From 569d5c902f3a9cc58a8e6271dbab403414d929f8 Mon Sep 17 00:00:00 2001
From: Martin Beseda <martin.beseda@vsb.cz>
Date: Tue, 27 Aug 2019 16:05:11 +0200
Subject: [PATCH] [WIP] Fixing ACSF NN constructor.
---
src/LearningMethods/LevenbergMarquardt.cpp | 1 +
src/Network/ACSFNeuralNetwork.cpp | 80 ++++++++++++-
src/Network/NeuralNetwork.cpp | 125 +--------------------
src/Network/NeuralNetwork.h | 6 -
src/examples/acsf2.cpp | 62 +++++++---
5 files changed, 127 insertions(+), 147 deletions(-)
diff --git a/src/LearningMethods/LevenbergMarquardt.cpp b/src/LearningMethods/LevenbergMarquardt.cpp
index 273ccbcc..be12f7bf 100644
--- a/src/LearningMethods/LevenbergMarquardt.cpp
+++ b/src/LearningMethods/LevenbergMarquardt.cpp
@@ -170,6 +170,7 @@ namespace lib4neuro {
gradient_norm = 0;
+ // TODO parallelize with OpenMP?
for (size_t ci = 0; ci < n_parameters; ++ci) {
mem_double = rhs[ci];
mem_double *= mem_double;
diff --git a/src/Network/ACSFNeuralNetwork.cpp b/src/Network/ACSFNeuralNetwork.cpp
index 2236c356..2549aff9 100644
--- a/src/Network/ACSFNeuralNetwork.cpp
+++ b/src/Network/ACSFNeuralNetwork.cpp
@@ -14,6 +14,11 @@ lib4neuro::ACSFNeuralNetwork::ACSFNeuralNetwork(std::unordered_map<ELEMENT_SYMBO
std::vector<size_t> inputs;
std::vector<size_t> subnet_outputs;
size_t neuron_idx;
+
+ std::unordered_map<ELEMENT_SYMBOL, std::vector<size_t>> subnet_neuron_idxs;
+ std::unordered_map<ELEMENT_SYMBOL, std::vector<size_t>> subnet_connection_idxs;
+
+
for(size_t i = 0; i < elements_list.size(); i++) {
std::vector<size_t> previous_layer;
std::vector<size_t> new_layer;
@@ -35,9 +40,20 @@ lib4neuro::ACSFNeuralNetwork::ACSFNeuralNetwork(std::unordered_map<ELEMENT_SYMBO
inputs.emplace_back(neuron_idx);
}
+ /* Create subnet for the current element */
+ bool new_subnet = false;
+ if(subnet_neuron_idxs.find(elements_list.at(i)) == subnet_neuron_idxs.end()) {
+ new_subnet = true;
+ subnet_neuron_idxs[elements_list.at(i)] = std::vector<size_t>();
+ subnet_connection_idxs[elements_list.at(i)] = std::vector<size_t>();
+ }
+
/* Create hidden layers in sub-net */
std::vector<unsigned int> n_neurons = n_hidden_neurons[elements_list.at(i)];
std::vector<NEURON_TYPE> types = type_hidden_neurons[elements_list.at(i)];
+ size_t local_neuron_idx = 0;
+ size_t local_connection_idx = 0;
+
for(size_t j = 0; j < n_neurons.size(); j++) { /* Iterate over hidden layers */
/* Create hidden neurons */
for(size_t k = 0; k < n_neurons.at(j); k++) {
@@ -61,18 +77,74 @@ lib4neuro::ACSFNeuralNetwork::ACSFNeuralNetwork(std::unordered_map<ELEMENT_SYMBO
}
}
- neuron_idx = this->add_neuron(hid_n, BIAS_TYPE::NEXT_BIAS);
+ if(new_subnet) {
+ neuron_idx = this->add_neuron(hid_n,
+ BIAS_TYPE::NEXT_BIAS);
+ subnet_neuron_idxs[elements_list.at(i)].emplace_back(neuron_idx);
+ } else {
+ neuron_idx = this->add_neuron(hid_n,
+ BIAS_TYPE::EXISTING_BIAS, subnet_neuron_idxs[elements_list.at(i)].at(0)+local_neuron_idx);
+ }
+ local_neuron_idx++;
new_layer.emplace_back(neuron_idx);
/* Connect hidden neuron to the previous layer */
for(auto prev_n : previous_layer) {
- this->add_connection_simple(prev_n, neuron_idx, SIMPLE_CONNECTION_TYPE::NEXT_WEIGHT);
+ if(new_subnet) {
+ subnet_connection_idxs[elements_list.at(i)].emplace_back(this->add_connection_simple(prev_n,
+ neuron_idx,
+ SIMPLE_CONNECTION_TYPE::NEXT_WEIGHT));
+
+ } else {
+ this->add_connection_simple(prev_n,
+ neuron_idx,
+ SIMPLE_CONNECTION_TYPE::EXISTING_WEIGHT, subnet_connection_idxs[elements_list.at(i)].at(0)+local_connection_idx);
