SmartGuitarAmp

Guitar plugin made with JUCE that uses neural networks to emulate a tube amplifier
Log | Files | Refs | Submodules | README

ConvolutionLayer.cpp (2763B)

      1 /*
      2   ==============================================================================
      3 
      4     ConvolutionLayer.cpp
      5     Created: 10 Jan 2019 5:04:39pm
      6     Author:  Damskägg Eero-Pekka
      7 
      8   ==============================================================================
      9 */
     10 
     11 #include "ConvolutionLayer.h"
     12 
     13 ConvolutionLayer::ConvolutionLayer(size_t inputChannels,
     14                                    size_t outputChannels,
     15                                    int filterWidth,
     16                                    int dilation,
     17                                    bool residual,
     18                                    std::string activationName):
     19 conv(inputChannels,
     20      Activations::isGated(activationName) ? outputChannels * 2 : outputChannels,
     21      filterWidth,
     22      dilation),
     23 out1x1(outputChannels, outputChannels, 1, 1),
     24 residual(residual),
     25 usesGating(Activations::isGated(activationName)),
     26 activation(Activations::getActivationFuncArray(activationName))
     27 {
     28 }
     29 
     30 void ConvolutionLayer::process(float* data, int numSamples) 
     31 {
     32     conv.process(data, numSamples);
     33     activation(data, conv.getNumOutputChannels(), numSamples);
     34     if (residual) {
     35         out1x1.process(data, numSamples);
     36     }
     37 }
     38 
     39 void ConvolutionLayer::process(float* data, float* skipData, int numSamples) 
     40 {
     41     conv.process(data, numSamples); 
     42     activation(data, conv.getNumOutputChannels(), numSamples); 
     43     copySkipData(data, skipData, numSamples); 
     44     if (residual) {
     45         out1x1.process(data, numSamples); 
     46     }
     47 }
     48 
     49 void ConvolutionLayer::copySkipData(float *data, float *skipData, int numSamples)
     50 {
     51     size_t skipChannels = usesGating ? conv.getNumOutputChannels()/2 : conv.getNumOutputChannels();
     52     for (size_t i = 0; i < (size_t)numSamples*skipChannels; ++i)
     53         skipData[i] = data[i];
     54 }
     55 
     56 void ConvolutionLayer::setParams(size_t newInputChannels, size_t newOutputChannels,
     57                                  int newFilterWidth, int newDilation, bool newResidual,
     58                                  std::string newActivationName)
     59 {
     60     activation = Activations::getActivationFuncArray(newActivationName);
     61     usesGating = Activations::isGated(newActivationName);
     62     size_t internalChannels = usesGating ? newOutputChannels * 2 : newOutputChannels;
     63     conv.setParams(newInputChannels, internalChannels, newFilterWidth, newDilation);
     64     out1x1.setParams(newOutputChannels, newOutputChannels, 1, 1);
     65     residual = newResidual;
     66 }
     67 
     68 void ConvolutionLayer::setWeight(std::vector<float> W, std::string name)
     69 {
     70     if ((name == "W_conv") || (name == "W"))
     71         conv.setWeight(W, "W");
     72     else if ((name == "b_conv") || (name == "b"))
     73         conv.setWeight(W, "b");
     74     else if (name == "W_out")
     75         out1x1.setWeight(W, "W");
     76     else if (name == "b_out")
     77         out1x1.setWeight(W, "b");
     78 }