def subSetMLP(self,
filename,
percentage,
layer_width,
layer_height,
learning_rate,
dropout_rate,
fitness_threshold,
batch_size,
display_step,
save_option=True):
self.staticController = self.importer.readStaticController(filename)
fullSet = DataSet()
fullSet.readSetFromController(self.staticController)
fullSet.formatToBinary()
subSet = DataSet()
subSet.readSubsetFromController(self.staticController, percentage)
subSet.formatToBinary()
self.nnm.setDebugMode(True)
self.nnm.setType(NNTypes.MLP)
self.nnm.setTrainingMethod(NNOptimizer.Adam)
self.nnm.setActivationFunction(NNActivationFunction.Sigmoid)
self.nnm.setDataSet(subSet)
self.nnm.setDropoutRate(dropout_rate)
self.nnm.rectangularHiddenLayers(layer_width, layer_height)
self.nnm.initialize(learning_rate, fitness_threshold, batch_size,
display_step, -1, 5000)
self.nnm.getDataSize()
# Train model and visualize performance
self.nnm.train()
self.nnm.plot()
fitness, wrong_states = self.nnm.checkFitness(fullSet)
self.nnm.randomCheck(fullSet)
if (save_option):
self.exporter.saveNetwork(self.nnm)
self.exporter.saveWrongStates(wrong_states)
self.exporter.saveMatlabMLP(self.staticController, self.nnm)
self.nnm.close()
self.cleanMemory()
def importMLP(self,
import_path,
filename,
layer_width,
layer_height,
learning_rate,
dropout_rate,
fitness_threshold,
batch_size,
display_step,
save_option=True):
self.staticController = self.importer.readStaticController(filename)
fullSet = DataSet()
fullSet.readSetFromController(self.staticController)
fullSet.formatToBinary()
self.nnm.setDebugMode(True)
self.nnm.setType(NNTypes.MLP)
self.nnm.setTrainingMethod(NNOptimizer.Adam)
self.nnm.setActivationFunction(NNActivationFunction.Sigmoid)
self.nnm.setDataSet(fullSet)
# Option to adjust parameters for new training session
self.nnm.setDropoutRate(dropout_rate)
self.nnm.rectangularHiddenLayers(layer_width, layer_height)
self.nnm.initialize(learning_rate, fitness_threshold, batch_size,
display_step)
self.nnm.getDataSize()
# Restore Network from saved file:
self.importer.restoreNetwork(self.nnm, import_path)
# Train model and visualize performance
self.nnm.train()
self.nnm.plot()
fitness, wrong_states = self.nnm.checkFitness(fullSet)
self.nnm.randomCheck(fullSet)
# Save Network or Variables
if (save_option):
self.exporter.saveNetwork(self.nnm)
self.exporter.saveWrongStates(wrong_states)
self.exporter.saveMatlabMLP(self.staticController, self.nnm)
self.nnm.close()
def scoutLearningRateConvergence(self, filename, layer_width, layer_height,
epoch_threshold, rates, batch_size,
display_step):
self.staticController = self.importer.readStaticController(filename)
dataSet = DataSet()
dataSet.readSetFromController(self.staticController)
dataSet.formatToBinary()
self.nnm.setDebugMode(False)
fitnesses = []
for r in rates:
print("\nLearning rate: " + str(r))
self.nnm.setType(NNTypes.MLP)
self.nnm.setTrainingMethod(NNOptimizer.Adam)
self.nnm.setActivationFunction(NNActivationFunction.Sigmoid)
self.nnm.setDataSet(dataSet)
self.nnm.rectangularHiddenLayers(layer_width, layer_height)
self.nnm.initializeNeuralNetwork()
self.nnm.initializeTraining(r, 1.0, batch_size, display_step,
epoch_threshold)
self.nnm.train()
fitness, wrong_states = self.nnm.checkFitness(dataSet)
self.fitnesses.append(fitness)
self.nnm.close()
# Plot
plt.semilogx(rates, fitnesses, 'r-')
plt.xlabel("Rates")
plt.ylabel("Fitness")
plt.grid()
(x1, x2, y1, y2) = plt.axis()
plt.axis((min(rates), max(rates), 0.0, y2 + 0.1))
plt.show()
self.cleanMemory()