+ }
+ local_connection_idx++;
}
- previous_layer = new_layer;
- new_layer.clear();
}
+ previous_layer = new_layer;
+ new_layer.clear();
}
+ /* Create hidden layers in sub-net */
+// std::vector<unsigned int> n_neurons = n_hidden_neurons[elements_list.at(i)];
+// std::vector<NEURON_TYPE> types = type_hidden_neurons[elements_list.at(i)];
+// for(size_t j = 0; j < n_neurons.size(); j++) { /* Iterate over hidden layers */
+// /* Create hidden neurons */
+// for(size_t k = 0; k < n_neurons.at(j); k++) {
+// std::shared_ptr<Neuron> hid_n;
+// switch(types.at(j)) {
+// case NEURON_TYPE::LOGISTIC: {
+// hid_n = std::make_shared<NeuronLogistic>();
+// break;
+// }
+// case NEURON_TYPE::BINARY: {
+// hid_n = std::make_shared<NeuronBinary>();
+// break;
+// }
+// case NEURON_TYPE::CONSTANT: {
+// hid_n = std::make_shared<NeuronConstant>();
+// break;
+// }
+// case NEURON_TYPE::LINEAR: {
+// hid_n = std::make_shared<NeuronLinear>();
+// break;
+// }
+// }
+//
+// neuron_idx = this->add_neuron(hid_n, BIAS_TYPE::NEXT_BIAS);
+// new_layer.emplace_back(neuron_idx);
+//
+// /* Connect hidden neuron to the previous layer */
+// for(auto prev_n : previous_layer) {
+// this->add_connection_simple(prev_n, neuron_idx, SIMPLE_CONNECTION_TYPE::NEXT_WEIGHT);
+// }
+// previous_layer = new_layer;
+// new_layer.clear();
+// }
+// }
+
/* Create output neurons for sub-net */
std::shared_ptr<NeuronLinear> sub_out_n = std::make_shared<NeuronLinear>();
neuron_idx = this->add_neuron(sub_out_n, BIAS_TYPE::NO_BIAS);
diff --git a/src/Network/NeuralNetwork.cpp b/src/Network/NeuralNetwork.cpp
index 0f5abd45..7a9ed274 100644
--- a/src/Network/NeuralNetwork.cpp
+++ b/src/Network/NeuralNetwork.cpp
@@ -260,7 +260,8 @@ namespace lib4neuro {
}
if (this->input_neuron_indices.size() != input.size()) {
- THROW_INVALID_ARGUMENT_ERROR("Data input size != Network input size");
+ THROW_INVALID_ARGUMENT_ERROR("Network input size(" + std::to_string(this->input_neuron_indices.size())
+ + ") != Data input size(" + std::to_string(input.size()) + ")");
}
if (this->output_neuron_indices.size() != output.size()) {
@@ -1051,127 +1052,5 @@ namespace lib4neuro {
error_partial[i] = 0;
}
}
-
-// NeuralNetwork::NeuralNetwork(std::unordered_map<ELEMENT_SYMBOL, Element*, ELEMENT_SYMBOL_HASH>* elements,
-// std::vector<std::pair<ELEMENT_SYMBOL, std::vector<double>>>* particles) {
-//
-// }
-
- NeuralNetwork::NeuralNetwork(std::unordered_map<ELEMENT_SYMBOL, Element*>& elements,
- std::vector<std::pair<ELEMENT_SYMBOL,
- std::vector<double>>>& particles) {
- /* Create a new set of Atomic-Centered Symmetry Functions serving as coordinates */
-// std::vector<std::vector<double>> acsf_coords(particles.size());
-// std::vector<double> acsf_coords_single;
-// for(size_t i = 0; i < particles.size(); i++) { /* Iterate over all the particles */
-// for (auto sym_func : elements[particles.at(i).first]->getSymmetryFunctions()) {
-// acsf_coords_single.emplace_back(sym_func->eval(i, particles));
-// }
-//
-// acsf_coords.at(i) = acsf_coords_single;
-// acsf_coords_single.clear();
-// }
-//
-// /* Check for duplicates in new coordinates */
-// std::vector<std::vector<double>> unique_coords = acsf_coords;
-// std::sort(unique_coords.begin(), unique_coords.end());
-// unique_coords.erase(std::unique(unique_coords.begin(), unique_coords.end()), unique_coords.end());
-//
-// if(unique_coords.size() != acsf_coords.size()) {
