void mlp::set_activation_function(int activation_function, mlp_layer layer)
{
if (grt_mlp.validateActivationFunction(activation_function) == false)
{
flext::error("activation function %d is invalid, hint should be between 0-%d", activation_function, GRT::Neuron::NUMBER_OF_ACTIVATION_FUNCTIONS - 1);
return;
}
GRT::Neuron::ActivationFunctions activation_function_ = (GRT::Neuron::ActivationFunctions)activation_function;
switch (layer)
{
case LAYER_INPUT:
input_activation_function = activation_function_;
break;
case LAYER_HIDDEN:
hidden_activation_function = activation_function_;
break;
case LAYER_OUTPUT:
output_activation_function = activation_function_;
break;
default:
ml::error("no activation function for layer: " + std::to_string(layer));
return;
}
post("activation function set to " + grt_mlp.activationFunctionToString(activation_function_));
}
// Methods
// NOTE: ANN is special since it supports both regression and classification, we therefore override these methods
void ann::train()
{
const data_type data_type = get_data_type();
GRT::UINT numSamples = data_type == LABELLED_CLASSIFICATION ? classification_data.getNumSamples() : regression_data.getNumSamples();
if (numSamples == 0)
{
flext::error("no observations added, use 'add' to add training data");
return;
}
bool success = false;
if (data_type == LABELLED_CLASSIFICATION)
{
grt_ann.init(
classification_data.getNumDimensions(),
num_hidden_neurons,
classification_data.getNumClasses(),
input_activation_function,
hidden_activation_function,
output_activation_function
);
success = grt_ann.train(classification_data);
}
else if (data_type == LABELLED_REGRESSION)
{
grt_ann.init(
regression_data.getNumInputDimensions(),
num_hidden_neurons,
regression_data.getNumTargetDimensions(),
input_activation_function,
hidden_activation_function,
output_activation_function
);
success = grt_ann.train(regression_data);
}
if (!success)
{
flext::error("training failed");
}
t_atom a_success;
SetInt(a_success, success);
ToOutAnything(1, get_s_train(), 1, &a_success);
}
void ann::error()
{
if (!grt_ann.getTrained())
{
flext::error("model not yet trained, send the \"train\" message to train");
return;
}
float error_f = grt_ann.getTrainingError();
t_atom error_a;
SetFloat(error_a, error_f);
ToOutAnything(0, get_s_error(), 1, &error_a);
}
void ann::set_null_rejection(bool null_rejection)
{
bool success = grt_ann.setNullRejection(null_rejection);
if (success == false)
{
flext::error("unable to set null_rejection");
}
}
void ann::set_gamma(float gamma)
{
bool success = grt_ann.setGamma(gamma);
if (success == false)
{
flext::error("unable to set gamma");
}
}
void ann::set_momentum(float momentum)
{
bool success = grt_ann.setMomentum(momentum);
if (success == false)
{
flext::error("unable to set momentum, hint: should be between 0-1");
}
}
void ann::set_training_rate(float training_rate)
{
bool success = grt_ann.setTrainingRate(training_rate);
if (success == false)
{
flext::error("unable to set training_rate, hint: should be between 0-1");
}
}
void ann::set_null_rejection_coeff(float null_rejection_coeff)
{
bool success = grt_ann.setNullRejectionCoeff(null_rejection_coeff);
if (success == false)
{
flext::error("unable to set null_rejection_coeff, hint: should be greater than 0");
}
}
void ann::set_validation_set_size(int validation_set_size)
{
bool success = grt_ann.setValidationSetSize(validation_set_size);
if (success == false)
{
flext::error("unable to set validation_set_size, hint: should be between 0-100");
}
}
void ann::set_randomise_training_order(bool randomise_training_order)
{
bool success = grt_ann.setRandomiseTrainingOrder(randomise_training_order);
if (success == false)
{
flext::error("unable to set randomise_training_order, hint: should be 0 or 1");
}
}
void ann::set_min_epochs(int min_epochs)
{
bool success = grt_ann.setMinNumEpochs(min_epochs);
if (success == false)
{
flext::error("unable to set min_epochs, hint: should be greater than 0");
}
}
void ann::set_min_change(float min_change)
{
bool success = grt_ann.setMinChange(min_change);
if (success == false)
{
flext::error("unable to set min_change, hint: should be greater than 0");
}
}
void ann::set_use_validation_set(bool use_validation_set)
{
bool success = grt_ann.setUseValidationSet(use_validation_set);
if (success == false)
{
flext::error("unable to set use_validation_set, hint: should be 0 or 1");
}
}
void ann::set_rand_training_iterations(int rand_training_iterations)
{
bool success = grt_ann.setNumRandomTrainingIterations(rand_training_iterations);
if (success == false)
{
flext::error("unable to set rand_training_iterations, hint: should be greater than 0");
}
}
