Added distributions::InvCDF for weighted sampling of features to make…

… the split. Unit tests added as well.
IlyaLab · Oct 22, 2012 · 6ff5b96 · 6ff5b96
1 parent 91953f3
commit 6ff5b96
Show file tree

Hide file tree

Showing 7 changed files with 199 additions and 7 deletions.
diff --git a/Makefile b/Makefile
@@ -1,4 +1,4 @@
-COMPILER = g++
+COMPILER = g++44
 CFLAGS = -O2 -std=c++0x -Wall -Wextra -pedantic -Isrc/
 TFLAGS = -pthread
 SOURCEFILES = src/progress.cpp src/statistics.cpp src/math.cpp src/stochasticforest.cpp src/rootnode.cpp src/node.cpp src/treedata.cpp src/datadefs.cpp src/utils.cpp src/distributions.cpp

diff --git a/src/distributions.cpp b/src/distributions.cpp
@@ -28,3 +28,34 @@ num_t distributions::RandInt::uniform() {
 
 }
 
+distributions::InvCDF::InvCDF(const vector<num_t>& weights) {
+
+  num_t sum = 0.0;
+
+  for ( size_t i = 0; i < weights.size(); ++i ) {
+    assert( weights[i] >= 0.0 );
+    sum += weights[i];
+  }
+
+  num_t cumProb = 0.0;
+
+  assert( sum > 0.0 );
+
+  for ( size_t i = 0; i < weights.size(); ++i ) {
+    if ( weights[i] > 0.0 ) {
+      cumProb += weights[i] / sum;
+      icdf_[cumProb] = i;
+    }
+  }
+
+}
+
+distributions::InvCDF::~InvCDF() {
+
+}
+
+size_t distributions::InvCDF::at(const num_t prob) {
+  assert(prob >= 0.0 && prob < 1.0);
+
+  return( icdf_.upper_bound(prob)->second );
+}
diff --git a/src/distributions.hpp b/src/distributions.hpp
@@ -45,8 +45,15 @@ namespace distributions {
     tr1::uniform_int<size_t> rand_;
 
   };
-
-  typedef map<datadefs::num_t,size_t> icdf_t;
+
+  class InvCDF {
+  public:
+    InvCDF(const vector<datadefs::num_t>& weights);
+    ~InvCDF();
+    size_t at(const datadefs::num_t prob);
+  private:
+    map<datadefs::num_t,size_t> icdf_;
+  };
 
   //icdf_t 
 

diff --git a/test/distributions_test.hpp b/test/distributions_test.hpp
@@ -9,6 +9,9 @@ class DistributionsTest : public CppUnit::TestFixture {
   CPPUNIT_TEST_SUITE( DistributionsTest );
   CPPUNIT_TEST( test_randint );
   CPPUNIT_TEST( test_uniform );
+  CPPUNIT_TEST( test_invcdf1 );
+  CPPUNIT_TEST( test_invcdf2 );
+  CPPUNIT_TEST( test_invcdf3 );
   CPPUNIT_TEST_SUITE_END();
 
 public:
@@ -17,6 +20,9 @@ class DistributionsTest : public CppUnit::TestFixture {
 
   void test_randint();
   void test_uniform();
+  void test_invcdf1();
+  void test_invcdf2();
+  void test_invcdf3();
 
 private:
 
@@ -84,7 +90,7 @@ void DistributionsTest::test_uniform() {
 
   for ( size_t i = 0; i < 100000; ++i ) {
     datadefs::num_t r = randInt_.uniform();
-    CPPUNIT_ASSERT( 0.0 <= r && r <= 1.0 );
+    CPPUNIT_ASSERT( 0.0 <= r && r < 1.0 );
 
     //cout << " " << r; 
 
