p.238 in NNFS

examples/dense-neural-network/CMakeLists.txt

@@ -0,0 +1,5 @@
+add_executable(dense-neural-network
+    main.cpp
+)
+
+target_link_libraries(dense-neural-network PRIVATE flux)

examples/dense-neural-network/clean.sh

@@ -0,0 +1,13 @@
+#!/bin/bash
+set -e
+
+SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+ROOT_DIR="$(cd "$SCRIPT_DIR/../.." && pwd)"
+BUILD_DIR="$ROOT_DIR/build"
+
+if [ -d "$BUILD_DIR" ]; then
+    echo "Cleaning build directory: $BUILD_DIR"
+    rm -rf "$BUILD_DIR"/*
+else
+    echo "No build directory to clean."
+fi

examples/dense-neural-network/main.cpp

@@ -0,0 +1,149 @@
+#include "./core/omp_config.h"
+
+#include "utils/utils.h"
+#include "numerics/numerics.h"
+#include "decomp/decomp.h"
+
+#include "modules/neural_networks/neural_networks.h"
+
+
+
+//#include <iostream>
+//#include <stdexcept>
+//#include <chrono>
+
+
+
+
+
+int main(int argc, char const *argv[])
+{   
+
+    utils::Mf X(10,2, 0);
+    utils::Matrix<int64_t> y(10,1, 0);
+    utils::Vector<int64_t> class_targets;
+    float loss;
+    float accuracy;
+
+    //neural_networks::create_spital_data<float, uint64_t>(10000, 3, X, y);
+    neural_networks::create_vertical_data<float, int64_t>(100, 3, X, y);
+
+    neural_networks::Dense_Layer<float> dense1(2, 3);
+    neural_networks::Activation_ReLU<float> activation1;
+    neural_networks::Dense_Layer<float> dense2(3, 3);
+    neural_networks::Activation_Softmax<float> activation2;
+    
+    neural_networks::Loss_CategoricalCrossentrophy<float, int64_t> loss_funtion;
+
+    float lowest_loss = 9999999;
+    utils::Mf best_dense_1_weights = dense1.weights;
+    utils::Vf best_dense_1_biases = dense1.biases;
+    utils::Mf best_dense_2_weights = dense2.weights;
+    utils::Vf best_dense_2_biases = dense2.biases;
+
+    utils::Vf vectRND;
+
+    utils::Vector<int64_t> predections;
+
+
+
+
+
+        for (uint64_t i = 0; i < 10; ++i){
+
+
+            // Generate a new set of weights for iteration
+            numerics::inplace_matrandom_mul(dense1.weights,0.98f, 1.02f);
+            numerics::inplace_vecrandom_mul(dense1.biases,0.98f, 1.02f);
+
+            numerics::inplace_matrandom_mul(dense2.weights,0.98f, 1.02f);
+            numerics::inplace_vecrandom_mul(dense2.biases,0.98f, 1.02f);
+
+            // Perform a forward pass of the training data through this layer
+            dense1.forward(X);
+            activation1.forward(dense1.outputs);
+            dense2.forward(activation1.outputs);
+            activation2.forward(dense2.outputs);
+
+            // Perform a farward pass through activation function
+            // it takes the output of the second dense layer here and returns loss
+            loss = loss_funtion.calculate(activation2.outputs, y);
+
+            predections = numerics::matargmax_row<int64_t, float>(activation2.outputs);
+
+            if (y.cols() < 1){
+                class_targets = numerics::matargmax_row<int64_t, int64_t>(y);
+            }else{
+                class_targets = y.get_col(0);
+            }
+
+            accuracy = numerics::vecmean_equal<float>(predections, class_targets);
+
+            if (loss < lowest_loss){
+                //std::cout << "New set of weights found, iteration:" << i << ", loss:" << loss << ", acc:" << accuracy << std::endl;
+                best_dense_1_weights = dense1.weights;
+                best_dense_1_biases = dense1.biases;
+                best_dense_2_weights = dense2.weights;
+                best_dense_2_biases = dense2.biases;
+                lowest_loss = loss;
+            } else{
+                //std::cout << "HERE" << std::endl;
+                dense1.weights = best_dense_1_weights;
+                dense1.biases = best_dense_1_biases;
+                dense2.weights = best_dense_2_weights;
+                dense2.biases = best_dense_2_biases;
+            }
+
+
+        }
+
+
+    //std::cout << loss << std::endl;
+    //std::cout << accuracy << std::endl;
+
+    utils::Matrix<float> softmax_outputs{{0.7, 0.1, 0.2},
+                                {0.1, 0.5, 0.4},
+                                {0.02, 0.9, 0.08}};
+    utils::Matrix<int64_t> clas_targets{{0},{1},{1}};
+
+    neural_networks::Activation_Softmax_Loss_CategoricalCrossentropy<float, int64_t> softmax_loss;
+    softmax_loss.backward(softmax_outputs, clas_targets);
+    utils::Matrix<float> dvalues1 = softmax_loss.dinputs;
+
+    neural_networks::Activation_Softmax<float> activation;
+    activation.outputs = softmax_outputs;
+    
+    //neural_networks::Loss_CategoricalCrossentrophy<float, int64_t> loss;
+
+
+    dvalues1.print();
+
+
+
+
+
+    /*
+    utils::Vd a = utils::linspace<double>(1, 10, 10, true);
+    a.print();
+    mesh::Mesh1D<double> mesh(a);
+    mesh.generate_vertices(0.5, 10.5);
+    double Gamma = 1.0;
+
+    
+    utils::Md A;
+    utils::Vd b, s(10,1);
+
+
+    core::Configs<double>& cfg = core::Configs<double>::defaults();
+    cfg.grid = core::GridKind::Uniform;
+    cfg.left = {core::FDKind::Forward, core::BCKind::Neumann, 0.0};
+    cfg.right = {core::FDKind::Backward, core::BCKind::Neumann, 0.0};
+    cfg.solver    = core::SolverKind::LU;
+
+
+    fluids::Diffusion1D<double> diffusion(cfg, mesh, Gamma);
+    diffusion.assemble(A, b, s);
+*/
+    
+    return 0;
+}

examples/dense-neural-network/run.sh

@@ -0,0 +1,27 @@
+#!/bin/bash
+set -e
+
+TARGET="dense-neural-network"
+
+SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+ROOT_DIR="$(cd "$SCRIPT_DIR/../.." && pwd)"
+BUILD_DIR="$ROOT_DIR/build"
+
+mkdir -p "$BUILD_DIR"
+cd "$BUILD_DIR"
+
+cmake .. #-DCMAKE_BUILD_TYPE=Debug
+cmake --build . --target "$TARGET"
+
+# Load omp.cfg
+if [ -f "$ROOT_DIR/omp.cfg" ]; then
+  export $(grep -v '^[[:space:]]*#' "$ROOT_DIR/omp.cfg" | grep -v '^[[:space:]]*$' | xargs)
+fi
+
+echo "=== CPU / OpenMP info ==="
+echo "System cores (nproc): $(nproc)"
+echo "OMP_NUM_THREADS=${OMP_NUM_THREADS:-"(not set)"}"
+echo "========================="
+
+
+"$BUILD_DIR/bin/$TARGET"