Added radialBasisFunc and ReLU for regModel (#619)

* Added the ReLU activation func. * Added the radialBasisFunc for regModel.
mdolab · Apr 6, 2024 · 35e0a4c · 35e0a4c
1 parent 1b56851
commit 35e0a4c
Show file tree

Hide file tree

Showing 2 changed files with 75 additions and 2 deletions.
diff --git a/src/adjoint/DARegression/DARegression.C b/src/adjoint/DARegression/DARegression.C
@@ -49,6 +49,18 @@ DARegression::DARegression(
     {
         regSubDict.readEntry<labelList>("hiddenLayerNeurons", hiddenLayerNeurons_);
         regSubDict.readEntry<word>("activationFunction", activationFunction_);
+        if (activationFunction_ == "ReLU")
+        {
+            leakyCoeff_ = regSubDict.lookupOrDefault<scalar>("leakyCoeff", 0.0);
+        }
+    }
+    else if (modelType_ == "radialBasisFunction")
+    {
+        nRBFs_ = regSubDict.getLabel("nRBFs");
+    }
+    else
+    {
+        FatalErrorIn("") << "modelType_: " << modelType_ << " not supported. Options are: neuralNetwork and radialBasisFunction" << abort(FatalError);
     }
 
     // initialize parameters and give it large values
@@ -323,9 +335,16 @@ label DARegression::compute()
                     {
                         layerVals[layerI][neuronI] = (1 - exp(-2 * layerVals[layerI][neuronI])) / (1 + exp(-2 * layerVals[layerI][neuronI]));
                     }
+                    else if (activationFunction_ == "ReLU")
+                    {
+                        if (layerVals[layerI][neuronI] < 0)
+                        {
+                            layerVals[layerI][neuronI] = leakyCoeff_ * layerVals[layerI][neuronI];
+                        }
+                    }
                     else
                     {
-                        FatalErrorIn("") << "activationFunction not valid. Options are: sigmoid and tanh" << abort(FatalError);
+                        FatalErrorIn("") << "activationFunction not valid. Options are: sigmoid, tanh, and ReLU" << abort(FatalError);
                     }
                 }
             }
@@ -350,9 +369,44 @@ label DARegression::compute()
 
         outputField.correctBoundaryConditions();
     }
+    else if (modelType_ == "radialBasisFunction")
+    {
+        List<List<scalar>> inputFields;
+        inputFields.setSize(inputNames_.size());
+
+        this->calcInput(inputFields);
+
+        label nInputs = inputNames_.size();
+
+        // increment of the parameters for each RBF basis
+        label dP = 2 * nInputs + 1;
+
+        forAll(mesh_.cells(), cellI)
+        {
+            scalar outputVal = 0.0;
+            for (label i = 0; i < nRBFs_; i++)
+            {
+                scalar expCoeff = 0.0;
+                for (label j = 0; j < nInputs; j++)
+                {
+                    scalar A = (inputFields[j][cellI] - parameters_[dP * i + 2 * j]) * (inputFields[j][cellI] - parameters_[dP * i + 2 * j]);
+                    scalar B = 2 * parameters_[dP * i + 2 * j + 1] * parameters_[dP * i + 2 * j + 1];
+                    expCoeff += A / B;
+                }
+                outputVal += parameters_[(dP + 1) * i + dP] * exp(-expCoeff);
+            }
+
+            outputField[cellI] = outputScale_ * (outputVal + outputShift_);
+        }
+
+        // check if the output values are valid otherwise fix/bound them
+        fail = this->checkOutput(outputField);
+
+        outputField.correctBoundaryConditions();
+    }
     else
     {
-        FatalErrorIn("") << "modelType_: " << modelType_ << " not supported. Options are: neuralNetwork" << abort(FatalError);
+        FatalErrorIn("") << "modelType_: " << modelType_ << " not supported. Options are: neuralNetwork and radialBasisFunction" << abort(FatalError);
     }
 
     return fail;
@@ -397,6 +451,19 @@ label DARegression::nParameters()
 
         return nParameters;
     }
+    else if (modelType_ == "radialBasisFunction")
+    {
+        label nInputs = inputNames_.size();
+
+        // each RBF has a weight, nInputs mean, and nInputs std
+        label nParameters = nRBFs_ * (2 * nInputs + 1);
+
+        return nParameters;
+    }
+    else
+    {
+        FatalErrorIn("") << "modelType_: " << modelType_ << " not supported. Options are: neuralNetwork and radialBasisFunction" << abort(FatalError);
+    }
 }
 
 label DARegression::checkOutput(volScalarField& outputField)

diff --git a/src/adjoint/DARegression/DARegression.H b/src/adjoint/DARegression/DARegression.H
@@ -81,6 +81,9 @@ protected:
     /// neural network activation function
     word activationFunction_;
 
+    /// if the ReLU activation function is used we can prescribe a potentially leaky coefficient
+    scalar leakyCoeff_ = 0.0;
+
     /// the upper bound for the output
     scalar outputUpperBound_;
 
@@ -93,6 +96,9 @@ protected:
     /// default output values
     scalar defaultOutputValue_;
 
+    /// number of radial basis function
+    label nRBFs_;
+
 public:
     /// Constructors
     DARegression(