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

Starting on the model.h, but need to make layer structs and structs for loss and optimizers

Browse Source
This commit is contained in:
Michelle
2026-05-30 09:13:52 +02:00
parent cb65174cf4
commit edad247227
30 changed files with 1879 additions and 159 deletions
Download Patch File Download Diff File Expand all files Collapse all files
examples/dense-neural-network/main.cpp
+81 -55
View File
@@ -21,93 +21,107 @@ int main(int argc, char const *argv[])
{
uint64_t number_of_classes = 2;
uint64_t number_of_samples = 150;
uint64_t number_of_epochs = 1000;
uint64_t number_of_samples = 1000;
uint64_t number_of_epochs = 10000;
utils::Mf X;
utils::Mf X_test;
utils::Matrix<int64_t> y;
utils::Matrix<int64_t> y_test;
utils::Matrix<float> y;
utils::Matrix<float> y_test;
float data_loss;
float regularization_loss;
float loss;
float accuracy;
utils::Vector<uint64_t> class_targets;
utils::Vector<float> predictions;
utils::Matrix<float> predictions;
// Create dataset
neural_networks::create_spital_data<float, int64_t>(number_of_samples, number_of_classes, X, y);
//neural_networks::create_spital_data<float, int64_t>(number_of_samples, number_of_classes, X, y);
//neural_networks::create_vertical_data<float, int64_t>(number_of_samples, number_of_classes, X, y);
neural_networks::create_sine_data(number_of_samples, // samples
2.0f* 3.1415f, // length
X,
y);
neural_networks::Model<float> model;
// Create Dense layer with 2 input featues and 3 output values
neural_networks::Dense_Layer<float> dense1(
2, 16, // input/output
1, 64, // input/output
0.0f, // weight L1
5e-4f, // weight L2
5e-5f, // weight L2
0.0f, // bias L1
5e-4f // bias L2
5e-5f // bias L2
);
// Create ReLU activation (to be used with Dense layer)
neural_networks::Activation_ReLU<float> activation1;
neural_networks::Dropout_Layer<float> dropout1(0.1);
// Create a second Dense layer with 16 inputs (as we take the vlaues from the last layer)
// and 16 output values
neural_networks::Dense_Layer<float> dense2(
16, 16, // input/output
64, 64, // input/output
0.0f, // weight L1
5e-4f, // weight L2
5e-5f, // weight L2
0.0f, // bias L1
5e-4f // bias L2
5e-5f // bias L2
);
neural_networks::Activation_Softmax<float> activation2;
neural_networks::Activation_ReLU<float> activation2;
// Create a second Dense layer with 3 inputs (as we take the vlaues from the last layer)
// and 3 output values
neural_networks::Dense_Layer<float> dense3(
16, 1, // input/output
64, 1, // input/output
0.0f, // weight L1
5e-4f, // weight L2
5e-5f, // weight L2
0.0f, // bias L1
5e-4f // bias L2
5e-5f // bias L2
);
neural_networks::Activation_Sigmoid<float> activation3;
neural_networks::Activation_Linear<float> activation3;
// Create a Sfotmax classifier's combined loss and activation
//neural_networks::Activation_Softmax_Loss_CategoricalCrossentropy<float, int64_t> loss_activation;
neural_networks::Loss_BinaryCrossentropy<float, int64_t> loss_activation;
// Create optimizer
//neural_networks::Optimizer_SGD<float> optimizer(1, 1e-3, 0.5);
//neural_networks::Optimizer_Adagrad<float> optimizer(1, 1e-3, 1e-6);
//neural_networks::Optimizer_RMSprop<float> optimizer(1, 1e-3, 1e-6, 0.9);
neural_networks::Loss_MeanSquaredError<float> loss_function;
neural_networks::Optimizer_Adam<float> optimizer(
0.05, // Learning-rate
5e-5, // Learning-rate decay
1e-6, // epsilons
0.001, // Learning-rate
1e-3, // Learning-rate decay
1e-7, // epsilons
0.9, // beta 1
0.999 // beta 2
);
model.add(dense1);
model.add(activation1);
model.add(dropout1);
model.add(dense2);
model.add(activation2);
model.add(dense3);
model.add(activation3);
/* Accuracy precision for accuracy calculation