-// THROW_RUNTIME_ERROR("Not all descriptors are unique with currently specified symmetry functions!");
-// }
-//
-// for(auto e : acsf_coords) {
-// for(auto ee : e) {
-// std::cout << ee << " ";
-// }
-// std::cout << std::endl;
-// }
-
- /* Construct the neural network */
-// std::vector<size_t> inputs;
-// std::vector<size_t> subnet_outputs;
-// size_t neuron_idx;
-// for(size_t i = 0; i < particles.size(); i++) {
-// std::vector<size_t> previous_layer;
-// std::vector<size_t> new_layer;
-//
-// /* Create input neurons for sub-net */
-// std::shared_ptr<NeuronLinear> inp_n;
-// for(size_t j = 0; j < acsf_coords.at(i).size(); j++) {
-// inp_n = std::make_shared<NeuronLinear>();
-// neuron_idx = this->add_neuron(inp_n, BIAS_TYPE::NO_BIAS);
-// previous_layer.emplace_back(neuron_idx);
-// inputs.emplace_back(neuron_idx);
-// }
-//
-// /* Add an additional input neuron for charge, if provided */
-// if(elements[particles.at(i).first]->isCharge()) {
-// inp_n = std::make_shared<NeuronLinear>();
-// neuron_idx = this->add_neuron(inp_n, BIAS_TYPE::NO_BIAS);
-// previous_layer.emplace_back(neuron_idx);
-// inputs.emplace_back(neuron_idx);
-// }
-//
-// /* Create hidden layers in sub-net */
-// std::vector<unsigned int> n_neurons = elements[particles.at(i).first]->getNHiddenNeurons();
-// std::vector<NEURON_TYPE> types = elements[particles.at(i).first]->getHiddenNeuronTypes();
-// for(size_t j = 0; j < n_neurons.size(); j++) { /* Iterate over hidden layers */
-// /* Create hidden neurons */
-// for(size_t k = 0; k < n_neurons.at(j); k++) {
-// std::shared_ptr<Neuron> hid_n;
-// switch(types.at(j)) {
-// case NEURON_TYPE::LOGISTIC: {
-// hid_n = std::make_shared<NeuronLogistic>();
-// break;
-// }
-// case NEURON_TYPE::BINARY: {
-// hid_n = std::make_shared<NeuronBinary>();
-// break;
-// }
-// case NEURON_TYPE::CONSTANT: {
-// hid_n = std::make_shared<NeuronConstant>();
-// break;
-// }
-// case NEURON_TYPE::LINEAR: {
-// hid_n = std::make_shared<NeuronLinear>();
-// break;
-// }
-// }
-//
-// neuron_idx = this->add_neuron(hid_n, BIAS_TYPE::NEXT_BIAS);
-// new_layer.emplace_back(neuron_idx);
-//
-// /* Connect hidden neuron to the previous layer */
-// for(auto prev_n : previous_layer) {
-// this->add_connection_simple(prev_n, neuron_idx, SIMPLE_CONNECTION_TYPE::NEXT_WEIGHT);
-// }
-// previous_layer = new_layer;
-// new_layer.clear();
-// }
-// }
-//
-// /* Create output neurons for sub-net */
-// std::shared_ptr<NeuronLinear> sub_out_n = std::make_shared<NeuronLinear>();
-// neuron_idx = this->add_neuron(sub_out_n, BIAS_TYPE::NO_BIAS);
-// subnet_outputs.emplace_back(neuron_idx);
-// for(auto prev_n : previous_layer) {
-// this->add_connection_simple(prev_n, neuron_idx);
-// }
-// }
-//
-// /* Specify network inputs */
-// this->specify_input_neurons(inputs);
-//
-// /* Create final output layer */
-// std::shared_ptr<NeuronLinear> final_out_n = std::make_shared<NeuronLinear>();
-// neuron_idx = this->add_neuron(final_out_n, BIAS_TYPE::NO_BIAS);
-// for(auto subnet_output : subnet_outputs) {
-// this->add_connection_constant(subnet_output, neuron_idx, 1);
-// }
-// std::vector<size_t> outputs = {neuron_idx};
-// this->specify_output_neurons(outputs);
- }
-
}
diff --git a/src/Network/NeuralNetwork.h b/src/Network/NeuralNetwork.h
index f9c2c600..499252ec 100644
--- a/src/Network/NeuralNetwork.h
+++ b/src/Network/NeuralNetwork.h
@@ -235,12 +235,6 @@ namespace lib4neuro {
*/
LIB4NEURO_API explicit NeuralNetwork(std::string filepath);
