Bausager/Flux
1
0
Fork 0
Code Issues 36 Pull Requests Actions Packages Projects 1 Releases Wiki Activity

Fittet new functions to everying in neural networks. Still need to optimise for uint64_t vs int64_t and vec vs mat in some places.

Browse Source
This commit is contained in:
Michelle
2026-05-16 20:37:05 +02:00
parent 412a854c65
commit d2fe8aa65c
50 changed files with 489 additions and 1482 deletions
Download Patch File Download Diff File Expand all files Collapse all files
include/modules/neural_networks/loss/Loss.h
+8 -9
View File
@@ -1,13 +1,12 @@
#pragma once
#include "./core/omp_config.h"
#include "core/omp_config.h"
#include "./utils/vector.h"
#include "./utils/matrix.h"
#include "utils/vector.h"
#include "utils/matrix.h"
#include "numerics/vecmean.h"
#include "numerics/matabs.h"
#include "numerics/matmean.h"
#include "numerics/mean.h"
#include "numerics/abs.h"
namespace neural_networks{
@@ -28,7 +27,7 @@ namespace neural_networks{
sample_losses = forward(output, y);
// Calculate mean loss
data_loss = numerics::vecmean(sample_losses);
data_loss = numerics::mean(sample_losses);
return data_loss;
@@ -42,12 +41,12 @@ namespace neural_networks{
// L1 regularization - weights
// calculate only when factor greater than 0
if (layer.weight_regularizer_l1){
regularization_losss += layer.weight_regularizer_l1 * numerics::matsum_coeff(numerics::matabs(layer.weights));
regularization_losss += layer.weight_regularizer_l1 * numerics::sum(numerics::abs(layer.weights));
}
// L2 regularization - weights
if (layer.weight_regularizer_l2){
regularization_losss += layer.weight_regularizer_l2 * numerics::matsum_coeff(numerics::matmul(layer.weights,layer.weights)); // elementwise!
regularization_losss += layer.weight_regularizer_l2 * numerics::sum(numerics::mul(layer.weights,layer.weights)); // elementwise!
}
// L1 regularization - biases
include/modules/neural_networks/loss/Loss_CategoricalCrossentrophy.h
+16 -12
View File
@@ -1,15 +1,15 @@
#pragma once
#include "./core/omp_config.h"
#include "core/omp_config.h"
#include "./utils/vector.h"
#include "./utils/matrix.h"
#include "./utils/matcast.h"
#include "utils/vector.h"
#include "utils/matrix.h"
#include "utils/matcast.h"
#include "./numerics/matclip.h"
#include "./numerics/veclog.h"
#include "numerics/clip.h"
#include "numerics/log.h"
#include "./Loss.h"
#include "Loss.h"
namespace neural_networks{
@@ -30,7 +30,7 @@ namespace neural_networks{
// Clip data to prevent dividning by 0
// Clip both sides to not drag mean towards any value
utils::Matrix<Td> y_pred_clipped = numerics::matclip(y_pred, Td{1e-7}, Td{1.0} - Td{1e-7});
utils::Matrix<Td> y_pred_clipped = numerics::clip(y_pred, Td{1e-7}, Td{1.0} - Td{1e-7});
// Probabilities for taget values
// only if categorical labes
@@ -40,7 +40,8 @@ namespace neural_networks{
correct_confidences[i] = y_pred_clipped(i, idx);
}
}else{ // Mask values - only for one-hot encoded labels
correct_confidences = numerics::matdot_row(y_pred_clipped, cast_y_true);
correct_confidences = numerics::sum_rowwise(numerics::mul(y_pred_clipped, cast_y_true));
//correct_confidences = numerics::matdot_row(y_pred_clipped, cast_y_true);
}
// Losses
@@ -72,9 +73,12 @@ namespace neural_networks{
// Calculate the gradient
numerics::inplace_matscalar(y_temp,Ti{-1});
dinputs = numerics::matdiv(utils::matcast<Td, Ti>(y_temp), dvalues);
numerics::inplace_matdiv(dinputs, samples);
//numerics::inplace_matscalar(y_temp,Ti{-1});
y_temp = numerics::neg(y_temp);
//dinputs = numerics::matdiv(utils::matcast<Td, Ti>(y_temp), dvalues);
dinputs = numerics::div(utils::matcast<Td, Ti>(y_temp), dvalues);
//numerics::inplace_matdiv(dinputs, samples);
dinputs = numerics::div(dinputs, samples);
}