PLearn 0.1
NatGradNNet.h
Go to the documentation of this file.
00001 // -*- C++ -*-
00002 
00003 // NatGradNNet.h
00004 //
00005 // Copyright (C) 2007 Yoshua Bengio
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: Yoshua Bengio
00036 
00040 #ifndef NatGradNNet_INC
00041 #define NatGradNNet_INC
00042 
00043 #include <plearn_learners/generic/PLearner.h>
00044 #include <plearn_learners/generic/GradientCorrector.h>
00045 #include <plearn/sys/Profiler.h>
00046 //#include "CorrelationProfiler.h" // *stat*
00047 
00048 namespace PLearn {
00049 
00053 class NatGradNNet : public PLearner
00054 {
00055     typedef PLearner inherited;
00056 
00057 public:
00058     //#####  Public Build Options  ############################################
00059 
00060     int noutputs;
00061 
00063     TVec<int> hidden_layer_sizes;
00064 
00067     TVec<Mat> layer_params;
00069     TVec<Mat> layer_mparams;
00070 
00072     real params_averaging_coeff;
00074     int params_averaging_freq;
00075 
00077     real init_lrate;
00078 
00080     real lrate_decay;
00081 
00083     real output_layer_L1_penalty_factor;
00084 
00086     real output_layer_lrate_scale;
00087 
00089     int minibatch_size;
00090 
00093     PP<GradientCorrector> neurons_natgrad_template;
00094     TVec<PP<GradientCorrector> > neurons_natgrad_per_layer;
00095 
00098     PP<GradientCorrector> params_natgrad_template;
00106     PP<GradientCorrector> params_natgrad_per_input_template;
00107 
00109     TVec<PP<GradientCorrector> > params_natgrad_per_group;
00110 
00114     PP<GradientCorrector> full_natgrad;
00115 
00117     string output_type;
00118 
00123     real input_size_lrate_normalization_power;
00124 
00129     real lrate_scale_factor;
00130     int lrate_scale_factor_max_power;
00131     int lrate_scale_factor_min_power;
00132 
00136     bool self_adjusted_scaling_and_bias;
00137     real target_mean_activation;
00138     real target_stdev_activation;
00139     // the mean and variance of the activations is estimated by a moving
00140     // average with this coefficient (near 0 for very slow averaging)
00141     real activation_statistics_moving_average_coefficient;
00142 
00143     // *stat*
00144     // Temporary stuff for getting a clue as to what's going on
00145     // Look for the marker '*stat*' in the code
00146 
00147     // -Options-
00149     //int corr_profiling_start, corr_profiling_end;
00150 
00151     // -Not options-
00152     //PP<CorrelationProfiler> g_corrprof, ng_corrprof;    // for optional gradient correlation profiling
00153     //real sum_gradient_norms;     // holds sum of the gradient norms - reset at each epoch
00154     //Vec all_params_cum_gradient; // holds the sum of the gradients - reset at each epoch
00155 
00157     //TVec<VecStatsCollector> pa_gradstats;   // one VecStatsCollector per output class
00158 
00160     //TVec<int> pv_all_nsamples;
00161     //TVec< TMat<int> > pv_layer_nsamples;
00162 
00163     // *stat* end
00164 
00165 public:
00166     //#####  Public Member Functions  #########################################
00167 
00168     NatGradNNet();
00169 
00170     //#####  PLearner Member Functions  #######################################
00171 
00174     // (PLEASE IMPLEMENT IN .cc)
00175     virtual int outputsize() const;
00176 
00180     // (PLEASE IMPLEMENT IN .cc)
00181     virtual void forget();
00182 
00186     // (PLEASE IMPLEMENT IN .cc)
00187     virtual void train();
00188 
00190     // (PLEASE IMPLEMENT IN .cc)
00191     virtual void computeOutput(const Vec& input, Vec& output) const;
00192     virtual void computeOutputs(const Mat& input, Mat& output) const;
00193 
00194     virtual void computeOutputsAndCosts(const Mat& input, const Mat& target, 
00195                                         Mat& output, Mat& costs) const;
00196 
00197 
00199     // (PLEASE IMPLEMENT IN .cc)
00200     virtual void computeCostsFromOutputs(const Vec& input, const Vec& output,
00201                                          const Vec& target, Vec& costs) const;
00202 
00205     // (PLEASE IMPLEMENT IN .cc)
00206     virtual TVec<std::string> getTestCostNames() const;
00207 
00210     // (PLEASE IMPLEMENT IN .cc)
00211     virtual TVec<std::string> getTrainCostNames() const;
00212 
00213 
00214     // *** SUBCLASS WRITING: ***
00215     // While in general not necessary, in case of particular needs
00216     // (efficiency concerns for ex) you may also want to overload
00217     // some of the following methods:
00218     // virtual void computeOutputAndCosts(const Vec& input, const Vec& target,
00219     //                                    Vec& output, Vec& costs) const;
00220     // virtual void computeCostsOnly(const Vec& input, const Vec& target,
00221     //                               Vec& costs) const;
00222     // virtual void test(VMat testset, PP<VecStatsCollector> test_stats,
00223     //                   VMat testoutputs=0, VMat testcosts=0) const;
00224     // virtual int nTestCosts() const;
00225     // virtual int nTrainCosts() const;
00226     // virtual void resetInternalState();
00227     // virtual bool isStatefulLearner() const;
00228 
00229 
00230     //#####  PLearn::Object Protocol  #########################################
00231 
00232     // Declares other standard object methods.
