commit 0fa05461cd89f8045ec7c501a3f4924b17c423a5
parent 4f0fd9d3d0c991da6b1d2620cf93b060aa76c492
Author: Jatin Chowdhury <[email protected]>
Date: Thu, 16 Dec 2021 21:49:53 +0000
Remove stale code from STNModel.h
Diffstat:
2 files changed, 13 insertions(+), 317 deletions(-)
diff --git a/Plugin/Source/Processors/Hysteresis/STNModel.cpp b/Plugin/Source/Processors/Hysteresis/STNModel.cpp
@@ -2,147 +2,28 @@
namespace STNSpace
{
-Dense54::Dense54()
-{
- for (size_t i = 0; i < 3; ++i)
- ins[i] = v_type (0.0);
-
- for (size_t i = 0; i < 16; ++i)
- weights[i] = v_type (0.0);
-
- for (size_t i = 0; i < 2; ++i)
- bias[i] = v_type (0.0);
-}
-
-void Dense54::setWeights (std::vector<std::vector<double>>& w)
-{
- for (size_t i = 0; i < out_size; ++i)
- for (size_t j = 0; j < in_size; ++j)
- weights[i * out_size + j / 2].set (j % 2, w[i][j]);
-}
-
-void Dense54::setBias (std::vector<double>& b)
-{
- for (size_t i = 0; i < out_size; ++i)
- bias[i / 2].set (i % 2, b[i]);
-}
-
-//===========================================================
-Dense44::Dense44()
-{
- for (size_t i = 0; i < 16; ++i)
- weights[i] = v_type (0.0);
-
- for (size_t i = 0; i < 2; ++i)
- bias[i] = v_type (0.0);
-}
-
-void Dense44::setWeights (std::vector<std::vector<double>>& w)
-{
- for (size_t i = 0; i < out_size; ++i)
- for (size_t j = 0; j < in_size; ++j)
- weights[i * out_size + j / 2].set (j % 2, w[i][j]);
-}
-
-void Dense44::setBias (std::vector<double>& b)
-{
- for (size_t i = 0; i < out_size; ++i)
- bias[i / 2].set (i % 2, b[i]);
-}
-
-//===========================================================
-Dense41::Dense41()
-{
- for (size_t i = 0; i < 2; ++i)
- weights[i] = v_type (0.0);
-
- bias = 0.0;
-}
+static bool printed = false;
-void Dense41::setWeights (std::vector<std::vector<double>>& w)
+STNModel::STNModel()
{
- for (size_t i = 0; i < out_size; ++i)
- for (size_t j = 0; j < in_size; ++j)
- weights[j / 2].set (j % 2, w[i][j]);
-}
-
-void Dense41::setBias (std::vector<double>& b)
-{
- bias = b[0];
+ // useful for debugging...
+ if (! printed)
+ {
+#if USE_RTNEURAL_STATIC
+ DBG ("Using RTNeural ModelT STN");
+#elif USE_RTNEURAL_POLY
+ DBG ("Using RTNeural polymorphic STN");
+#endif
+ printed = true;
+ }
}
-//===========================================================
void STNModel::loadModel (const nlohmann::json& modelJ)
{
#if USE_RTNEURAL_STATIC
model.parseJson (modelJ);
#elif USE_RTNEURAL_POLY
model = RTNeural::json_parser::parseJson<double> (modelJ);
-#else
- auto layers = modelJ["layers"];
-
- const auto weights_l0 = layers.at (0)["weights"];
- {
- // load weights
- std::vector<std::vector<double>> denseWeights (4);
- for (auto& w : denseWeights)
- w.resize (5, 0.0);
-
- auto layerWeights = weights_l0[0];
- for (size_t i = 0; i < layerWeights.size(); ++i)
- {
- auto lw = layerWeights[i];
- for (size_t j = 0; j < lw.size(); ++j)
- denseWeights[j][i] = lw[j].get<double>();
- }
- dense54.setWeights (denseWeights);
-
- // load biases
- std::vector<double> denseBias = weights_l0[1].get<std::vector<double>>();
- dense54.setBias (denseBias);
- }
-
- const auto weights_l1 = layers.at (1)["weights"];
- {
- // load weights
