class NeuralNetworkManager:
def __init__(self):
self.type = None
self.nn = MLP()
self.training_method = None
self.hidden_activation_function = None
self.output_activation_function = None
self.dropout_rate = 0.0
self.training = True
self.learning_rate = 0.1
self.fitness_threshold = 0.75
self.epoch_threshold = -1
self.batch_size = 100
self.shuffle_rate = 2500
self.display_step = 1000
self.epoch = 0
self.layers = []
self.data_set = None
self.bed = BinaryEncoderDecoder()
self.utils = Utilities()
self.controller = StaticController()
self.debug_mode = False
# Plotting variables
self.losses = []
self.fitnesses = []
self.iterations = []
self.save_location = './nn/log/'
self.encode = None
self.non_det = False
# Getters and setters
def getType(self): return self.type
def getTrainingMethod(self): return self.training_method
def getActivationFunctionHidden(self): return self.hidden_activation_function
def getActivationFunctionOutput(self): return self.output_activation_function
def getLearningRate(self): return self.learning_rate
def getFitnessThreshold(self): return self.fitness_threshold
def getBatchSize(self): return self.batch_size
def getDisplayStep(self): return self.display_step
def getEpoch(self): return self.epoch
def getEpochThreshold(self): return self.epoch_threshold
def getDropoutRate(self): return self.dropout_rate
def getShuffleRate(self): return self.shuffle_rate
def getSaveLocation(self): return self.save_location
def setType(self, value): self.type = value
def setTrainingMethod(self, optimizer): self.training_method = optimizer
def setActivationFunctionHidden(self, activation_function): self.hidden_activation_function = activation_function
def setActivationFunctionOutput(self, activation_function): self.output_activation_function = activation_function
def setLearningRate(self, value): self.learning_rate = value
def setFitnessThreshold(self, value): self.fitness_threshold = value
def setBatchSize(self, value): self.batch_size = value
def setDisplayStep(self, value): self.display_step = value
def setEpochThreshold(self, value): self.epoch_threshold = value
def setDropoutRate(self, value): self.dropout_rate = value
def setShuffleRate(self, value): self.shuffle_rate = value
def setSaveLocation(self, value): self.save_location = value
def setDebugMode(self, value): self.debug_mode = value
def setDataSet(self, data_set): self.data_set = data_set
def setEncodeTypes(self, value): self.encode = value
# Hidden layer generation functions
# Linearly increase/decrease neurons per hidden layer based on the input and ouput neurons
def linearHiddenLayers(self, num_hidden_layers):
self.layers = []
x_dim = self.data_set.getXDim()
y_dim = self.data_set.getYDim()
a = (y_dim - x_dim)/(num_hidden_layers + 1)
self.layers.append(x_dim)
for i in range(1, num_hidden_layers + 1):
self.layers.append(round(x_dim + a*i))
self.layers.append(y_dim)
return self.layers
# Rectangular hidden layer
def rectangularHiddenLayers(self, width, height):
self.layers = []
self.layers.append(self.data_set.getXDim())
for i in range(width):
self.layers.append(height)
self.layers.append(self.data_set.getYDim())
#Customize layer sturcture
def customHiddenLayers(self, layer):
self.layers = []
x_dim = self.data_set.getXDim()
y_dim = self.data_set.getYDim()
self.layers.append(x_dim)
for i in range(1, len(layer)+1):
self.layers.append(layer[i-1])
self.layers.append(y_dim)
return self.layers
# Initialize neural network
def initializeNeuralNetwork(self):
if(self.debug_mode):
print("\nNeural network initialization:")
