PLearn 0.1
|
00001 // -*- C++ -*- 00002 00003 // TransposedDoubleProductVariable.cc 00004 // 00005 // Copyright (C) 2007 Simon Lemieux, 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: Simon Lemieux, Pascal Vincent 00036 00040 #include "TransposedDoubleProductVariable.h" 00041 00042 namespace PLearn { 00043 using namespace std; 00044 00047 PLEARN_IMPLEMENT_OBJECT( 00048 TransposedDoubleProductVariable, 00049 "Let H, W and M be the inputs and nw the length of W. Then output(n,k) = sum_i{ sum_j { W(i,k)*M(j,k)*H(n,i+j*nw) } }", 00050 "MULTI LINE\nHELP FOR USERS" 00051 ); 00052 00053 TransposedDoubleProductVariable::TransposedDoubleProductVariable() 00054 {} 00055 00056 TransposedDoubleProductVariable::TransposedDoubleProductVariable(Var h, Var w, Var m) 00057 : inherited(h & w & m, h.length(), w.width()) 00058 { 00059 build_(); 00060 } 00061 00062 00063 void TransposedDoubleProductVariable::recomputeSize(int& l, int& w) const 00064 { 00065 if (varray.size() > 0) { 00066 // l = varH()->length(); // the computed length of this Var 00067 //w = varW()->width(); // the computed width 00068 l = varray[0].length(); 00069 w = varray[1].width(); 00070 } else 00071 l = w = 0; 00072 } 00073 00074 // ### computes value from varray values 00075 void TransposedDoubleProductVariable::fprop() 00076 { 00077 Mat w = varW()->matValue, 00078 m = varM()->matValue, 00079 h = varH()->matValue; 00080 00081 int nw = w.length(), 00082 nm = m.length(), 00083 nx = h.length(), 00084 d = w.width(); 00085 00086 for(int n=0; n<nx; n++) 00087 for(int k=0; k<d; k++) 00088 { 00089 matValue(n,k)=0.; 00090 for(int i=0; i<nw; i++) 00091 for(int j=0; j<nm; j++) 00092 matValue(n,k) += w(i,k)*m(j,k)*h(n,i+j*nw); 00093 } 00094 } 00095 00096 // ### computes varray gradients from gradient 00097 void TransposedDoubleProductVariable::bprop() 00098 { 00099 Mat w = varW()->matValue, 00100 m = varM()->matValue, 00101 h = varH()->matValue, 00102 w_grad = varW()->matGradient, 00103 m_grad = varM()->matGradient, 00104 h_grad = varH()->matGradient, 00105 x_grad = matGradient; 00106 00107 int nw = w.length(), 00108 nm = m.length(), 00109 nx = h.length(), 00110 d = w.width(); 00111 00112 for(int n=0; n<nx; n++) 00113 for(int k=0; k<d; k++) 00114 for(int i=0; i<nw; i++) 00115 for(int j=0; j<nm; j++) 00116 { 00117 w_grad(i,k) += x_grad(n,k)*m(j,k)*h(n,i+j*nw); 00118 m_grad(j,k) += x_grad(n,k)*w(i,k)*h(n,i+j*nw); 00119 h_grad(n,i+j*nw) += x_grad(n,k)*w(i,k)*m(j,k); 00120 } 00121 } 00122 00123 // ### You can implement these methods: 00124 // void TransposedDoubleProductVariable::bbprop() {} 00125 // void TransposedDoubleProductVariable::symbolicBprop() {} 00126 // void TransposedDoubleProductVariable::rfprop() {} 00127 00128 00129 // ### Nothing to add here, simply calls build_ 00130 void TransposedDoubleProductVariable::build() 00131 { 00132 inherited::build(); 00133 build_(); 00134 } 00135 00136 void TransposedDoubleProductVariable::makeDeepCopyFromShallowCopy(CopiesMap& copies) 00137 { 00138 inherited::makeDeepCopyFromShallowCopy(copies); 00139 00140 // ### Call deepCopyField on all "pointer-like" fields 00141 // ### that you wish to be deepCopied rather than 00142 // ### shallow-copied. 00143 // ### ex: 00144 // deepCopyField(trainvec, copies); 00145 00146 // ### If you want to deepCopy a Var field: 00147 // varDeepCopyField(somevariable, copies); 00148 } 00149 00150 void TransposedDoubleProductVariable::declareOptions(OptionList& ol) 00151 { 00152 // ### Declare all of this object's options here. 00153 // ### For the "flags" of each option, you should typically specify 00154 // ### one of OptionBase::buildoption, OptionBase::learntoption or 00155 // ### OptionBase::tuningoption. If you don't provide one of these three, 00156 // ### this option will be ignored when loading values from a script. 00157 // ### You can also combine flags, for example with OptionBase::nosave: 00158 // ### (OptionBase::buildoption | OptionBase::nosave) 00159 00160 // ### ex: 00161 // declareOption(ol, "myoption", &TransposedDoubleProductVariable::myoption, 00162 // OptionBase::buildoption, 00163 // "Help text describing this option"); 00164 // ... 00165 00166 // Now call the parent class' declareOptions 00167 inherited::declareOptions(ol); 00168 } 00169 00170 void TransposedDoubleProductVariable::build_() 00171 { 00172 // ### This method should do the real building of the object, 00173 // ### according to set 'options', in *any* situation. 00174 // ### Typical situations include: 00175 // ### - Initial building of an object from a few user-specified options 00176 // ### - Building of a "reloaded" object: i.e. from the complete set of 00177 // ### all serialised options. 00178 // ### - Updating or "re-building" of an object after a few "tuning" 00179 // ### options have been modified. 00180 // ### You should assume that the parent class' build_() has already been 00181 // ### called. 00182 00183 if (varH().width() != varW().length()*varM().length()) 00184 PLERROR("The width of matrix H is incompatible with the lengths of matrix W and M in TranposedDoubleProductVariable"); 00185 if (varM().width() != varW().width()) 00186 PLERROR("Matrix W and M must have the same width in TranposedDoubleProduct"); 00187 00188 00189 } 00190 00191 00192 } // end of namespace PLearn 00193 00194 00195 /* 00196 Local Variables: 00197 mode:c++ 00198 c-basic-offset:4 00199 c-file-style:"stroustrup" 00200 c-file-offsets:((innamespace . 0)(inline-open . 0)) 00201 indent-tabs-mode:nil 00202 fill-column:79 00203 End: 00204 */ 00205 // vim: filetype=cpp:expandtab:shiftwidth=4:tabstop=8:softtabstop=4:encoding=utf-8:textwidth=79 :