- std::vector<std::vector<double>> denseWeights (4);
- for (auto& w : denseWeights)
- w.resize (4, 0.0);
-
- auto layerWeights = weights_l1[0];
- for (size_t i = 0; i < layerWeights.size(); ++i)
- {
- auto lw = layerWeights[i];
- for (size_t j = 0; j < lw.size(); ++j)
- denseWeights[j][i] = lw[j].get<double>();
- }
- dense44.setWeights (denseWeights);
-
- // load biases
- std::vector<double> denseBias = weights_l1[1].get<std::vector<double>>();
- dense44.setBias (denseBias);
- }
-
- const auto weights_l2 = layers.at (2)["weights"];
- {
- // load weights
- std::vector<std::vector<double>> denseWeights (1);
- for (auto& w : denseWeights)
- w.resize (4, 0.0);
-
- auto layerWeights = weights_l2[0];
- for (size_t i = 0; i < layerWeights.size(); ++i)
- {
- auto lw = layerWeights[i];
- for (size_t j = 0; j < lw.size(); ++j)
- denseWeights[j][i] = lw[j].get<double>();
- }
- dense41.setWeights (denseWeights);
-
- // load biases
- std::vector<double> denseBias = weights_l2[1].get<std::vector<double>>();
- dense41.setBias (denseBias);
- }
#endif
}
diff --git a/Plugin/Source/Processors/Hysteresis/STNModel.h b/Plugin/Source/Processors/Hysteresis/STNModel.h
@@ -3,188 +3,15 @@
#include <JuceHeader.h>
#include <RTNeural/RTNeural.h>
-// include <Accelerate> on Apple devices so we can use vvtanh
-#if JUCE_MAC || JUCE_IOS
-#define Point CarbonDummyPointName
-#define Component CarbonDummyCompName
-#include <Accelerate/Accelerate.h>
-#undef Point
-#undef Component
-#endif
-
#define USE_RTNEURAL_POLY 0
#define USE_RTNEURAL_STATIC 1
namespace STNSpace
{
-using v_type = dsp::SIMDRegister<double>;
-constexpr auto v_size = v_type::SIMDNumElements;
-static_assert (v_size == 2, "SIMD double size is required to be 2.");
-
-class Dense54
-{
-public:
- Dense54();
-
- void setWeights (std::vector<std::vector<double>>& w);
- void setBias (std::vector<double>& b);
-
- inline void forward (const double* input) noexcept
- {
- ins[0] = v_type::fromRawArray (input);
- ins[1] = v_type::fromRawArray (input + v_size);
- ins[2] = v_type::fromRawArray (input + 2 * v_size);
-
- outs[0] = v_type (0.0);
- outs[1] = v_type (0.0);
- for (size_t k = 0; k < 3; ++k)
- {
- // output 0
- outs[0].set (0, outs[0].get (0) + (ins[k] * weights[0 * out_size + k]).sum());
- // output 1
- outs[0].set (1, outs[0].get (1) + (ins[k] * weights[1 * out_size + k]).sum());
- // output 2
- outs[1].set (0, outs[1].get (0) + (ins[k] * weights[2 * out_size + k]).sum());
- // output 3
- outs[1].set (1, outs[1].get (1) + (ins[k] * weights[3 * out_size + k]).sum());
- }
-
- outs[0] += bias[0];
- outs[1] += bias[1];
- }
-
- v_type outs[2];
-
-private:
- static constexpr size_t in_size = 5;
- static constexpr size_t out_size = 4;
-
- v_type ins[3];
- v_type bias[2];
- v_type weights[16];
-};
-
-//===========================================================
-class Dense44
-{
-public:
- Dense44();
-
- void setWeights (std::vector<std::vector<double>>& w);
- void setBias (std::vector<double>& b);
-
- inline void forward (const v_type* ins) noexcept
- {
- outs[0] = v_type (0.0);
- outs[1] = v_type (0.0);
- for (size_t k = 0; k < 2; ++k)
- {
- // output 0