if(self.type == NNTypes.MLP):
self.nn = MLP()
self.nn.setDebugMode(self.debug_mode)
if(self.debug_mode):
print("Neural network type: MLP")
# Initialize network and loss function
self.nn.setNeurons(self.layers)
self.nn.setDropoutRate(self.dropout_rate)
self.nn.setActivationFunctionHidden(self.hidden_activation_function)
self.nn.setActivationFunctionOutput(self.output_activation_function)
self.nn.setTaskTypes(self.encode)
self.nn.initializeNetwork()
# Print neural network status
if(self.debug_mode):
print("Generated network neuron topology: " + str(self.layers) + " with dropout rate: " + str(self.nn.getDropoutRate()))
# Initialize training function
def initializeTraining(self, learning_rate, fitness_threshold, batch_size, display_step, epoch_threshold = -1, shuffle_rate = 10000):
self.learning_rate = learning_rate
self.fitness_threshold = fitness_threshold
self.batch_size = batch_size
self.display_step = display_step
self.epoch_threshold = epoch_threshold
self.shuffle_rate = shuffle_rate
self.nn.initializeLossFunction()
self.nn.initializeTrainFunction(self.training_method, self.learning_rate)
# Initialize fitness function
def initializeFitnessFunction(self):
with tf.name_scope("fitness"):
eta = self.data_set.getYEta()
size = self.data_set.getSize()
lower_bound = tf.subtract(self.nn.y, eta)
upper_bound = tf.add(self.nn.y, eta)
is_fit = tf.logical_and(tf.greater_equal(self.nn.predictor, lower_bound), tf.less(self.nn.predictor, upper_bound))
non_zero = tf.to_float(tf.count_nonzero(tf.reduce_min(tf.cast(is_fit, tf.int8), 1)))
self.fitness = non_zero/size
tf.summary.scalar("fitness", self.fitness)
# General initialization function to call all functions
def initialize(self, learning_rate, fitness_threshold, batch_size, display_step, epoch_threshold = -1, shuffle_rate = 10000):
self.initializeNeuralNetwork()
self.initializeFitnessFunction()
self.initializeTraining(learning_rate, fitness_threshold, batch_size, display_step, epoch_threshold, shuffle_rate)
self.train_writer = tf.summary.FileWriter(self.save_location, self.nn.session.graph)
# Check a state against the dataset and nn by using its id in the dataset
def checkByIndex(self, index, out):
x = self.data_set.x[index]
estimation = self.nn.estimate([x])[0]
y = self.data_set.getY(index)
y_eta = self.data_set.getYEta()
equal = True
for i in range(self.data_set.getYDim()):
if(not((y[i] - y_eta[i]) <= estimation[i] and (y[i] + y_eta[i]) > estimation[i])):
equal = False
if(out):
print("u: " + str(y) + " u_: " + str(np.round(estimation,2)) + " within etas: " + str(equal))
return equal
# Check fitness of the neural network for a specific dataset and return wrong states
# as of right now it assumes a binary encoding of the dataset
def checkFitness(self, data_set):
self.data_set = data_set
size = self.data_set.getSize()
fit = size
wrong = []
x, y = self.data_set.x, self.data_set.y
y_eta = self.data_set.getYEta()
y_dim = self.data_set.getYDim()
estimation = self.nn.estimate(self.data_set.x)
for i in range(size):
equal = True
for j in range(y_dim):
if(not((y[i][j] - y_eta[j]) <= estimation[i][j] and (y[i][j] + y_eta[j]) > estimation[i][j]) and equal):
wrong.append(self.bed.baton(x[i]))
fit -= 1
equal = False
fitness = fit/size*100
return fitness, wrong
# Check fitness of the neural network for a specific dataset and return wrong states
# as of right now it assumes a binary encoding of the dataset
def checkFitnessAllGridPoint(self, data_set):
print("\nCalculating fitness and storing wrong states")
self.data_set = data_set
size = self.data_set.getSize()