void ann::set_activation_function(int activation_function, ann_layer layer)
{
GRT::Neuron::Type activation_function_ = GRT::Neuron::Type::LINEAR;
try
{
activation_function_ = get_grt_neuron_type(activation_function);
}
catch (std::exception& e)
{
flext::error(e.what());
return;
}
if (grt_ann.validateActivationFunction(activation_function_) == false)
{
flext::error("activation function %d is invalid, hint should be between 0-%d", activation_function, GRT::Neuron::NUMBER_OF_ACTIVATION_FUNCTIONS - 1);
return;
}
switch (layer)
{
case LAYER_INPUT:
input_activation_function = activation_function_;
break;
case LAYER_HIDDEN:
hidden_activation_function = activation_function_;
break;
case LAYER_OUTPUT:
output_activation_function = activation_function_;
break;
default:
ml::error("no activation function for layer: " + std::to_string(layer));
return;
}
post("activation function set to " + grt_ann.activationFunctionToString(activation_function_));
}
void ann::get_use_validation_set(bool &use_validation_set) const
{
use_validation_set = grt_ann.getUseValidationSet();
}
void ann::get_rand_training_iterations(int &rand_training_iterations) const
{
rand_training_iterations = grt_ann.getNumRandomTrainingIterations();
}
void ann::get_null_rejection_coeff(float &null_rejection_coeff) const
{
null_rejection_coeff = grt_ann.getNullRejectionCoeff();
}
void ann::get_null_rejection(bool &null_rejection) const
{
null_rejection = grt_ann.getNullRejectionEnabled();
}
void ann::get_gamma(float &gamma) const
{
gamma = grt_ann.getGamma();
}
void ann::get_momentum(float &momentum) const
{
momentum = grt_ann.getMomentum();
}
void ann::get_training_rate(float &training_rate) const
{
training_rate = grt_ann.getTrainingRate();
}
void ann::map(int argc, const t_atom *argv)
{
const data_type data_type = get_data_type();
GRT::UINT numSamples = data_type == LABELLED_CLASSIFICATION ? classification_data.getNumSamples() : regression_data.getNumSamples();
if (numSamples == 0)
{
flext::error("no observations added, use 'add' to add training data");
return;
}
if (grt_ann.getTrained() == false)
{
flext::error("model has not been trained, use 'train' to train the model");
return;
}
GRT::UINT numInputNeurons = grt_ann.getNumInputNeurons();
GRT::VectorDouble query(numInputNeurons);
if (argc < 0 || (unsigned)argc != numInputNeurons)
{
flext::error("invalid input length, expected %d, got %d", numInputNeurons, argc);
}
for (uint32_t index = 0; index < (uint32_t)argc; ++index)
{
double value = GetAFloat(argv[index]);
query[index] = value;
}
bool success = grt_ann.predict(query);
if (success == false)
{
flext::error("unable to map input");
return;
}
if (grt_ann.getClassificationModeActive())
{
const GRT::VectorDouble likelihoods = grt_ann.getClassLikelihoods();
const GRT::Vector<GRT::UINT> labels = classification_data.getClassLabels();
const GRT::UINT predicted = grt_ann.getPredictedClassLabel();
const GRT::UINT classification = predicted == 0 ? 0 : get_class_id_for_index(predicted);
if (likelihoods.size() != labels.size())
{
flext::error("labels / likelihoods size mismatch");
}
else if (probs)
{
AtomList probs_list;
for (unsigned count = 0; count < labels.size(); ++count)
{
t_atom label_a;
t_atom likelihood_a;
SetFloat(likelihood_a, static_cast<float>(likelihoods[count]));
SetInt(label_a, get_class_id_for_index(labels[count]));
probs_list.Append(label_a);
probs_list.Append(likelihood_a);
}
ToOutAnything(1, get_s_probs(), probs_list);
}
ToOutInt(0, classification);
}
else if (grt_ann.getRegressionModeActive())
{
GRT::VectorDouble regression_data = grt_ann.getRegressionData();
GRT::VectorDouble::size_type numOutputDimensions = regression_data.size();
if (numOutputDimensions != grt_ann.getNumOutputNeurons())
{
flext::error("invalid output dimensions: %d", numOutputDimensions);
return;
}
AtomList result;
for (uint32_t index = 0; index < numOutputDimensions; ++index)
{
t_atom value_a;
double value = regression_data[index];
SetFloat(value_a, value);
result.Append(value_a);
}
ToOutList(0, result);
}
}
void ann::get_validation_set_size(int &validation_set_size) const
{
validation_set_size = grt_ann.getValidationSetSize();
}
void ann::get_randomise_training_order(bool &randomise_training_order) const
{
randomise_training_order = grt_ann.getRandomiseTrainingOrder();
}
void mlp::clear()
{
grt_mlp.clear();
ml::clear();
}
void ann::clear()
{
grt_ann.clear();
ml::clear();
clear_index_maps();
}
void ann::get_min_change(float &min_change) const
{
min_change = grt_ann.getMinChange();
}
void ann::get_max_epochs(int &max_epochs) const
{
max_epochs = grt_ann.getMaxNumEpochs();
}
void ann::set_max_epochs(int max_epochs)
{
grt_ann.setMaxNumEpochs(max_epochs);
}