- outs[0].set (0, outs[0].get (0) + (ins[k] * weights[0 * out_size + k]).sum());
- // output 1
- outs[0].set (1, outs[0].get (1) + (ins[k] * weights[1 * out_size + k]).sum());
- // output 2
- outs[1].set (0, outs[1].get (0) + (ins[k] * weights[2 * out_size + k]).sum());
- // output 3
- outs[1].set (1, outs[1].get (1) + (ins[k] * weights[3 * out_size + k]).sum());
- }
-
- outs[0] += bias[0];
- outs[1] += bias[1];
- }
-
- v_type outs[2];
-
-private:
- static constexpr size_t in_size = 4;
- static constexpr size_t out_size = 4;
-
- v_type bias[2];
- v_type weights[16];
-};
-
-//===========================================================
-class Dense41
-{
-public:
- Dense41();
-
- void setWeights (std::vector<std::vector<double>>& w);
- void setBias (std::vector<double>& b);
-
- inline double forward (const v_type* ins) const noexcept
- {
- double out = 0.0;
- for (size_t k = 0; k < 2; ++k)
- out += (ins[k] * weights[k]).sum();
-
- return out + bias;
- }
-
-private:
- static constexpr size_t in_size = 4;
- static constexpr size_t out_size = 1;
-
- double bias;
- v_type weights[2];
-};
-
-//===========================================================
-class Tanh
-{
-public:
- Tanh() = default;
-
- inline void forward (const v_type* input) noexcept
- {
-#if defined(_M_ARM64) || defined(__arm64__) || defined(__aarch64__)
- alignas (16) double x[4];
- input[0].copyToRawArray (x);
- input[1].copyToRawArray (&x[2]);
-
- vvtanh (x, x, &size);
-
- outs[0] = v_type::fromRawArray (x);
- outs[1] = v_type::fromRawArray (x + 2);
-#elif USE_XSIMD
- using x_type = xsimd::simd_type<double>;
- outs[0] = v_type (xsimd::tanh (static_cast<x_type> (input[0].value)));
- outs[1] = v_type (xsimd::tanh (static_cast<x_type> (input[1].value)));
-#else
- // fallback
- outs[0].set (0, std::tanh (input[0].get (0)));
- outs[0].set (1, std::tanh (input[0].get (1)));
- outs[1].set (0, std::tanh (input[1].get (0)));
- outs[1].set (1, std::tanh (input[1].get (1)));
-#endif
- }
-
- v_type outs[2];
-
-private:
- static constexpr int size = 4;
-};
-
-static bool printed = false;
-
class STNModel
{
public:
- STNModel()
- {
- // useful for debugging...
- if (! printed)
- {
-#if USE_RTNEURAL_STATIC
- DBG ("Using RTNeural ModelT STN");
-#elif USE_RTNEURAL_POLY
- DBG ("Using RTNeural polymorphic STN");
-#else
- DBG ("Using hand-coded STN");
-#endif
- printed = true;
- }
- }
+ STNModel();
inline double forward (const double* input) noexcept
{
@@ -192,12 +19,6 @@ public:
return model.forward (input);
#elif USE_RTNEURAL_POLY
return model->forward (input);
-#else
- dense54.forward (input);
- tanh1.forward (dense54.outs);
- dense44.forward (tanh1.outs);
- tanh2.forward (dense44.outs);
- return dense41.forward (tanh2.outs);
#endif
}
@@ -208,12 +29,6 @@ private:
RTNeural::ModelT<double, 5, 1, RTNeural::DenseT<double, 5, 4>, RTNeural::TanhActivationT<double, 4>, RTNeural::DenseT<double, 4, 4>, RTNeural::TanhActivationT<double, 4>, RTNeural::DenseT<double, 4, 1>> model;
#elif USE_RTNEURAL_POLY
std::unique_ptr<RTNeural::Model<double>> model;
-#else
- Dense54 dense54;
- Tanh tanh1;
- Dense44 dense44;
- Tanh tanh2;
- Dense41 dense41;
#endif
JUCE_DECLARE_NON_COPYABLE_WITH_LEAK_DETECTOR (STNModel)
