PLearn 0.1
|
00001 // -*- C++ -*- 00002 00003 // SumVarianceOfLinearTransformedCategoricals.cc 00004 // 00005 // Copyright (C) 2009 Pascal Vincent 00006 // 00007 // Redistribution and use in source and binary forms, with or without 00008 // modification, are permitted provided that the following conditions are met: 00009 // 00010 // 1. Redistributions of source code must retain the above copyright 00011 // notice, this list of conditions and the following disclaimer. 00012 // 00013 // 2. Redistributions in binary form must reproduce the above copyright 00014 // notice, this list of conditions and the following disclaimer in the 00015 // documentation and/or other materials provided with the distribution. 00016 // 00017 // 3. The name of the authors may not be used to endorse or promote 00018 // products derived from this software without specific prior written 00019 // permission. 00020 // 00021 // THIS SOFTWARE IS PROVIDED BY THE AUTHORS ``AS IS'' AND ANY EXPRESS OR 00022 // IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES 00023 // OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN 00024 // NO EVENT SHALL THE AUTHORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, 00025 // SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED 00026 // TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR 00027 // PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF 00028 // LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING 00029 // NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS 00030 // SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 00031 // 00032 // This file is part of the PLearn library. For more information on the PLearn 00033 // library, go to the PLearn Web site at www.plearn.org 00034 00035 // Authors: Pascal Vincent 00036 00040 #include "SumVarianceOfLinearTransformedCategoricals.h" 00041 00042 namespace PLearn { 00043 using namespace std; 00044 00047 PLEARN_IMPLEMENT_OBJECT( 00048 SumVarianceOfLinearTransformedCategoricals, 00049 "Computes the sum of the variances of the elements of a linear transformation of a concatenation of independent random variables following a categorical distribution.", 00050 "By categorical distribution we mean multinomials with the number-of-trials parameter set to n=1.\n" 00051 "Let P=inpu1 a (l,d') matrix. Each row contains the parameters of groupsize.length() categoricals.\n" 00052 " ( the sum of the groupsize elements equals d').\n" 00053 "Let H ~ MultipleCategorical(P) a (l,d') corresponding random variable, \n" 00054 " this correponds to drawings from independent categorical variables whose probablity parameters are those in P \n" 00055 "Let W=input2 a (d,d') linear transformation matrix. \n" 00056 "Let X=H W^t a (l,d) random variable matrix corresponding to applying the transformation. \n" 00057 " i.e. X_i = sum_j H_ij W^t_j \n" 00058 "SumVarianceOfLinearTransformedCategoricals computes the sum of the variances of the elements of X.\n" 00059 "i.e. sum_ij Var[X_ij] \n" 00060 ); 00061 00062 SumVarianceOfLinearTransformedCategoricals::SumVarianceOfLinearTransformedCategoricals() 00063 : inherited(0,0,1,1) 00064 {} 00065 00066 SumVarianceOfLinearTransformedCategoricals::SumVarianceOfLinearTransformedCategoricals(Variable* input1, Variable* input2, 00067 bool call_build_) 00068 : inherited(input1, input2, 1, 1, call_build_) 00069 { 00070 if (call_build_) 00071 build_(); 00072 } 00073 00074 // constructor from input variable and parameters 00075 // SumVarianceOfLinearTransformedCategoricals::SumVarianceOfLinearTransformedCategoricals(Variable* input1, Variable* input2, 00076 // param_type the_parameter, ..., 00077 // bool call_build_) 00078 // ### replace with actual parameters 00079 // : inherited(input1, input2, this_variable_length, this_variable_width, 00080 // call_build_), 00081 // parameter(the_parameter), 00082 // ... 00083 //{ 00084 // if (call_build_) 00085 // build_(); 00086 //} 00087 00088 void SumVarianceOfLinearTransformedCategoricals::recomputeSize(int& l, int& w) const 00089 { 00090 l = 1; 00091 w = 1; 00092 } 00093 00094 // ### computes value from input1 and input2 values 00095 void SumVarianceOfLinearTransformedCategoricals::fprop() 00096 { 00097 if(input1.width()!