def simpleTrain(self, net=None, nConvergence=0):
if (net == None):
net = self.net
if (nConvergence == 0):
nConvergence = self.nIters
Network.train(net, nIters=self.nIters, nConvergence=nConvergence)
return self.summary(net)
def progressiveTrain(self, net=None):
if (net == None):
net = self.net
for train in self.trainPlane:
net.unfreezAllTrainableVar()
for layer in train["freez_layers"]:
net.freezTrainableVar(layer)
Network.train(net, nIters=train["n_iters"])
return self.summary(net)
def __init__(self, n_of_inputs, hidden_layers, n_of_outputs,max_error):
rev = hidden_layers
rev = rev[::-1]
hidden_layers.append(n_of_outputs)
hidden_layers.extend(rev)
self.autoencoder = Network(n_of_inputs, hidden_layers, n_of_inputs,max_error)
self.n_of_inputs = n_of_inputs
self.n_of_outputs = n_of_outputs
self.hl = hidden_layers
self.neurons_of_layer = [self.n_of_inputs] + self.hl + [self.n_of_inputs]
print("neurons_of_layer:", self.neurons_of_layer)
def readMyFile(filename):
data = []
labels = []
with open(filename) as csvDataFile:
csvReader = csv.reader(csvDataFile)
for row in csvReader:
n = row.__len__()
data.append([row[index] for index in range(0, n - 1)])
labels.append(row[n - 1])
print(data)
print(labels)
return np.asarray(data, dtype=float32), np.asarray(labels, dtype=float32)
# # Ricordare di cambiare i path dentro la radice che sceglierete dovrete creare la cartella config con dentro i json, consiglio di lasciare tutto come lo fornisco io
ROOT_DIR = os.path.abspath(os.curdir)
os.chdir(ROOT_DIR)
jsonStructureDir = ROOT_DIR + "\config\structure.txt"
jsonTrainDir = ROOT_DIR + "\config\\train.txt"
jsonProgressiveTrainDir = ROOT_DIR + "\config\\progressiveTrain.txt"
#2 esempio di test classe Trainer + Network (simple train) (file "structure.txt" +train.txt)
trainer = Trainer(jsonTrainDir)
# prendo un dataset di test dalla repository di sklearn
data, labels = readMyFile(trainer.path)
xTrain, xTest, yTrain, yTest = trainer.split(data, labels)
netTest_2 = Network(jsonStructureDir, xTrain, yTrain)
trainer.simpleTrain(netTest_2)
netTest_2.saveNetwork()
print(data)
print(labels)
return np.asarray(data, dtype=float32), np.asarray(labels, dtype=float32)
# # Ricordare di cambiare i path dentro la radice che sceglierete dovrete creare la cartella config con dentro i json, consiglio di lasciare tutto come lo fornisco io
ROOT_DIR = os.path.abspath(os.curdir)
os.chdir(ROOT_DIR)
jsonStructureDir = ROOT_DIR + "\config\structure.txt"
jsonTrainDir = ROOT_DIR + "\config\\train.txt"
jsonProgressiveTrainDir = ROOT_DIR + "\config\\progressiveTrain.txt"
#1 esempio di test: classe Network stand-alone effettua solo un training di tutta la rete secondo le informazioni contenute nel json "structure.txt"
datasetPath = 'xor_test_100.csv'
data, labels = readMyFile(datasetPath)
netTest_1 = Network(jsonStructureDir, [], [], data, labels)
Network.train(netTest_1)
predProb = netTest_1.predict()
yHat = np.where(predProb < 0.5, 0, 1)
acc, trueVector, oneHotYTest = netTest_1.acc(yHat)
netTest_1.saveNetwork()
print("Test Accuracy %.2f" % acc)
#2 esempio di test classe Trainer + Network (simple train) (file "structure.txt" +train.txt)
trainer = Trainer(jsonTrainDir)
# prendo un dataset di test dalla repository di sklearn
data, labels = readMyFile(trainer.path)
xTrain, xTest, yTrain, yTest = trainer.split(data, labels)
netTest_2 = Network(jsonStructureDir, xTrain, yTrain)
