def performNN(all_extracted_features, all_targets):
from pyneurgen.neuralnet import NeuralNet
#from pyneurgen.nodes import BiasNode, Connection
net = NeuralNet()
net.init_layers(len(all_extracted_features[0]), [2], 1)
net.randomize_network()
net.set_halt_on_extremes(True)
# Set to constrain beginning weights to -5 to 5
# Just to show we can
#net.set_random_constraint(.5)
net.set_learnrate(.001)
net.set_all_inputs(all_extracted_features)
net.set_all_targets(all_targets)
length = len(all_extracted_features)
learn_end_point = int(length * .8)
net.set_learn_range(0, learn_end_point)
net.set_test_range(learn_end_point + 1, length - 1)
net.layers[1].set_activation_type('tanh')
net.learn(epochs=150, show_epoch_results=True, random_testing=True)
mse = net.test()
print mse
net.set_all_targets(all_targets)
net.set_learn_range(indexes_learn)
net.get_learn_range()
net.set_test_range(indexes_test)
net.get_test_range()
net.layers[0].set_activation_type('tanh')
net.layers[1].set_activation_type('tanh')
net.layers[2].set_activation_type('tanh')
net.layers[3].set_activation_type('tanh')
###training network
net.learn(epochs=1200, show_epoch_results=True,
random_testing=True)
mse = net.test()
#extract predicted velues
all_learn = [item[1][0] for item in net.get_learn_data()]
learn_positions = [item[0][3] for item in net.get_learn_data()]
test_positions = [item[0][3] for item in net.get_test_data()]
all_targets1 = [item[0][0] for item in net.test_targets_activations]
allactuals = [item[1][0] for item in net.test_targets_activations]
# This is quick and dirty, but it will show the results
subplot(3, 1, 1)
plot(learn_positions, all_learn,'bo')
title("all_targets")
grid(True)
x_input = np.concatenate(
(x_train, x_test, np.zeros((1, np.shape(x_train)[1]))))
y_input = np.concatenate((y_train, y_test, np.zeros((1, 1))))
#elaboracao do modelo de rede neural com os parametros definidos
fit1 = NeuralNet()
fit1.init_layers(input_nodes, [hidden_nodes], output_nodes,
ElmanSimpleRecurrent())
fit1.randomize_network()
fit1.layers[1].set_activation_type('sigmoid')
fit1.set_learnrate(0.05)
fit1.set_all_inputs(x_input)
fit1.set_all_targets(y_input)
fit1.set_learn_range(0, i)
fit1.set_test_range(i, i + 1)
fit1.learn(epochs=100, show_epoch_results=True, random_testing=False)
mse = fit1.test()
all_mse.append(mse)
print("test set MSE = ", np.round(mse, 6))
target = [item[0][0] for item in fit1.test_targets_activations]
target = scaler_y.inverse_transform(
np.array(target).reshape((len(target), 1)))
pred = [item[1][0] for item in fit1.test_targets_activations]
pred = scaler_y.inverse_transform(np.array(pred).reshape((len(pred), 1)))
real_y_test.append(target[0][0])
predicted_y_test.append(pred[0][0])
filehandler = open('objects/elman/el_' + str(i) + '.obj', 'w')
pickle.dump(fit1, filehandler)
filehandler.close()
#calculo do erro dos minimos quadrados
total_mse = mean_squared_error(real_y_test, predicted_y_test)
net.set_all_inputs(all_inputs)
net.set_all_targets(all_targets)
length = len(all_inputs)
learn_end_point = int(length * .8)
net.set_learn_range(0, learn_end_point)
net.set_test_range(learn_end_point + 1, length - 1)
net.layers[1].set_activation_type('tanh')
net.learn(epochs=125, show_epoch_results=True,
random_testing=False)
mse = net.test()
test_positions = [item[0][1] * 1000.0 for item in net.get_test_data()]
all_targets1 = [item[0][0] for item in net.test_targets_activations]
allactuals = [item[1][0] for item in net.test_targets_activations]
fig = plt.figure()
ax1 = fig.add_subplot(311)
ax1.plot([i[1] for i in population])
ax1.set_title("Population")
ax1.grid(True)
ax2 = fig.add_subplot(312)
ax2.plot(test_positions, all_targets1, 'bo', label='targets')
ax2.plot(test_positions, allactuals, 'ro', label='actuals')
input_nodes = 1
hidden_nodes = 5
output_nodes = 1
output_order = 20
incoming_weight_from_output = .5
input_order = 20
incoming_weight_from_input = .5
net = NeuralNet()
net.init_layers(
input_nodes, [hidden_nodes], output_nodes,