00233     // ### If your class is not instantiatable (it has pure virtual methods)
00234     // ### you should replace this by PLEARN_DECLARE_ABSTRACT_OBJECT_METHODS
00235     PLEARN_DECLARE_OBJECT(NatGradNNet);
00236 
00237     // Simply calls inherited::build() then build_()
00238     virtual void build();
00239 
00241     // (PLEASE IMPLEMENT IN .cc)
00242     virtual void makeDeepCopyFromShallowCopy(CopiesMap& copies);
00243 
00244 protected:
00245     //#####  Protected Options  ###############################################
00246 
00247     // ### Declare protected option fields (such as learned parameters) here
00248 
00250     int n_layers;
00251 
00253     TVec<int> layer_sizes;
00254 
00256     TVec<Mat> biases;
00257     TVec<Mat> weights,mweights;
00258     TVec<Vec> activations_scaling; // output = tanh(activations_scaling[layer][neuron] * (biases[layer][neuron] + weights[layer]*input[layer-1])
00259     TVec<Vec> mean_activations;
00260     TVec<Vec> var_activations;
00261     real cumulative_training_time;
00262 
00263 protected:
00264     //#####  Protected Member Functions  ######################################
00265 
00267     // (PLEASE IMPLEMENT IN .cc)
00268     static void declareOptions(OptionList& ol);
00269 
00271     void onlineStep(int t, const Mat& targets, Mat& train_costs, Vec example_weights);
00272 
00275     void fpropNet(int n_examples, bool during_training) const;
00276 
00279     void fbpropLoss(const Mat& output, const Mat& target, const Vec& example_weights, Mat& train_costs) const;
00280 
00281 
00282 private:
00283     //#####  Private Member Functions  ########################################
00284 
00286     // (PLEASE IMPLEMENT IN .cc)
00287     void build_();
00288 
00289 private:
00290     //#####  Private Data Members  ############################################
00291 
00292     // The rest of the private stuff goes here
00293 
00294     Vec all_params; // all the parameters in one vector
00295     Vec all_params_delta; // update direction
00296     Vec all_params_gradient; // all the parameter gradients in one vector
00297     Vec all_mparams; // mean parameters (moving-averaged over past values)
00298     TVec<Mat> layer_params_gradient;
00299     TVec<Vec> layer_params_delta;
00300     TVec<Vec> group_params; // params of each group (pointing in all_params)
00301     TVec<Vec> group_params_delta; // params_delta of each group (pointing in all_params_delta)
00302     TVec<Vec> group_params_gradient; // params_delta of each group (pointing in all_params_gradient)
00303     Mat neuron_gradients; // one row per example of a minibatch, has concatenation of layer 0, layer 1, ... gradients.
00304     TVec<Mat> neuron_gradients_per_layer; // pointing into neuron_gradients (one row per example of a minibatch)
00305     mutable TVec<Mat> neuron_outputs_per_layer;  // same structure
00306     mutable TVec<Mat> neuron_extended_outputs_per_layer;  // with 1's in the first pseudo-neuron, for doing biases
00307     Mat targets; // one target row per example in a minibatch
00308     Vec example_weights; // one element per example in a minibatch
00309     Mat train_costs; // one row per example in a minibatch
00310     
00311 };
00312 
00313 // Declares a few other classes and functions related to this class
00314 DECLARE_OBJECT_PTR(NatGradNNet);
00315 
00316 } // end of namespace PLearn
00317 
00318 #endif
00319 
00320 
00321 /*
00322   Local Variables:
00323   mode:c++
00324   c-basic-offset:4
00325   c-file-style:"stroustrup"
00326   c-file-offsets:((innamespace . 0)(inline-open . 0))
00327   indent-tabs-mode:nil
00328   fill-column:79
00329   End:
00330 */
00331 // vim: filetype=cpp:expandtab:shiftwidth=4:tabstop=8:softtabstop=4:encoding=utf-8:textwidth=79 :
 All Classes Namespaces Files Functions Variables Typedefs Enumerations Enumerator Friends Defines