fit = size
wrong = []
x, y = self.data_set.x, self.data_set.y
y_eta = self.data_set.getYEta()
y_dim = self.data_set.getYDim()
estimation = self.nn.estimate(self.data_set.x)
# for binary
labels = np.array(y)
# predictions = np.array(np.round(estimation))
# invalid_flag = (labels[:,-1] == 0)
if(not self.non_det):
# check the control input prediction and invalid flag based on controller type (D or ND)
if(self.controller.con_det):
predictions = np.array(np.round(estimation))
# array of boolean with output flag equal to invalid control
invalid_flag = (labels[:,-1] == 0)
# check the control input prediction
same_inputs = (predictions[:,:-1] == labels[:,:-1]).all(axis = 1)
# check if the prediction and true output is equal
same_flag = (predictions[:,-1] == labels[:,-1])
else:
predictions = np.array(estimation)
invalid_flag = (labels[:,-1] == 1)
predictions_soft = np.zeros_like(predictions)
predictions_soft[np.arange(len(predictions)), predictions.argmax(1)] = 1
same_inputs = (labels[np.arange(len(labels)), predictions_soft.argmax(1)])
# check if the prediction flag and true output flag is equal
car_u = self.controller.input_total_gp
flag_one = np.logical_and(np.argmax(predictions_soft, 1) == (car_u), labels[:,-1] == 1)
flag_zero = np.logical_and(np.argmax(predictions_soft, 1) != (car_u), labels[:,-1] == 0)
same_flag = np.logical_or(flag_one, flag_zero)
# same_flag = (predictions[:,-1] == labels[:,-1])
early_check = np.logical_and(invalid_flag, same_flag)
# if it is same flag and same input, we have the performance here
same_all = np.logical_and(same_flag, same_inputs)
# we need to or to compensate incorrect prediction but does not matter because it is not winning domain
logic_all = np.logical_or(early_check, same_all)
# average the value to get fitness in the range of [0,1]
fitness = (np.mean(logic_all))
# valid inputs but wrong prediction
wrong_idx = np.where(np.logical_and((logic_all == 0), np.logical_not(invalid_flag)))
else:
predictions = np.array(np.round(estimation))
fitness = (np.mean(predictions == labels))
wrong_idx = np.where((np.logical_not(predictions == labels)).any(axis=1))
# get the index of wrong states
states = np.array(self.data_set.x)
if(self.data_set.encode == EncodeTypes.Classification):
wrong = np.squeeze(states[wrong_idx])
elif(self.data_set.encode == EncodeTypes.Boolean):
if(self.data_set.order == Ordering.Original):
wrong = states[wrong_idx]
wrong_temp = list(map(self.bed.baton, wrong))
wrong_x = list(map(self.controller.stox, wrong_temp))
wrong = []
for i in range(len(wrong_x)):
wrong.append([wrong_temp[i], wrong_x[i]])
else:
ss_dim = int(self.controller.state_space_dim)
wrong_states = states[wrong_idx]
if(wrong_states.size == 0):
return fitness, wrong
# len_one_dim = int(len(wrong_states[0])/ss_dim)
bit_dim = self.controller.bit_dim
temp = []
for i in range(ss_dim):
# bin_i = wrong_states[:,i*len_one_dim:(i+1)*len_one_dim]
bin_i = wrong_states[:,i*bit_dim[i]:(i+1)*bit_dim[i]]
array_i_add_dim = np.array(list(map(self.bed.baton, bin_i)))[:,None]
temp.append(array_i_add_dim)
wrong_temp = temp[0]
for i in range(ss_dim-1):
wrong_temp = np.concatenate((wrong_temp, temp[i+1]), axis = 1)
wrong_x = list(map(self.controller.sstox, wrong_temp))
wrong_s = list(map(self.controller.sstos, wrong_temp))
wrong = []
for i in range(len(wrong_s)):
wrong.append([wrong_s[i], wrong_x[i]])
return fitness, wrong
# Fitness modification for including as wel the non-winning domain to the NN
# the last bit as the valid flag change the calculation quite a lot
def allGridPointFitness(self, data_set):