trainer.simpleTrain(netTest_2)
netTest_2.saveNetwork()
return np.asarray(data, dtype=float32), np.asarray(labels, dtype=float32)
# # Ricordare di cambiare i path dentro la radice che sceglierete dovrete creare la cartella config con dentro i json, consiglio di lasciare tutto come lo fornisco io
ROOT_DIR = os.path.abspath(os.curdir)
os.chdir(ROOT_DIR)
jsonStructureDir = ROOT_DIR + "\config\structure.txt"
jsonTrainDir = ROOT_DIR + "\config\\train.txt"
jsonProgressiveTrainDir = ROOT_DIR + "\config\\progressiveTrain.txt"
#1 esempio di test: classe Network stand-alone effettua solo un training di tutta la rete secondo le informazioni contenute nel json "structure.txt"
datasetPath = 'xor_test_100.csv'
data,labels = readMyFile(datasetPath)
netTest_1 = Network(jsonStructureDir, [],[], data, labels)
Network.train(netTest_1)
predProb = netTest_1.predict()
yHat = np.where(predProb < 0.5, 0, 1)
acc, trueVector, oneHotYTest = netTest_1.acc(yHat)
netTest_1.saveNetwork()
print("Test Accuracy %.2f" % acc)
# #2 esempio di test classe Trainer + Network (simple train) (file "structure.txt" +train.txt)
# trainer = Trainer(jsonTrainDir)
# # prendo un dataset di test dalla repository di sklearn
# data,labels = readMyFile(trainer.path)
# xTrain, xTest, yTrain, yTest = trainer.split( data, labels)
# netTest_2 = Network(jsonStructureDir,xTrain,yTrain)
# trainer.simpleTrain(netTest_2)
def readMyFile(filename):
data = []
labels = []
with open(filename) as csvDataFile:
csvReader = csv.reader(csvDataFile)
for row in csvReader:
n = row.__len__()
data.append([row[index] for index in range(0, n - 1)])
labels.append(row[n - 1])
print(data)
print(labels)
return np.asarray(data, dtype=float32), np.asarray(labels, dtype=float32)
# # Ricordare di cambiare i path dentro la radice che sceglierete dovrete creare la cartella config con dentro i json, consiglio di lasciare tutto come lo fornisco io
ROOT_DIR = os.path.abspath(os.curdir)
os.chdir(ROOT_DIR)
jsonStructureDir = ROOT_DIR + "\config\structure.txt"
jsonTrainDir = ROOT_DIR + "\config\\train.txt"
#2 esempio di test classe Trainer + Network (simple train) (file "structure.txt" +train.txt)
trainer = Trainer(jsonTrainDir)
data, labels = readMyFile(trainer.path)
xTrain, xTest, yTrain, yTest = trainer.split(data, labels)
#8 Uso della rete salvata, per usare una rete salvata basta mettere il suo nome nel file jsonStructureDir e istanzianziarla con rebuild
netTest_6 = Network(jsonStructureDir, rebuild=True)
predProb = netTest_6.predict(xTest)
yHat = np.where(predProb < 0.5, 0, 1)
acc, trueVector, oneHotYTest = netTest_6.acc(yHat, yTest)
print("Test Accuracy %.2f" % acc)
# Turn training and testing datasets into lists X_train_numerical = df_train.values X_train_binary = df_train_missing.values X_test_numerical = df_test.values X_test_binary = df_test_missing.values # Helper functions for pre and post processing the output min_y, max_y = min(y_train), max(y_train) min_max_y = lambda a: (a-min_y) / (max_y-min_y) reverse_y = lambda a: a * (max_y - min_y) + min_y """ NETWORK SETUP """ # Create the network models: 2 sensor, 1 decision # Note: existing architecture may be loaded using the load method from MLP library numerical_sensor = Network(architecture=[Layer(11), Layer(11, 'sigmoid'), Layer(1, 'sigmoid')], l_rate=0.001) #Numerical Sensor binary_sensor = Network(architecture=[Layer(11), Layer(4, 'sigmoid'), Layer(1, 'sigmoid')], l_rate=0.001) #Binary