@@ -101,7 +107,53 @@ void DistributionsTest::test_uniform() {
 
 }
 
+void DistributionsTest::test_invcdf1() {
 
+  distributions::InvCDF icdf({0.2,0.2,0.2,0.2,0.2});
+
+  CPPUNIT_ASSERT( icdf.at(0.00) == 0 );
+  CPPUNIT_ASSERT( icdf.at(0.05) == 0 );
+  CPPUNIT_ASSERT( icdf.at(0.15) == 0 );
+  CPPUNIT_ASSERT( icdf.at(0.25) == 1 );
+  CPPUNIT_ASSERT( icdf.at(0.35) == 1 );
+  CPPUNIT_ASSERT( icdf.at(0.45) == 2 );
+  CPPUNIT_ASSERT( icdf.at(0.55) == 2 );
+  CPPUNIT_ASSERT( icdf.at(0.65) == 3 );
+  CPPUNIT_ASSERT( icdf.at(0.75) == 3 );
+  CPPUNIT_ASSERT( icdf.at(0.85) == 4 );
+  CPPUNIT_ASSERT( icdf.at(0.95) == 4 );
+  CPPUNIT_ASSERT( icdf.at(0.999999999999) == 4 );
+
+}
+
+void DistributionsTest::test_invcdf2() {
+
+  distributions::InvCDF icdf({1,2,3,0,4});
+
+  CPPUNIT_ASSERT( icdf.at(0.00) == 0 );
+  CPPUNIT_ASSERT( icdf.at(0.05) == 0 );
+  CPPUNIT_ASSERT( icdf.at(0.15) == 1 );
+  CPPUNIT_ASSERT( icdf.at(0.25) == 1 );
+  CPPUNIT_ASSERT( icdf.at(0.35) == 2 );
+  CPPUNIT_ASSERT( icdf.at(0.45) == 2 );
+  CPPUNIT_ASSERT( icdf.at(0.55) == 2 );
+  CPPUNIT_ASSERT( icdf.at(0.65) == 4 );
+  CPPUNIT_ASSERT( icdf.at(0.75) == 4 );
+  CPPUNIT_ASSERT( icdf.at(0.85) == 4 );
+  CPPUNIT_ASSERT( icdf.at(0.95) == 4 );
+  CPPUNIT_ASSERT( icdf.at(0.99999999999) == 4 );
+
+} 
+
+void DistributionsTest::test_invcdf3() {
+
+  distributions::InvCDF icdf({0,0,1,0,0});
+
+  for ( size_t i = 0; i < 100; ++i ) {
+    CPPUNIT_ASSERT( icdf.at(static_cast<datadefs::num_t>(0.01*i)) == 2 );
+  }
+
+}
 
 
 // Registers the fixture into the test 'registry'

diff --git a/test/stochasticforest_test.hpp b/test/stochasticforest_test.hpp
@@ -56,7 +56,7 @@ void StochasticForestTest::test_treeDataPercolation() {
   Treedata::Treedata testData("testdata.tsv",&parameters_);
 
   // Next load a test predictor
-  StochasticForest SF(&parameters_);
+  StochasticForest SF(parameters_);
 
   vector<num_t> prediction,confidence;
 
@@ -109,7 +109,7 @@ void StochasticForestTest::test_CART() {
 
   parameters_.nMaxLeaves = 2;
 
-  StochasticForest CART(&treeData,&parameters_);
+  StochasticForest CART(&treeData,parameters_);
 
   CPPUNIT_ASSERT( CART.rootNodes_[0]->splitter_.name == "N:input" );
 

diff --git a/test/treedata_test.hpp b/test/treedata_test.hpp
@@ -767,7 +767,7 @@ void treeDataTest::test_fullSplitterSweep() {
   //size_t targetIdx = treeData->getFeatureIdx("N:output");
   size_t minSamples = 3;
 
-  vector<string> sweepFiles = {"test_fullSplitterSweep.txt","test_fullSplitterSweep_C:class.txt"};
+  vector<string> sweepFiles = {"test_fullSplitterSweep.txt","test_fullSplitterSweep_class.txt"};
 
