PLearn 0.1
|
00001 // -*- C++ -*- 00002 00003 // RBMParameters.cc 00004 // 00005 // Copyright (C) 2006 Pascal Lamblin 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 Lamblin 00036 00041 #include "RBMParameters.h" 00042 #include <plearn/math/TMat_maths.h> 00043 #include <plearn/base/stringutils.h> 00044 //#include "RBMLayer.h" 00045 00046 namespace PLearn { 00047 using namespace std; 00048 00049 PLEARN_IMPLEMENT_ABSTRACT_OBJECT( 00050 RBMParameters, 00051 "Virtual class for the parameters between two layers of an RBM", 00052 ""); 00053 00054 RBMParameters::RBMParameters( real the_learning_rate ) : 00055 learning_rate(the_learning_rate), 00056 going_up(true), 00057 pos_count(0), 00058 neg_count(0) 00059 { 00060 } 00061 00062 RBMParameters::RBMParameters( string down_types, string up_types, 00063 real the_learning_rate ) : 00064 up_units_types(up_types), 00065 down_units_types(down_types), 00066 learning_rate(the_learning_rate), 00067 going_up(true), 00068 pos_count(0), 00069 neg_count(0) 00070 { 00071 // We're not sure inherited::build() has been called 00072 build(); 00073 } 00074 00075 void RBMParameters::declareOptions(OptionList& ol) 00076 { 00077 // ### Declare all of this object's options here. 00078 // ### For the "flags" of each option, you should typically specify 00079 // ### one of OptionBase::buildoption, OptionBase::learntoption or 00080 // ### OptionBase::tuningoption. If you don't provide one of these three, 00081 // ### this option will be ignored when loading values from a script. 00082 // ### You can also combine flags, for example with OptionBase::nosave: 00083 // ### (OptionBase::buildoption | OptionBase::nosave) 00084 00085 declareOption(ol, "up_units_types", &RBMParameters::up_units_types, 00086 OptionBase::buildoption, 00087 "Each character of this string describes the type of an" 00088 " up unit:\n" 00089 " - 'l' if the energy function of this unit is linear\n" 00090 " (binomial or multinomial unit),\n" 00091 " - 'q' if it is quadratic (for a gaussian unit).\n"); 00092 00093 declareOption(ol, "down_units_types", &RBMParameters::down_units_types, 00094 OptionBase::buildoption, 00095 "Same meaning as 'up_units_types', but with down units"); 00096 00097 declareOption(ol, "learning_rate", &RBMParameters::learning_rate, 00098 OptionBase::buildoption, 00099 "The learning rate, used both in update() and bpropUpdate() " 00100 "methods\n"); 00101 00102 declareOption(ol, "initialization_method", 00103 &RBMParameters::initialization_method, 00104 OptionBase::buildoption, 00105 "The method used to initialize the weights:\n" 00106 " - \"uniform_linear\" = a uniform law in [-1/d, 1/d]\n" 00107 " - \"uniform_sqrt\" = a uniform law in [-1/sqrt(d)," 00108 " 1/sqrt(d)]\n" 00109 " - \"zero\" = all weights are set to 0,\n" 00110 "where d = max( up_layer_size, down_layer_size ).\n"); 00111 00112 // Now call the parent class' declareOptions 00113 inherited::declareOptions(ol); 00114 } 00115 00116 void RBMParameters::build_() 00117 { 00118 up_layer_size = int(up_units_types.size()); 00119 down_layer_size = int(down_units_types.size()); 00120 if( up_layer_size == 0 || down_layer_size == 0 ) 00121 return; 00122 00123 string im = lowerstring( initialization_method ); 00124 if( im == "" || im == "uniform_sqrt" ) 00125 initialization_method = "uniform_sqrt"; 00126 else if( im == "uniform_linear" ) 00127 initialization_method = im; 00128 else if( im == "zero" ) 00129 initialization_method = im; 00130 else 00131 PLERROR( "RBMParameters::build_ - initialization_method\n" 00132 "\"%s\" unknown.\n", initialization_method.c_str() ); 00133 00134 input_size = down_layer_size; 00135 output_size = up_layer_size; 00136 } 00137 00138 void RBMParameters::build() 00139 { 00140 inherited::build(); 00141 build_(); 00142 } 00143 00144 00145 void RBMParameters::makeDeepCopyFromShallowCopy(CopiesMap& copies) 00146 { 00147 inherited::makeDeepCopyFromShallowCopy(copies); 00148 00149 // ### Call deepCopyField on all "pointer-like" fields 00150 // ### that you wish to be deepCopied rather than 00151 // ### shallow-copied. 00152 00153 deepCopyField(input_vec, copies); 00154 } 00155 00156 void RBMParameters::setAsUpInput( const Vec& input ) const 00157 { 00158 PLASSERT( input.size() == up_layer_size ); 00159 input_vec = input; 00160 going_up = false; 00161 } 00162 00163 void RBMParameters::setAsDownInput( const Vec& input ) const 00164 { 00165 PLASSERT( input.size() == down_layer_size ); 00166 input_vec = input; 00167 going_up = true; 00168 } 00169 00170 void RBMParameters::update( const Vec& pos_down_values, 00171 const Vec& pos_up_values, 00172 const Vec& neg_down_values, 00173 const Vec& neg_up_values) 00174 { 00175 // Not-so-efficient implementation 00176 accumulatePosStats( pos_down_values, pos_up_values ); 00177 accumulateNegStats( neg_down_values, neg_up_values ); 00178 update(); 00179 } 00180 00182 void RBMParameters::fprop(const Vec& input, Vec& output) const 00183 { 00184 // propagates the activations. 00185 setAsDownInput( input ); 00186 output.resize( output_size ); 00187 computeUnitActivations( 0, output_size, output ); 00188 } 00189 00190 } // end of namespace PLearn 00191 00192 00193 /* 00194 Local Variables: 00195 mode:c++ 00196 c-basic-offset:4 00197 c-file-style:"stroustrup" 00198 c-file-offsets:((innamespace . 0)(inline-open . 0)) 00199 indent-tabs-mode:nil 00200 fill-column:79 00201 End: 00202 */ 00203 // vim: filetype=cpp:expandtab:shiftwidth=4:tabstop=8:softtabstop=4:encoding=utf-8:textwidth=79 :