# There are no really accuracy factor for regression problem,
# but we can simulate/approximate it. We'll calculate it by checking
# how many values have a difference to their ground truth equivalent
# less than given precision
# We'll calculate this precision as a fraction of standard deviation
# of al the ground truth values */
// accuracy_precision = np.std(y) / 250
/*
float accuracy_precision = numerics::standard_deviation(y)/ 250.0f;
// Train in loop
for (uint64_t epoch = 0; epoch < number_of_epochs+1; ++epoch){
// Perform a forward pass of our training data through this layer
dense1.forward(X);
activation1.forward(dense1.outputs);
dropout1.forward(activation1.outputs);
//dropout1.forward(activation1.outputs);
dense2.forward(dropout1.outputs);
dense2.forward(activation1.outputs);
activation2.forward(dense2.outputs);
dense3.forward(activation2.outputs);
@@ -115,25 +129,21 @@ int main(int argc, char const *argv[])
// Perform a foard pass through the activation/loss function
// takes the output of the second dense layer here and returns loss
data_loss = loss_activation.calculate(activation3.outputs, y);
data_loss = loss_function.calculate(activation3.outputs, y);
// Calculate regularization penalty
regularization_loss = loss_activation.regularization_loss(dense1) +
loss_activation.regularization_loss(dense2) +
loss_activation.regularization_loss(dense3);
regularization_loss = loss_function.regularization_loss(dense1) +
loss_function.regularization_loss(dense2);
loss = data_loss + regularization_loss;
// Calculate accuracy from output of activation3 and targets
// Part in the brackets returns a binary mask - array consisting
// of True/False values, multiplying it by 1 changes it into array
// of 1s and 0s
predictions = numerics::greater_than(activation3.outputs, 0.5f).get_col(0);
accuracy = numerics::mean(numerics::equal_elementwise_serial(predictions, utils::veccast<float, int64_t>(y.get_col(0))));
predictions = activation3.outputs;
accuracy = numerics::mean(numerics::less( numerics::abs( numerics::sub(predictions, y)), accuracy_precision));
//accuracy = numerics::mean(numerics::equal_elementwise_serial(predictions, utils::veccast<float, int64_t>(y.get_col(0))));
if (!(epoch%100)){
if (!(epoch%100) && epoch != 0){
std::cout << "epoch: " << epoch;
std::cout << ", acc: " << accuracy;
std::cout << ", loss: " << loss;
@@ -145,16 +155,15 @@ int main(int argc, char const *argv[])
}
// Backward pass
loss_activation.backward(activation3.outputs, y);
loss_function.backward(activation3.outputs, y);
activation3.backward(loss_activation.dinputs);
activation3.backward(loss_function.dinputs);
dense3.backward(activation3.dinputs);
activation2.backward(dense3.dinputs);
dense2.backward(activation2.dinputs);
dropout1.backward(dense2.dinputs);
activation1.backward(dropout1.dinputs);
activation1.backward(dense2.dinputs);
dense1.backward(activation1.dinputs);
// Update weights and biases
@@ -166,6 +175,17 @@ int main(int argc, char const *argv[])
}
std::cout << "X, y, pred:" << std::endl;
for (uint64_t i = 0; i < X.rows(); ++i) {
std::cout << X(i, 0)
<< ", "
<< y(i, 0)
<< ", "
<< activation3.outputs(i, 0)
<< std::endl;
}
// Validate the model
// Create dataset
@@ -193,6 +213,12 @@ int main(int argc, char const *argv[])
loss = data_loss + regularization_loss;
// skal flyttes ned under loss functions.
predictions = activation3.outputs();
predictions = numerics::mean(numerics::abs(numerics::sub(predictions, y)));
std::cout << predictions << std::endl;
// Calculate accuracy from output of activation2 and targets
predictions = numerics::greater_than(activation3.outputs, 0.5f).get_col(0);
@@ -200,6 +226,6 @@ int main(int argc, char const *argv[])
std::cout << "validation, acc: " << accuracy << ", loss: " << loss << std::endl;
*/
return 0;
}