   for ( size_t f = 0; f < sweepFiles.size(); ++f ) {
 

diff --git a/test_fullSplitterSweep_class.txt b/test_fullSplitterSweep_class.txt
@@ -0,0 +1,102 @@
+C:class	N:output	8.277247e-02	132	146
+C:class	N:input	1.080478e-02	273	4
+C:class	N:noise_1	1.186730e-02	145	131
+C:class	N:noise_2	9.047567e-03	149	131
+C:class	N:noise_3	5.632269e-03	37	240
+C:class	N:noise_4	2.012800e-02	101	176
+C:class	N:noise_5	4.715754e-03	144	132
+C:class	C:noise_6	1.915083e-03	93	178
+C:class	N:noise_7	8.512985e-03	263	19
+C:class	N:noise_8	4.338452e-03	17	260
+C:class	N:noise_9	7.568886e-03	157	128
+C:class	N:noise_10	5.742623e-03	212	67
+C:class	N:noise_11	8.089438e-03	267	7
+C:class	N:noise_12	6.164748e-03	162	107
+C:class	N:noise_13	5.167658e-03	144	132
+C:class	N:noise_14	7.158856e-03	265	13
+C:class	N:noise_15	8.629035e-03	31	245
+C:class	N:noise_16	1.294779e-02	270	6
+C:class	N:noise_17	3.829856e-03	43	233
+C:class	C:noise_18	1.939052e-02	89	183
+C:class	N:noise_19	7.300695e-03	13	262
+C:class	N:noise_20	1.300216e-02	88	188
+C:class	N:noise_21	7.683229e-03	227	43
+C:class	N:noise_22	1.820842e-02	8	268
+C:class	C:noise_23	4.497780e-03	77	195
+C:class	N:noise_24	1.553144e-02	31	240
+C:class	N:noise_25	7.985333e-03	28	244
+C:class	N:noise_26	1.021783e-02	213	64
+C:class	N:noise_27	5.208401e-03	32	239
+C:class	N:noise_28	8.072889e-03	250	28
+C:class	N:noise_29	7.635241e-03	10	265
+C:class	N:noise_30	1.439265e-02	225	50
+C:class	N:noise_31	9.134501e-03	269	10
+C:class	N:noise_32	1.537890e-02	185	89
+C:class	N:noise_33	7.579281e-03	3	278
+C:class	N:noise_34	1.615210e-02	41	231
+C:class	N:noise_35	9.442723e-03	52	227
+C:class	N:noise_36	9.529369e-03	205	70
+C:class	N:noise_37	8.874968e-03	182	90
+C:class	N:noise_38	6.306706e-03	83	191
+C:class	N:noise_39	6.690488e-03	13	257
+C:class	N:noise_40	6.732478e-03	156	123
+C:class	N:noise_41	6.800112e-03	258	16
+C:class	N:noise_42	1.330721e-02	5	271
+C:class	N:noise_43	6.459754e-03	251	22
+C:class	N:noise_44	1.463259e-02	86	190
+C:class	N:noise_45	1.074883e-02	57	221
+C:class	N:noise_46	4.948310e-03	249	27
+C:class	N:noise_47	1.185683e-02	74	201
+C:class	N:noise_48	5.228776e-03	203	68
+C:class	N:noise_49	8.015075e-03	270	3
+C:class	N:noise_50	3.827547e-03	245	26
+C:class	N:noise_51	5.623361e-03	8	260
+C:class	N:noise_52	1.132672e-02	264	11
+C:class	N:noise_53	1.474747e-02	198	72
+C:class	N:noise_54	1.464262e-02	184	93
+C:class	N:noise_55	1.061787e-02	57	218
+C:class	N:noise_56	9.102942e-03	275	7
+C:class	N:noise_57	8.660064e-03	4	270
+C:class	N:noise_58	1.468407e-02	7	273
+C:class	N:noise_59	4.682519e-03	254	14
+C:class	N:noise_60	8.742796e-03	87	189
+C:class	N:noise_61	6.415251e-03	276	3
+C:class	C:noise_62	3.877685e-03	91	182
+C:class	N:noise_63	9.286570e-03	252	26
+C:class	N:noise_64	7.245304e-03	6	272
+C:class	N:noise_65	8.578843e-03	274	4
+C:class	C:noise_66	3.314331e-03	100	172
+C:class	N:noise_67	8.746950e-03	268	4
+C:class	N:noise_68	8.593788e-03	63	215
+C:class	C:noise_69	4.556471e-03	83	198
+C:class	N:noise_70	1.385411e-02	269	5
+C:class	N:noise_71	7.371832e-03	45	234
+C:class	N:noise_72	5.581014e-03	266	7
+C:class	N:noise_73	5.526464e-03	222	50
+C:class	C:noise_74	1.235055e-03	86	186
+C:class	C:noise_75	2.357321e-03	81	190
+C:class	N:noise_76	5.492676e-03	231	42
+C:class	N:noise_77	1.658936e-02	207	66
+C:class	N:noise_78	9.358209e-03	38	239
+C:class	N:noise_79	5.969206e-03	185	95
+C:class	N:noise_80	1.487118e-02	17	260
+C:class	N:noise_81	5.329469e-03	268	5
+C:class	N:noise_82	7.679314e-03	246	26
+C:class	N:noise_83	9.127313e-03	7	274
+C:class	N:noise_84	4.509829e-03	137	143
+C:class	N:noise_85	4.530825e-03	110	169
+C:class	C:noise_86	1.859930e-03	78	199
+C:class	N:noise_87	1.187371e-02	30	244
+C:class	N:noise_88	5.670918e-03	200	80
+C:class	N:noise_89	7.086722e-03	21	256
+C:class	N:noise_90	1.036187e-02	207	59
+C:class	N:noise_91	1.046930e-02	4	274
+C:class	N:noise_92	1.057821e-02	239	38
+C:class	N:noise_93	7.872152e-03	45	233
+C:class	N:noise_94	8.085322e-03	23	250
+C:class	C:noise_95	1.761905e-03	100	180
+C:class	N:noise_96	9.914971e-03	144	131
+C:class	N:noise_97	1.001201e-02	17	261
+C:class	N:noise_98	7.998670e-03	274	3
+C:class	N:noise_99	6.701586e-03	221	60
+C:class	N:noise_100	1.181143e-02	191	82