NARXRecurrent(output_order, incoming_weight_from_output, input_order,
incoming_weight_from_input))
net.randomize_network()
X = np.linspace(0, 10.0, num=10001)
Y = simpleWeierstrassTimeSeries(X)
Y = Y.reshape(-1, 1)
net.set_all_inputs(Y[:-1])
net.set_all_targets(Y[1:])
net.set_learn_range(0, 8000)
net.set_test_range(8000, 9999)
print net.test()
net.learn(epochs=5)
print net.test()
def serNeural(sDay,nAhead,x0,hWeek):
nLin = sDay.shape[0] + nAhead
nFit = sDay.shape[0] if int(x0['obs_time']) <= 14 else int(x0['obs_time'])
predS = getHistory(sDay,nAhead,x0,hWeek)
weekS = [x.isocalendar()[1] for x in sDay.index]
population = [[float(i),sDay['y'][i],float(i%7),weekS[i]] for i in range(sDay.shape[0])]
all_inputs = []
all_targets = []
factorY = sDay['y'].mean()
factorT = 1.0 / float(len(population))*factorY
factorD = 1./7.*factorY
factorW = 1./52.*factorY
factorS = 4.*sDay['y'].std()
factorH = factorY/sDay['hist'].mean()
def population_gen(population):
pop_sort = [item for item in population]
# random.shuffle(pop_sort)
for item in pop_sort:
yield item
for t,y,y1,y2 in population_gen(population):
#all_inputs.append([t*factorT,(.5-random.random())*factorS+factorY,y1*factorD,y2*factorW])
all_inputs.append([y1*factorD,(.5-random.random())*factorS+factorY,y2*factorW])
all_targets.append([y])
if False:
plt.plot([x[0] for x in all_inputs],'-',label='targets0')
plt.plot([x[1] for x in all_inputs],'-',label='targets1')
plt.plot([x[2] for x in all_inputs],'-',label='targets2')
# plt.plot([x[3] for x in all_inputs],'-',label='targets3')
plt.plot([x[0] for x in all_targets],'-',label='actuals')
plt.legend(loc='lower left', numpoints=1)
plt.show()
net = NeuralNet()
net.init_layers(3,[10],1,NARXRecurrent(3,.6,2,.4))
net.randomize_network()
net.set_random_constraint(.5)
net.set_learnrate(.1)
net.set_all_inputs(all_inputs)
net.set_all_targets(all_targets)
#predS['pred'] = [item[0][0] for item in net.test_targets_activations]
length = len(all_inputs)
learn_end_point = int(length * .8)
# random.sample(all_inputs,10)
net.set_learn_range(0, learn_end_point)
net.set_test_range(learn_end_point + 1, length - 1)
net.layers[1].set_activation_type('tanh')
net.learn(epochs=125,show_epoch_results=True,random_testing=False)
mse = net.test()
#net.save(os.environ['LAV_DIR'] + "/out/train/net.txt")
test_positions = [item[0][0] for item in net.get_test_data()]
all_targets1 = [item[0][0] for item in net.test_targets_activations]
all_actuals = [item[1][0] for item in net.test_targets_activations]
# This is quick and dirty, but it will show the results
plt.subplot(3, 1, 1)
plt.plot([i for i in sDay['y']],'-')
plt.title("Population")
plt.grid(True)
plt.subplot(3, 1, 2)
plt.plot(test_positions, all_targets1, 'b-', label='targets')
plt.plot(test_positions, all_actuals, 'r-', label='actuals')
plt.grid(True)
plt.legend(loc='lower left', numpoints=1)
plt.title("Test Target Points vs Actual Points")
plt.subplot(3, 1, 3)
plt.plot(range(1, len(net.accum_mse) + 1, 1), net.accum_mse)
plt.xlabel('epochs')
plt.ylabel('mean squared error')
plt.grid(True)
plt.title("Mean Squared Error by Epoch")
plt.show()
NARXRecurrent(output_order, incoming_weight_from_output, input_order,
incoming_weight_from_input))
fit1.randomize_network()
fit1.layers[1].set_activation_type('sigmoid')
fit1.set_learnrate(0.35)
fit1.set_all_inputs(x)
fit1.set_all_targets(y)
length = len(x)
learn_end_point = int(length * 0.85)
fit1.set_learn_range(0, learn_end_point)
fit1.set_test_range(learn_end_point + 1, length - 1)
fit1.learn(epochs=10, show_epoch_results=True, random_testing=False)
mse = fit1.test()
print("MSE for test set: ", round(mse, 6))
plt.figure(figsize=(15, 6))
plt.plot(np.arange(len(fit1.accum_mse)), fit1.accum_mse)
plt.xlabel('Epochs')
plt.ylabel('Mean Squared Error')
plt.savefig('../figs/fig9.png')
yhat = [i[1][0] for i in fit1.test_targets_activations]
yhat = scaler.inverse_transform(np.array(yhat).reshape((len(yhat), 1)))
yhat = yhat.flatten()
l = len(yhat)
yend = y[-l:]
time = date[-l:]