Sensor decision = Network(architecture=[Layer(2), Layer(1, 'linear')], l_rate=0.001) #Decision MLP # Initialize the networks (this generates the initial weights) numerical_sensor.initialize() binary_sensor.initialize() decision.initialize() # Uncomment below for a small batch test # X_train_numerical = X_train_numerical[:600] # X_train_binary = X_train_binary[:600] # y_train = y_train[:600] # X_test_numerical = X_test_numerical[:300] # X_test_binary = X_test_binary[:300] # y_test = y_test[:300]
def createNet(self, X, Y, jsonStructureDir):
self.split(X, Y)
self.net = Network(jsonStructureDir, x=self.xTrain, y=self.yTrain)
return self.net
# TRAIN-TEST SPLIT
X_train, X_test, y_train, y_test = train_test_split(X,
y,
test_size=0.20,
random_state=1)
# MIN-MAX Output SCALING
min_y, max_y = min(y_train), max(
y_train) #keeping track of min and max to unscale the predicted output
y_train = [(y - min_y) / (max_y - min_y) for y in y_train]
print('> Data Preprocessed and Scaled')
'''NEURAL NETWORK'''
# Architecture, activation functions, and learning rate setup
net = Network(architecture=[Layer(11),
Layer(11, 'tanh'),
Layer(1, 'sigmoid')],
l_rate=0.1)
print('> Network setup.')
net.initialize() #network initilization
print('> Network initialized.')
# Training
trial_error = net.train(X_train,
y_train,
epochs=50,
verbose=True,
plot=True,
decay=0.1,
decay_rate=10)
print('> Training complete.')
class Autoencoder:
def __init__(self, n_of_inputs, hidden_layers, n_of_outputs,max_error):
rev = hidden_layers
rev = rev[::-1]
hidden_layers.append(n_of_outputs)
hidden_layers.extend(rev)
self.autoencoder = Network(n_of_inputs, hidden_layers, n_of_inputs,max_error)
self.n_of_inputs = n_of_inputs
self.n_of_outputs = n_of_outputs
self.hl = hidden_layers
self.neurons_of_layer = [self.n_of_inputs] + self.hl + [self.n_of_inputs]
print("neurons_of_layer:", self.neurons_of_layer)
# Epochs: cantidad de epocas a entrenar
# eta: learning rate
# K: si el error decrece K veces, entonces eta = eta + a*eta
# Q: si el error crece Q veces, entonces eta = eta + b*eta
# adaptative_lr: activa o desactiva el decaimiento exponencial del eta
def train(self, inputs, outputs, epochs, eta,K,a,Q,b, adaptive_lr=False):
self.autoencoder.train(inputs, outputs, epochs, eta,K,a,Q,b, adaptive_lr)
def decode(self, input_):
# return self.autoencoder.predict(input_)
weights = self.autoencoder.weights
weights = weights[int(len(weights) / 2):]
b = self.autoencoder.biases
b = b[int(len(b) / 2):]
neurons_of_layer = self.neurons_of_layer[int(len(self.neurons_of_layer) / 2):]
activations = []
for i in range(len(neurons_of_layer)):
a = np.zeros(neurons_of_layer[i])
activations.append(a)
activations[0] = input_
for i, w in enumerate(weights):
x = np.dot(w.T, input_) + b[i].T
x = x.reshape(x.shape[1])
input_ = sigmoid(x)
activations[i + 1] = input_
return activations[-1]
def encode(self, input_):
weights = self.autoencoder.weights
weights = weights[:int(len(weights) / 2)]
b = self.autoencoder.biases
b = b[:int(len(b) / 2)]
neurons_of_layer = self.neurons_of_layer[:int(len(self.neurons_of_layer) / 2) + 1]
activations = []
for i in range(len(neurons_of_layer)):
a = np.zeros(neurons_of_layer[i])
activations.append(a)
activations[0] = input_
for i, w in enumerate(weights):
x = np.dot(w.T, input_) + b[i].T
x = x.reshape(x.shape[1])
input_ = sigmoid(x)
activations[i + 1] = input_
return activations[-1]