self.data_set = data_set
size = self.data_set.getSize()
fit = size
wrong = []
x, y = self.data_set.x, self.data_set.y
y_eta = self.data_set.getYEta()
y_dim = self.data_set.getYDim()
estimation = self.nn.estimate(self.data_set.x)
# calculate the fitness by using np array to optimize computation
labels = np.array(y)
# predictions = np.array(np.round(estimation))
if(not self.non_det):
# check the control input prediction and invalid flag based on controller type (D or ND)
if(self.controller.con_det):
# print('deterministic')
predictions = np.array(np.round(estimation))
# array of boolean with output flag equal to invalid control
invalid_flag = (labels[:,-1] == 0)
# check the control input prediction
same_inputs = (predictions[:,:-1] == labels[:,:-1]).all(axis = 1)
# check if the prediction flag and true output flag is equal
same_flag = (predictions[:,-1] == labels[:,-1])
else:
# print('determinizing')
predictions = np.array(estimation)
invalid_flag = (labels[:,-1] == 1)
predictions_soft = np.zeros_like(predictions)
predictions_soft[np.arange(len(predictions)), predictions.argmax(1)] = 1
same_inputs = (labels[np.arange(len(labels)), predictions_soft.argmax(1)])
# check if the prediction flag and true output flag is equal
car_u = self.controller.input_total_gp
# print(car_u)
# check the correctness of the output prediction
# both for invalid and valid label
flag_one = np.logical_and(np.argmax(predictions_soft, 1) == (car_u), labels[:,-1] == 1)
flag_zero = np.logical_and(np.argmax(predictions_soft, 1) != (car_u), labels[:,-1] == 0)
same_flag = np.logical_or(flag_one, flag_zero)
# print(np.mean(same_flag))
# if it is invalid and it predicted right we have a flag that we do not have to check the rest of the bit
early_check = np.logical_and(invalid_flag, same_flag)
# if it is the same flag and same input, we have the performance here
same_all = np.logical_and(same_flag, same_inputs)
# we need to or to compensate incorrect prediction but does not matter because it is not winning domain
logic_all = np.logical_or(early_check, same_all)
# average the value to get fitness in the range of [0,1]
fitness = (np.mean(logic_all))
else:
# print('non-deterministic')
predictions = np.array(np.round(estimation))
fitness = (np.mean(predictions == labels))
# print(labels[:10])
# print(predictions[:10])
return fitness
# Fitness modification for including as wel the non-winning domain to the NN
# the last bit as the valid flag change the calculation quite a lot
def allGridPointFitnessDeterminizing(self, data_set):
self.data_set = data_set
size = self.data_set.getSize()
fit = size
wrong = []
x, y = self.data_set.x, self.data_set.y
y_eta = self.data_set.getYEta()
y_dim = self.data_set.getYDim()
estimation = self.nn.estimate(self.data_set.x)
# calculate the fitness by using np array to optimize computation
labels = np.array(y)
predictions = np.array(np.round(estimation))
# array of boolean with output flag equal to invalid control
invalid_flag = (labels[:,-1] == 1)
# check if the prediction and true output is equal
same_flag = (predictions[:,-1] == labels[:,-1])
# if it is invalid and it predicted right we have a flag that we do not have to check the rest of the bit
early_check = np.logical_and(invalid_flag, same_flag)
# TO DO branch condition regression and classification
# check the control input prediction
# same_inputs = (predictions[:,:-1] == labels[:,:-1]).all(axis = 1)
predictions_soft = np.zeros_like(predictions)
predictions_soft[np.arange(len(predictions)), predictions.argmax(1)] = 1
