jacobian

commit 9adbe6c93d1dcb5d1a91d11cf7d36943a4dec76c
parent a3114188ddc6ff1e853b331730bb7b3bad85eb0f
Author: David Freifeld <freifeld.david@gmail.com>
Date:   Tue, 14 Jul 2020 20:33:29 -0700

Attempt at using matrix functions reveals they are slow

Diffstat:
M
src/bpnn.cpp
 | 
3
+++
A
tests/activations.cpp
 | 
32
++++++++++++++++++++++++++++++++

2 files changed, 35 insertions(+), 0 deletions(-)
diff --git a/src/bpnn.cpp b/src/bpnn.cpp
@@ -9,6 +9,7 @@
 #include "utils.hpp"
 #include <ctime>
 #include <random>
+#include <Eigen/MatrixFunctions>
 
 #define SHUFFLED_PATH "./shuffled.txt"
 #define TEST_PATH "./test.txt"
@@ -146,6 +147,7 @@ void Network::set_activation(int index, std::function<float(float)> custom, std:
 void Network::feedforward()
 {
   for (int j = 0; j < layers[0].contents->rows(); j++) {
+    if (strcmp(layers[0].activation_str, "linear") == 0) break;
     for (int k = 0; k < layers[0].contents->cols(); k++) {
       (*layers[0].dZ)(j,k) = layers[0].activation_deriv((*layers[0].contents)(j,k));
       (*layers[0].contents)(j,k) = layers[0].activation((*layers[0].contents)(j,k));
@@ -160,6 +162,7 @@ void Network::feedforward()
   }
   for (int i = 1; i < length; i++) {
     for (int j = 0; j < layers[i].contents->rows(); j++) {
+      if (strcmp(layers[i].activation_str, "linear") == 0) break;
       for (int k = 0; k < layers[i].contents->cols(); k++) {
         (*layers[i].dZ)(j,k) = layers[i].activation_deriv((*layers[i].contents)(j,k));
         (*layers[i].contents)(j,k) = layers[i].activation((*layers[i].contents)(j,k));
diff --git a/tests/activations.cpp b/tests/activations.cpp
@@ -0,0 +1,32 @@
+#include <Eigen/Dense>
+#include <Eigen/MatrixFunctions>
+
+#include <iostream>
+
+std::complex<float> lecun_tanh(std::complex<float> x, int) {return (float)1.7159 * tanh(((float)2.0/3) * x);}
+//std::complex<float> lecun_tanh_deriv(std::complex<float> x, int) {return 1.14393 * pow(1.0/cosh(2.0/3 * x),2);}
+
+std::complex<float> mat_sigmoid(std::complex<float> x, int) {return (float)1.0/((float)1+exp(-x));}
+float sigmoid(float x) {return 1.0/(1+exp(-x));}
+
+int main()
+{
+  int size;
+  std::cin >> size;
+  Eigen::MatrixXf m = Eigen::MatrixXf::Random(size,size);
+  auto mat_start = std::chrono::high_resolution_clock::now();
+  m = m.matrixFunction(mat_sigmoid);
+  auto mat_end = std::chrono::high_resolution_clock::now();
+  Eigen::MatrixXf m2 = Eigen::MatrixXf::Random(size,size);
+  auto start = std::chrono::high_resolution_clock::now();
+  for (int j = 0; j < m2.rows(); j++) {
+    for (int k = 0; k < m2.cols(); k++) {
+      m2(j,k) = sigmoid(m2(j,k));
+    }
+  }
+  auto end = std::chrono::high_resolution_clock::now();
+  std::cout << "MATRIX: " << std::chrono::duration_cast<std::chrono::nanoseconds>(mat_end - mat_start).count() << " NORMAL: " << std::chrono::duration_cast<std::chrono::nanoseconds>(end - start).count() << "\n";
+
+}
+
+