- // TODO remove method
- LIB4NEURO_API explicit NeuralNetwork(std::unordered_map<ELEMENT_SYMBOL,
- Element*>& elements,
- std::vector<std::pair<ELEMENT_SYMBOL,
- std::vector<double>>>& particles);
-
/**
*
*/
diff --git a/src/examples/acsf2.cpp b/src/examples/acsf2.cpp
index a9be9165..f896a108 100644
--- a/src/examples/acsf2.cpp
+++ b/src/examples/acsf2.cpp
@@ -133,7 +133,7 @@ int main() {
elements[l4n::ELEMENT_SYMBOL::He] = &helium;
/* Read data */
- l4n::XYZReader reader("../../data/HE21+T4.xyz");
+ l4n::XYZReader reader("/home/martin/lib4neuro/data/HE21+T1.xyz");
reader.read();
std::cout << "Finished reading data" << std::endl;
@@ -142,18 +142,52 @@ int main() {
/* Create a neural network */
std::unordered_map<l4n::ELEMENT_SYMBOL, std::vector<unsigned int>> n_hidden_neurons;
- n_hidden_neurons[l4n::ELEMENT_SYMBOL::He] = {10};
+ n_hidden_neurons[l4n::ELEMENT_SYMBOL::He] = {2};
std::unordered_map<l4n::ELEMENT_SYMBOL, std::vector<l4n::NEURON_TYPE>> type_hidden_neurons;
type_hidden_neurons[l4n::ELEMENT_SYMBOL::He] = {l4n::NEURON_TYPE::LOGISTIC};
l4n::ACSFNeuralNetwork net(elements, *reader.get_element_list(), reader.contains_charge(), n_hidden_neurons, type_hidden_neurons);
+// l4n::NeuralNetwork net;
+// std::vector<std::shared_ptr<l4n::NeuronLinear>> inps;
+// std::vector<size_t> inps_inds;
+// for(unsigned int i = 0; i < 126; i++) {
+// std::shared_ptr<l4n::NeuronLinear> inp = std::make_shared<l4n::NeuronLinear>();
+// inps.emplace_back(inp);
+// inps_inds.emplace_back(net.add_neuron(inp, l4n::BIAS_TYPE::NO_BIAS));
+// }
+//
+// net.specify_input_neurons(inps_inds);
+//
+// std::vector<std::shared_ptr<l4n::NeuronLogistic>> hids;
+//
+// std::vector<unsigned int> hids_idxs;
+// size_t idx;
+// unsigned int n_hidden = 5;
+// for(unsigned int i = 0; i < n_hidden; i++) {
+// std::shared_ptr<l4n::NeuronLogistic> hid = std::make_shared<l4n::NeuronLogistic>();
+// hids.emplace_back(hid);
+// idx = net.add_neuron(hid, l4n::BIAS_TYPE::NEXT_BIAS);
+// hids_idxs.emplace_back(idx);
+//
+// for(unsigned int j = 0; j < 126; j++) {
+// net.add_connection_simple(j, idx);
+// }
+// }
+//
+// std::shared_ptr<l4n::NeuronLinear> out = std::make_shared<l4n::NeuronLinear>();
+// idx = net.add_neuron(out, l4n::BIAS_TYPE::NO_BIAS);
+// std::vector<size_t> out_inds = {idx};
+// for(unsigned int i = 0; i < n_hidden; i++) {
+// net.add_connection_simple(hids_idxs.at(i), idx);
+// }
+// net.specify_output_neurons(out_inds);
l4n::MSE mse(&net, ds.get());
net.randomize_parameters();
// optimize_via_particle_swarm(net, mse);
- // double err1 = optimize_via_LBMQ(net, mse);
+ double err1 = optimize_via_LBMQ(net, mse);
double err2 = optimize_via_gradient_descent(net, mse);
if(err2 > 0.00001) {
@@ -161,17 +195,17 @@ int main() {
}
/* Print fit comparison with real data */
-// std::vector<double> output;
-// output.resize(1);
-//
-// for(auto e : *ds->get_data()) {
-// for(auto inp_e : e.first) {
-// std::cout << inp_e << " ";
-// }
-// std::cout << e.second.at(0) << " ";
-// net.eval_single(e.first, output);
-// std::cout << output.at(0) << std::endl;
-// }
+ std::vector<double> output;
+ output.resize(1);
+
+ for(auto e : *ds->get_data()) {
+ for(auto inp_e : e.first) {
+ std::cout << inp_e << " ";
+ }
+ std::cout << e.second.at(0) << " ";
+ net.eval_single(e.first, output);
+ std::cout << output.at(0) << std::endl;
+ }
} catch (const std::exception& e) {
std::cerr << e.what() << std::endl;
--
GitLab