=input2.width()) 00098 PLERROR("Incompatible sizes: width of P (input1) must equal width of W (input2)"); 00099 00100 Mat P = input1->matValue; 00101 Mat W = input2->matValue; 00102 int d = W.length(); 00103 int l = P.length(); 00104 // int m = W.width(); // should equal sum of groupsizes 00105 00106 double simplesum = 0; 00107 double sqsum = 0; 00108 00109 int ngroups = groupsizes.length(); 00110 const int* pgroupsize = groupsizes.data(); 00111 00112 for(int t=0; t<l; t++) 00113 { 00114 const real* p = P[t]; 00115 for(int i=0; i<d; i++) 00116 { 00117 const real* Wi = W[i]; 00118 int k = 0; 00119 for(int groupnum=0; groupnum<ngroups; groupnum++) 00120 { 00121 double tmpsqsum = 0; 00122 int gs = pgroupsize[groupnum]; 00123 while(gs--) 00124 { 00125 real Wik = Wi[k]; 00126 real pk = p[k]; 00127 real Wik_pk = Wik*pk; 00128 simplesum += Wik*Wik_pk; 00129 tmpsqsum += Wik_pk; 00130 k++; 00131 } 00132 sqsum += tmpsqsum*tmpsqsum; 00133 } 00134 } 00135 } 00136 value[0] = simplesum-sqsum; 00137 } 00138 00139 // ### computes input1 and input2 gradients from gradient 00140 void SumVarianceOfLinearTransformedCategoricals::bprop() 00141 { 00142 Mat P = input1->matValue; 00143 Mat Pgrad = input1->matGradient; 00144 Mat W = input2->matValue; 00145 Mat Wgrad = input2->matGradient; 00146 int d = W.length(); 00147 int l = P.length(); 00148 // int m = W.width(); // should equal sum of groupsizes 00149 00150 real gr = gradient[0]; 00151 00152 int ngroups = groupsizes.length(); 00153 const int* pgroupsize = groupsizes.data(); 00154 00155 group_sum_wik_pk.resize(ngroups); 00156 real* p_group_sum_wik_pk = group_sum_wik_pk.data(); 00157 00158 for(int t=0; t<l; t++) 00159 { 00160 const real* p = P[t]; 00161 real* gp = Pgrad[t]; 00162 for(int i=0; i<d; i++) 00163 { 00164 const real* Wi = W[i]; 00165 int k = 0; 00166 for(int groupnum=0; groupnum<ngroups; groupnum++) 00167 { 00168 int gs = pgroupsize[groupnum]; 00169 double sum_wik_pk = 0; 00170 while(gs--) 00171 { 00172 sum_wik_pk += Wi[k]*p[k]; 00173 k++; 00174 } 00175 p_group_sum_wik_pk[groupnum] = sum_wik_pk; 00176 } 00177 00178 real* gWi = Wgrad[i]; 00179 k = 0; 00180 for(int groupnum=0; groupnum<ngroups; groupnum++) 00181 { 00182 int gs = pgroupsize[groupnum]; 00183 real grsum = p_group_sum_wik_pk[groupnum]; 00184 while(gs--) 00185 { 00186 real Wik = Wi[k]; 00187 gWi[k] += gr*2*p[k]*(Wik-grsum); 00188 gp[k] += gr*Wik*(Wik-2*grsum); 00189 k++; 00190 } 00191 } 00192 } 00193 } 00194 00195 } 00196 00197 // ### You can implement these methods: 00198 // void SumVarianceOfLinearTransformedCategoricals::bbprop() {} 00199 // void SumVarianceOfLinearTransformedCategoricals::symbolicBprop() {} 00200 // void SumVarianceOfLinearTransformedCategoricals::rfprop() {} 00201 00202 00203 // ### Nothing to add here, simply calls build_ 00204 void SumVarianceOfLinearTransformedCategoricals::build() 00205 { 00206 inherited::build(); 00207 build_(); 00208 } 00209 00210 void SumVarianceOfLinearTransformedCategoricals::makeDeepCopyFromShallowCopy(CopiesMap& copies) 00211 { 00212 inherited::makeDeepCopyFromShallowCopy(copies); 00213 00214 // ### Call deepCopyField on all "pointer-like" fields 00215 // ### that you wish to be deepCopied rather than 00216 // ### shallow-copied. 00217 // ### ex: 00218 // deepCopyField(trainvec, copies); 00219 deepCopyField(groupsizes, copies); 00220 deepCopyField(group_sum_wik_pk, copies); 00221 00222 // ### If you want to deepCopy a Var field: 00223 // varDeepCopyField(somevariable, copies); 00224 } 00225 00226 void SumVarianceOfLinearTransformedCategoricals::declareOptions(OptionList& ol) 00227 { 00228 // ### Declare all of this object's options here. 00229 // ### For the "flags" of each option, you should typically specify 00230 // ### one of OptionBase::buildoption, OptionBase::learntoption or 00231 // ### OptionBase::tuningoption. If you don't provide one of these three, 00232 // ### this option will be ignored when loading values from a script. 00233 // ### You can also combine flags, for example with OptionBase::nosave: 00234 // ### (OptionBase::buildoption | OptionBase::nosave) 00235 00236 declareOption(ol, "groupsizes", &SumVarianceOfLinearTransformedCategoricals::groupsizes, 00237 OptionBase::buildoption, 00238 "defines the dimensions of the categorical variables."); 00239 00240 // Now call the parent class' declareOptions 00241 inherited::declareOptions(ol); 00242 } 00243 00244 void SumVarianceOfLinearTransformedCategoricals::build_() 00245 { 00246 // ### This method should do the real building of the object, 00247 // ### according to set 'options', in *any* situation. 00248 // ### Typical situations include: 00249 // ### - Initial building of an object from a few user-specified options 00250 // ### - Building of a "reloaded" object: i.e. from the complete set of 00251 // ### all serialised options. 00252 // ### - Updating or "re-building" of an object after a few "tuning" 00253 // ### options have been modified. 00254 // ### You should assume that the parent class' build_() has already been 00255 // ### called. 00256 } 00257 00258 00259 } // end of namespace PLearn 00260 00261 00262 /* 00263 Local Variables: 00264 mode:c++ 00265 c-basic-offset:4 00266 c-file-style:"stroustrup" 00267 c-file-offsets:((innamespace . 0)(inline-open . 0)) 00268 indent-tabs-mode:nil 00269 fill-column:79 00270 End: 00271 */ 00272 // vim: filetype=cpp:expandtab:shiftwidth=4:tabstop=8:softtabstop=4:encoding=utf-8:textwidth=79 :