From 0c0351e2110bad051825f161fdc301774306c2b4 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?C=C3=A9dric=20Traizet?= <traizetc@cesbio.cnes.fr>
Date: Wed, 24 May 2017 17:07:10 +0200
Subject: [PATCH] added denoising autoencoder (noise strength is now a
 parameter of cbdimensionalityreductiontrainer

---
 include/AutoencoderModel.h     |  4 ++++
 include/AutoencoderModel.txx   | 11 ++++++-----
 include/cbTrainAutoencoder.txx |  7 +++++++
 3 files changed, 17 insertions(+), 5 deletions(-)

diff --git a/include/AutoencoderModel.h b/include/AutoencoderModel.h
index eb65fdb1df..81b55a5fbf 100644
--- a/include/AutoencoderModel.h
+++ b/include/AutoencoderModel.h
@@ -37,6 +37,9 @@ public:
 	itkGetMacro(Regularization,double);
 	itkSetMacro(Regularization,double);
 
+	itkGetMacro(Noise,double);
+	itkSetMacro(Noise,double);
+
 	bool CanReadFile(const std::string & filename);
 	bool CanWriteFile(const std::string & filename);
 
@@ -60,6 +63,7 @@ private:
 	unsigned int m_NumberOfHiddenNeurons;
 	unsigned int m_NumberOfIterations;
 	double m_Regularization;
+	double m_Noise;
 };
 } // end namespace otb
 
diff --git a/include/AutoencoderModel.txx b/include/AutoencoderModel.txx
index dad8324514..5c7bc5d97e 100644
--- a/include/AutoencoderModel.txx
+++ b/include/AutoencoderModel.txx
@@ -11,6 +11,8 @@
 #include <shark/Algorithms/GradientDescent/Rprop.h>// the RProp optimization algorithm
 #include <shark/ObjectiveFunctions/Loss/SquaredLoss.h> // squared loss used for regression
 #include <shark/ObjectiveFunctions/Regularizer.h> //L2 regulariziation
+#include <shark/Models/ImpulseNoiseModel.h>//noise source to corrupt the inputs
+#include <shark/Models/ConcatenatedModel.h>//to concatenate the noise with the model
 
 namespace otb
 {
@@ -32,7 +34,6 @@ AutoencoderModel<TInputValue,AutoencoderType>::~AutoencoderModel()
 template <class TInputValue, class AutoencoderType>
 void AutoencoderModel<TInputValue,AutoencoderType>::Train()
 {
-	
 	std::vector<shark::RealVector> features;
 	
 	Shark::ListSampleToSharkVector(this->GetInputListSample(), features);
@@ -42,10 +43,12 @@ void AutoencoderModel<TInputValue,AutoencoderType>::Train()
 	std::size_t inputs = dataDimension(inputSamples);
 	m_net.setStructure(inputs, m_NumberOfHiddenNeurons);
 	initRandomUniform(m_net,-0.1*std::sqrt(1.0/inputs),0.1*std::sqrt(1.0/inputs));
-	
+	shark::ImpulseNoiseModel noise(m_Noise,0.0);//set an input pixel with probability p to 0
+	shark::ConcatenatedModel<shark::RealVector,shark::RealVector> model = noise>> m_net;
+
 	shark::LabeledData<shark::RealVector,shark::RealVector> trainSet(inputSamples,inputSamples);//labels identical to inputs
 	shark::SquaredLoss<shark::RealVector> loss;
-	shark::ErrorFunction error(trainSet, &m_net, &loss);
+	shark::ErrorFunction error(trainSet, &model, &loss);
 	shark::TwoNormRegularizer regularizer(error.numberOfVariables());
 	error.setRegularizer(m_Regularization,&regularizer);
 
@@ -117,7 +120,6 @@ template <class TInputValue, class AutoencoderType>
 typename AutoencoderModel<TInputValue,AutoencoderType>::TargetSampleType
 AutoencoderModel<TInputValue,AutoencoderType>::DoPredict(const InputSampleType & value) const
 {  
-	std::cout << "SINGLE PIXEL " ;
 	shark::RealVector samples(value.Size());
 	for(size_t i = 0; i < value.Size();i++)
     {
@@ -144,7 +146,6 @@ template <class TInputValue, class AutoencoderType>
 void AutoencoderModel<TInputValue,AutoencoderType>
 ::DoPredictBatch(const InputListSampleType *input, const unsigned int & startIndex, const unsigned int & size, TargetListSampleType * targets) const
 {
-	std::cout << "BATCH" << std::endl;
 	std::vector<shark::RealVector> features;
 	Shark::ListSampleRangeToSharkVector(input, features,startIndex,size);
 	shark::Data<shark::RealVector> data = shark::createDataFromRange(features);
diff --git a/include/cbTrainAutoencoder.txx b/include/cbTrainAutoencoder.txx
index 82e21d17b6..e120ae5e9b 100644
--- a/include/cbTrainAutoencoder.txx
+++ b/include/cbTrainAutoencoder.txx
@@ -58,6 +58,12 @@ cbLearningApplicationBaseDR<TInputValue,TOutputValue>
   SetParameterFloat("model.autoencoder.regularization",0, false);
   SetParameterDescription("model.autoencoder.regularization", 
                          "Strength of the L2 regularization used during training");
+                         
+  //Noise strength
+  AddParameter(ParameterType_Float, "model.autoencoder.noise", "Strength of the noise");
+  SetParameterFloat("model.autoencoder.noise",0, false);
+  SetParameterDescription("model.autoencoder.noise", 
+                         "Strength of the noise");
 }
 
 
@@ -97,6 +103,7 @@ void cbLearningApplicationBaseDR<TInputValue,TOutputValue>
 		dimredTrainer->SetNumberOfHiddenNeurons(GetParameterInt("model.autoencoder.nbneuron"));
 		dimredTrainer->SetNumberOfIterations(GetParameterInt("model.autoencoder.nbiter"));
 		dimredTrainer->SetRegularization(GetParameterFloat("model.autoencoder.regularization"));
+		dimredTrainer->SetRegularization(GetParameterFloat("model.autoencoder.noise"));
 		dimredTrainer->SetInputListSample(trainingListSample);
 		dimredTrainer->Train();
 		dimredTrainer->Save(modelPath);
-- 
GitLab