From c930da157610d78d81d355e032563dc69b439c2d Mon Sep 17 00:00:00 2001
From: Martin Beseda <martin.beseda@vsb.cz>
Date: Tue, 22 Jan 2019 15:06:16 +0100
Subject: [PATCH] FIX: MSE is now evaluated againts de-normalized data
---
src/ErrorFunction/ErrorFunctions.cpp | 59 ++++++++++++++++++++++------
1 file changed, 48 insertions(+), 11 deletions(-)
diff --git a/src/ErrorFunction/ErrorFunctions.cpp b/src/ErrorFunction/ErrorFunctions.cpp
index 89f5cde2..e4b386d2 100644
--- a/src/ErrorFunction/ErrorFunctions.cpp
+++ b/src/ErrorFunction/ErrorFunctions.cpp
@@ -111,6 +111,7 @@ namespace lib4neuro {
size_t n_elements = data->size();
//TODO instead use something smarter
+ std::vector<std::vector<double>> outputs(data->size());
std::vector<double> output(dim_out);
COUT_DEBUG("Evaluation of the error function MSE on the given data-set (format 'data-set element index' 'input'"
@@ -127,8 +128,16 @@ namespace lib4neuro {
output,
weights);
- /* Compute difference for every element of the output vector */
+ outputs.at(i) = output;
+ }
+
+ if(data_set->is_normalized()) {
+ data_set->de_normalize();
+ }
+
+ for (auto i = 0; i < data->size(); i++) {
+ /* Compute difference for every element of the output vector */
#ifdef L4N_DEBUG
std::stringstream ss_input;
for(auto j = 0; j < dim_in-1; j++) {
@@ -139,16 +148,18 @@ namespace lib4neuro {
std::stringstream ss_real_output;
std::stringstream ss_predicted_output;
#endif
+ double denormalized_output;
for (size_t j = 0; j < dim_out; ++j) {
+ denormalized_output = data_set->get_normalization_strategy()->de_normalize(outputs.at(i).at(j));
#ifdef L4N_DEBUG
ss_real_output << data->at(i).second.at(j);
- ss_predicted_output << output.at(j);
+ ss_predicted_output << denormalized_output;
#endif
- val = output.at(j) - data->at(i).second.at(j);
+ val = denormalized_output - data->at(i).second.at(j);
error += val * val;
- output_norm += output.at(j) * output.at(j);
+ output_norm += denormalized_output * denormalized_output;
}
#ifdef L4N_DEBUG
COUT_DEBUG(i << ": "
@@ -184,6 +195,7 @@ namespace lib4neuro {
std::vector<std::pair<std::vector<double>, std::vector<double>>>* data = data_set->get_data();
//TODO instead use something smarter
+ std::vector<std::vector<double>> outputs(data->size());
std::vector<double> output(dim_out);
COUT_DEBUG(
@@ -204,23 +216,32 @@ namespace lib4neuro {
output,
weights);
+ outputs.at(i) = output;
+ }
+
+ if(data_set->is_normalized()) {
+ data_set->de_normalize();
+ }
+
+ for(auto i = 0; i < data->size(); i++) {
/* Compute difference for every element of the output vector */
+ double denormalized_output;
for (size_t j = 0; j < dim_out; ++j) {
+ denormalized_output = data_set->get_normalization_strategy()->de_normalize(outputs.at(i).at(j));
COUT_DEBUG("Element " << i << ": "
<< data->at(i).first.at(j) << " "
<< data->at(i).second.at(j) << " "
- << output.at(j) << std::endl);
+ << denormalized_output << std::endl);
ofs << data->at(i).first.at(j) << " "
<< data->at(i).second.at(j) << " "
- << output.at(j) << std::endl;
+ << denormalized_output << std::endl;
- val = output.at(j) - data->at(i).second.at(j);
+ val = denormalized_output - data->at(i).second.at(j);
error += val * val;
}
ofs << std::endl;
-
}
ofs.close();
@@ -236,12 +257,14 @@ namespace lib4neuro {
std::vector<std::pair<std::vector<double>, std::vector<double>>>* data = data_set->get_data();
//TODO instead use something smarter
+ std::vector<std::vector<double>> outputs(data->size());
std::vector<double> output(dim_out);
COUT_DEBUG(
"Evaluation of the error function MSE on the given data-set (format 'input' 'real output' 'predicted output'):"
<< std::endl);
+ /* Compute predicted outputs */
for (auto i = 0; i < data->size(); i++) { // Iterate through every element in the test set
/* Compute the net output and store it into 'output' variable */
@@ -249,17 +272,31 @@ namespace lib4neuro {
output,
weights);
+ outputs.at(i) = output;
+ }
+
+ /* De-normalize data-set, if it's normalized */
+ if(data_set->is_normalized()) {
+ data_set->de_normalize();
+ }
+
+ /* Evaluate the prediction error on de-normalized data */
+ for(auto i = 0; i < data->size(); i++) {
+
/* Compute difference for every element of the output vector */
- for (size_t j = 0; j < dim_out; ++j) {
+ double denormalized_output;
+ for (auto j = 0; j < dim_out; ++j) {
+ denormalized_output = data_set->get_normalization_strategy()->de_normalize(outputs.at(i).at(j));
COUT_DEBUG("Element " << i << ": "
<< data->at(i).first.at(j) << " "
<< data->at(i).second.at(j) << " "
- << output.at(j) << std::endl);
+ << denormalized_output << std::endl);
- val = output.at(j) - data->at(i).second.at(j);
+ val = denormalized_output - data->at(i).second.at(j);
error += val * val;
}
}
+
return error / n_elements;
}
--
GitLab