same_inputs = (labels[np.arange(len(labels)), predictions_soft.argmax(1)])
# if it is same flag and same input, we have the performance here
same_all = np.logical_and(same_flag, same_inputs)
# we need to or to compensate incorrect prediction but does not matter because it is not winning domain
logic_all = np.logical_or(early_check, same_all)
# average the value to get fitness in the range of [0,1]
fitness = (np.mean(logic_all))
return fitness
# Randomly check neural network against a dataset
def randomCheck(self, data_set):
self.data_set = data_set
self.initializeFitnessFunction()
print("\nValidating:")
for i in range(10):
r = round(random.random()*(self.data_set.getSize()-1))
self.checkByIndex(r, True)
# Train network
def train(self):
self.clear()
print("\nTraining (Ctrl+C to interrupt):")
signal.signal(signal.SIGINT, self.interrupt)
i, batch_index, loss, fit = 0,0,0,0.0
old_epoch = 0
stagnan = False
self.merged_summary = tf.summary.merge_all()
start_time = time.time()
while self.training:
batch = self.data_set.getBatch(self.batch_size, batch_index)
loss, summary = self.nn.trainStep(batch, self.merged_summary)
# if(i % self.shuffle_rate == 0 and i != 0): self.data_set.shuffle()
# if(fit > 0.999):
# fit = self.allGridPointFitness(self.data_set)
# self.batch_size = 4096
# if(i % self.display_step == 0 and i != 0):
if((self.epoch % self.display_step == 0) and (old_epoch != self.epoch)):
old_epoch = self.epoch
# fit = self.nn.runInSession(self.fitness, self.data_set.x, self.data_set.y, 1.0)
fit = self.allGridPointFitness(self.data_set)
# self.addToLog(loss, fit, i)
self.addToLog(loss, fit, self.epoch)
print("i = " + str(i) + "\tepoch = " + str(self.epoch) + "\tloss = " + str(float("{0:.4f}".format(loss))) + "\tfit = " + str(float("{0:.4f}".format(fit))))
self.train_writer.add_summary(summary, i)
if(self.epoch >= self.epoch_threshold and self.epoch_threshold > 0):
print("i = " + str(i) + "\tepoch = " + str(self.epoch) + "\tloss = " + str(float("{0:.4f}".format(loss))) + "\tfit = " + str(float("{0:.4f}".format(fit))))
print("Finished training, epoch threshold reached")
break
if(fit >= self.fitness_threshold):
print("Finished training")
break
if (len(self.fitnesses) > 40) and stagnan == False:
if((self.fitnesses[-1] - 0.001) <= self.fitnesses[-40]):
print("Finished training, fitness did not improve after "+str(40*self.display_step)+" epoch")
stagnan = True
if(math.isnan(loss)):
print("i = " + str(i) + "\tepoch = " + str(self.epoch) + "\tloss = " + str(float("{0:.3f}".format(loss))) + "\tfit = " + str(float("{0:.3f}".format(fit))))
print("Finished training, solution did not converge")
break
batch_index += self.batch_size
if(batch_index >= self.data_set.getSize()):
batch_index = batch_index % self.data_set.getSize()
self.data_set.shuffle()
self.epoch += 1
i += 1
end_time = time.time()
print("Time taken: " + self.utils.formatTime(end_time - start_time))
# Interrupt handler to interrupt the training while in progress
def interrupt(self, signal, frame):
self.training = False
# Plotting loss and fitness functions
def plot(self):
plt.figure(1)
plt.plot(self.iterations, self.losses, 'bo')
plt.xlabel("Iterations")
plt.ylabel("Loss")
plt.grid()
x1,x2,y1,y2 = plt.axis()
plt.axis((x1,x2,0,y2+0.1))
plt.figure(2)
plt.plot(self.iterations, self.fitnesses, 'r-')
plt.xlabel("Iterations")
plt.ylabel("Fitness")
plt.grid()
x1,x2,y1,y2 = plt.axis()
plt.axis((x1,x2,0,1))
plt.show()
# Add to log
def addToLog(self, loss, fit, iteration):
self.losses.append(loss)
