachine So one can actually do a cross-validation with multiclass linear machines Correct the equations in comment.

apply() had an extra SG_REF on the returned CLables

… given ones

use SGMatrix mat(dims, nvec) whereever possible

- make ref() function protected: when assigning SGVector/Matrix objects they will automagically call ref() so an extra ref() is not necessary - ensure that when calling unref() the ptr to the refcount and the data is set to NULL - this is OK since there must be one object that still exists that will then later free the data. this enables calling unref() twice w/o getting double free errors

in the multiclasslinearmachine example

store_model_features for KernelMulticlassMachine.

Use index_t instead of int32_t for index type.

CSet and CMap

CSet and CMap improvements

- Added simplified ported tree_leastR feature tree guided lasso - The code is licensed by GPLv3 license accepted by the authors of the SLEP.

- fixed wrong vector length in Homogeneous Kernel Map preprocessor

Closes #445 This reverts commit ba71ded.

Add the example to the Makefile and fix a typo in the example

Minor fixes for evaluation_cross_validation_multiclass

No need for copying the vector on the stack it's enough to pass a const reference to it as we are only reading from it.

HKM: fix warning and pass vector by reference

MC Accuracy, some reference results (liblinear as base classifier): * DiscriminantECOC: 0.8206 * RandomDense: 0.9043 * RandomSparse: 0.9274 and some reference number from mldata site: * KNN (k=1) 0.9800 * KNN (k=5) 0.9789 * KNN (k=10) 0.9755

OVR : 0.9452 ECOCOVR: 0.8911

interfaces

- drop SGVector& typemaps - whitespace fixes

-instead confidence parameter, test now computes p-value -moved bootstrapping to two-sample-test-statistic to have it available for two-sample based tests -some minor changes and documentation updates

some design updates

it causes a null ptr exception - swig/lua seem not to be able to convert the table into a CStringFeatures<char> object but 0x0 ...

CArray* to CDynamicArray conversion

Multiclass ecoc

Fixed codestyle and doc's for DynamicArrays

Minor fixes isolated

Multiclass ecoc

- drop it from file formats - typemaps - parameters

NOTE: does not correspond to real vector index in sparse fm

…binary subtask

CLabels -> CDenseLabels -> CRealLabels -> CBinaryLabels -> CMulticlassLabels -> CRealLabels - put labels under shogun/labels/ - change include paths in all files - CMachine still returns a generic CLabels object in its apply() function - but submachines create customized CRealLabels, CMulticlassLabels etc objects. - all machines now check if their train function is called with matching label type - type of label object can be determined via label->get_label_type() - introduce confidences in labels - introduce set_to_const function in labels

Minor changes in QDA

Multiclass ecoc

…constructor)

Fix compile warning.

…NULL)

…constructor

add type casting functions that enable CLabel -> CMulticlassLabel etc conversion

- apply() of CMachine now calls one of appropriate apply_binary(), apply_regression() or apply_multiclass() - apply(int32_t) renamed to apply_one(int32_t)

- Added NativeMulticlassMachine for classifiers without submachines

Add problem types to machines.

It is not proper to call this in constructor. For example, the return value from apply() might be an "invalid" BinaryLabels because not both positive and negative labels presented.

Ensure related kernels call labels->ensure_valid

…raining, made clustering algorithms do not require labels

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Multiclass ecoc #546

Multiclass ecoc #546

Commits on May 7, 2012

Commits on May 8, 2012

Commits on May 9, 2012

Commits on May 10, 2012

Commits on May 11, 2012

Commits on May 12, 2012

Commits on May 13, 2012

Commits on May 14, 2012

Commits on May 15, 2012

Commits on May 16, 2012

Commits on May 17, 2012

Commits on May 18, 2012

Commits on May 20, 2012

Commits on May 21, 2012

Commits on May 22, 2012