self.fitnesses.append(fit)
self.iterations.append(iteration)
# Get projected data size
def getDataSize(self):
size = self.nn.calculateDataSize()
print("Minimal neural network size of: " + self.utils.formatBytes(size))
return size
# Clear variables
def clear(self):
self.epoch = 0
self.training = True
self.fitnesses = []
self.iterations = []
self.losses = []
# Save network
def save(self, filename):
print("\nSaving neural network")
self.nn.save(filename)
# Close session
def close(self):
self.nn.close()
self.train_writer.close()
# create loosing points to be plotted on matlab
def createLoosingPoints(self, wrong_states):
print("\nStoring loosing states for matlab simulation")
# print("\nLoop", self.controller.state_total_gp)
# for i in range(self.controller.state_total_gp):
#if i not in self.controller.states or i in wrong:
# if i not in self.controller.states:
# loosing_states.append(i)
total = set(range(self.controller.state_total_gp))
winning = set(self.controller.states)
loosing_states = total - winning
# print(len(total), len(winning), len(loosing_states))
# print("\nConvert to x")
loosing_states_x = list(map(self.controller.stox, loosing_states))
return loosing_states_x
class NeuralNetworkManager:
def __init__(self):
self.type = None
self.nn = MLP()
self.training_method = None
self.activation_function = None
self.dropout_rate = 0.0
self.training = True
self.learning_rate = 0.1
self.fitness_threshold = 0.75
self.epoch_threshold = -1
self.batch_size = 100
self.shuffle_rate = 2500
self.display_step = 1000
self.epoch = 0
self.layers = []
self.data_set = None
self.bed = BinaryEncoderDecoder()
self.utils = Utilities()
self.debug_mode = False
# Plotting variables
self.losses = []
self.fitnesses = []
self.iterations = []
self.save_location = './nn/log/'
# Getters and setters
def getType(self):
return self.type
def getTrainingMethod(self):
return self.training_method
def getActivationFunction(self):
return self.activation_function
def getLearningRate(self):
return self.learning_rate
def getFitnessThreshold(self):
return self.fitness_threshold
def getBatchSize(self):
return self.batch_size
def getDisplayStep(self):
return self.display_step
def getEpoch(self):
return self.epoch
def getEpochThreshold(self):
return self.epoch_threshold
def getDropoutRate(self):
return self.dropout_rate
def getShuffleRate(self):
return self.shuffle_rate
def getSaveLocation(self):
return self.save_location
def setType(self, type):
self.type = type
def setTrainingMethod(self, optimizer):
self.training_method = optimizer
def setActivationFunction(self, activation_function):
self.activation_function = activation_function
def setLearningRate(self, value):
self.learning_rate = value
def setFitnessThreshold(self, value):
self.fitness_threshold = value
def setBatchSize(self, value):
self.batch_size = value
def setDisplayStep(self, value):
self.display_step = value
def setEpochThreshold(self, value):
self.epoch_threshold = value
def setDropoutRate(self, value):
self.dropout_rate = value
def setShuffleRate(self, value):
self.shuffle_rate = value
def setSaveLocation(self, value):
self.save_location = value
def setDebugMode(self, value):
self.debug_mode = value
def setDataSet(self, data_set):
self.data_set = data_set
# Hidden layer generation functions
# Linearly increase/decrease neurons per hidden layer based on the input and ouput neurons
def linearHiddenLayers(self, num_hidden_layers):
self.layers = []
x_dim = self.data_set.getXDim()
y_dim = self.data_set.getYDim()
a = (y_dim - x_dim) / (num_hidden_layers + 1)
self.layers.append(x_dim)
for i in range(1, num_hidden_layers + 1):
self.layers.append(round(x_dim + a * i))
self.layers.append(y_dim)
return self.layers
# Rectangular hidden layer
def rectangularHiddenLayers(self, width, height):
self.layers = []
self.layers.append(self.data_set.getXDim())
for i in range(width):
self.layers.append(height)
self.layers.append(self.data_set.getYDim())
#Customize layer sturcture
def customHiddenLayers(self, layer):
self.layers = []
x_dim = self.data_set.getXDim()
y_dim = self.data_set.getYDim()
self.layers.append(x_dim)
for i in range(1, len(layer) + 1):
self.layers.append(layer[i - 1])
self.layers.append(y_dim)
return self.layers
# Initialize neural network
def initializeNeuralNetwork(self):
if (self.debug_mode):
print("\nNeural network initialization:")
if (self.type == NNTypes.MLP):
self.nn = MLP()
self.nn.setDebugMode(self.debug_mode)
if (self.debug_mode):
print("Neural network type: MLP")
# Initialize network and loss function
self.nn.setNeurons(self.layers)
self.nn.setDropoutRate(self.dropout_rate)
self.nn.setActivationFunction(self.activation_function)
self.nn.initializeNetwork()
# Print neural network status
if (self.debug_mode):
print("Generated network neuron topology: " + str(self.layers) +
" with dropout rate: " + str(self.nn.getDropoutRate()))
# Initialize training function
def initializeTraining(self,
learning_rate,
fitness_threshold,
batch_size,
display_step,
epoch_threshold=-1,
shuffle_rate=10000):
self.learning_rate = learning_rate
self.fitness_threshold = fitness_threshold
self.batch_size = batch_size
self.display_step = display_step
self.epoch_threshold = epoch_threshold
self.shuffle_rate = shuffle_rate
self.nn.initializeLossFunction()
self.nn.initializeTrainFunction(self.training_method,
self.learning_rate)
# Initialize fitness function
def initializeFitnessFunction(self):
with tf.name_scope("fitness"):
eta = self.data_set.getYEta()
size = self.data_set.getSize()
lower_bound = tf.subtract(self.nn.y, eta)
upper_bound = tf.add(self.nn.y, eta)
is_fit = tf.logical_and(
tf.greater_equal(self.nn.predictor, lower_bound),
tf.less(self.nn.predictor, upper_bound))
non_zero = tf.to_float(
tf.count_nonzero(tf.reduce_min(tf.cast(is_fit, tf.int8), 1)))
self.fitness = non_zero / size
tf.summary.scalar("fitness", self.fitness)
# General initialization function to call all functions
def initialize(self,
learning_rate,
fitness_threshold,
batch_size,
display_step,
epoch_threshold=-1,
shuffle_rate=10000):
self.initializeNeuralNetwork()
self.initializeFitnessFunction()
self.initializeTraining(learning_rate, fitness_threshold, batch_size,
display_step, epoch_threshold, shuffle_rate)
self.train_writer = tf.summary.FileWriter(self.save_location,
self.nn.session.graph)
# Check a state against the dataset and nn by using its id in the dataset
def checkByIndex(self, index, out):
x = self.data_set.x[index]
estimation = self.nn.estimate([x])[0]
y = self.data_set.getY(index)
y_eta = self.data_set.getYEta()
equal = True
for i in range(self.data_set.getYDim()):
if (not ((y[i] - y_eta[i]) <= estimation[i] and
(y[i] + y_eta[i]) > estimation[i])):
equal = False
if (out):
print("u: " + str(y) + " u_: " + str(numpy.round(estimation, 2)) +
" within etas: " + str(equal))
return equal
# Check fitness of the neural network for a specific dataset and return wrong states
# as of right now it assumes a binary encoding of the dataset
def checkFitness(self, data_set):
self.data_set = data_set
size = self.data_set.getSize()
fit = size
wrong = []
x, y = self.data_set.x, self.data_set.y
y_eta = self.data_set.getYEta()
y_dim = self.data_set.getYDim()
estimation = self.nn.estimate(self.data_set.x)
for i in range(size):
equal = True
for j in range(y_dim):
if (not ((y[i][j] - y_eta[j]) <= estimation[i][j] and
(y[i][j] + y_eta[j]) > estimation[i][j]) and equal):
wrong.append(self.bed.baton(x[i]))
fit -= 1
equal = False
fitness = fit / size * 100
print("\nDataset fitness: " + str(float("{0:.3f}".format(fitness))) +
"%")
return fitness, wrong
# Randomly check neural network against a dataset
def randomCheck(self, data_set):
self.data_set = data_set
self.initializeFitnessFunction()
print("\nValidating:")
for i in range(10):
r = round(random.random() * (self.data_set.getSize() - 1))
self.checkByIndex(r, True)
# Train network
def train(self):
self.clear()
print("\nTraining (Ctrl+C to interrupt):")
signal.signal(signal.SIGINT, self.interrupt)
i, batch_index, loss, fit = 0, 0, 0, 0.0
self.merged_summary = tf.summary.merge_all()
start_time = time.time()
while self.training:
batch = self.data_set.getBatch(self.batch_size, batch_index)
loss, summary = self.nn.trainStep(batch, self.merged_summary)
if (i % self.shuffle_rate == 0 and i != 0): self.data_set.shuffle()
if (i % self.display_step == 0 and i != 0):
fit = self.nn.runInSession(self.fitness, self.data_set.x,
self.data_set.y, 1.0)
self.addToLog(loss, fit, i)
print("i = " + str(i) + "\tepoch = " + str(self.epoch) +
"\tloss = " + str(float("{0:.3f}".format(loss))) +
"\tfit = " + str(float("{0:.3f}".format(fit))))
self.train_writer.add_summary(summary, i)
if (self.epoch >= self.epoch_threshold
and self.epoch_threshold > 0):
print("i = " + str(i) + "\tepoch = " + str(self.epoch) +
"\tloss = " + str(float("{0:.3f}".format(loss))) +
"\tfit = " + str(float("{0:.3f}".format(fit))))
print("Finished training, epoch threshold reached")
break
if (fit >= self.fitness_threshold):
print("Finished training")
break
if (math.isnan(loss)):
print("i = " + str(i) + "\tepoch = " + str(self.epoch) +
"\tloss = " + str(float("{0:.3f}".format(loss))) +
"\tfit = " + str(float("{0:.3f}".format(fit))))
print("Finished training, solution did not converge")
break
batch_index += self.batch_size
if (batch_index >= self.data_set.getSize()):
batch_index = batch_index % self.data_set.getSize()
self.epoch += 1
i += 1
end_time = time.time()
print("Time taken: " + self.utils.formatTime(end_time - start_time))
# Interrupt handler to interrupt the training while in progress
def interrupt(self, signal, frame):
self.training = False
# Plotting loss and fitness functions
def plot(self):
plt.figure(1)
plt.plot(self.iterations, self.losses, 'bo')
plt.xlabel("Iterations")
plt.ylabel("Loss")
plt.grid()
x1, x2, y1, y2 = plt.axis()
plt.axis((x1, x2, 0, y2 + 0.1))
plt.figure(2)
plt.plot(self.iterations, self.fitnesses, 'r-')
plt.xlabel("Iterations")
plt.ylabel("Fitness")
plt.grid()
x1, x2, y1, y2 = plt.axis()
plt.axis((x1, x2, 0, 1))
plt.show()
# Add to log
def addToLog(self, loss, fit, iteration):
self.losses.append(loss)
self.fitnesses.append(fit)
self.iterations.append(iteration)
# Get projected data size
def getDataSize(self):
size = self.nn.calculateDataSize()
print("Minimal neural network size of: " +
self.utils.formatBytes(size))
return size
# Clear variables
def clear(self):
self.epoch = 0
self.training = True
self.fitnesses = []
self.iterations = []
self.losses = []
# Save network
def save(self, filename):
print("\nSaving neural network")
self.nn.save(filename)
# Close session
def close(self):
self.nn.close()
self.train_writer.close()
