def train(nnet, max_epochs, test_epoch_modulo=10, plot=False):
"""step through the training process manually (since we want to pull some
stats between epochs"""
i = 0
for epoch in xrange(max_epochs):
nnet.train()
NNUtils.save_nn_obj_to_file(nnet, './tmpnetwork.pkl')
if i % test_epoch_modulo == 0:
nnet.print_network_state()
(estimated, actual,) = test()
if plot != False:
directory = 'plots.%s' % suffix
if not os.path.exists(directory):
os.mkdir(directory)
filepath = '%s/%s.png' % (directory, str(nnet.epoch_i).zfill(6))
title = '%s\nepoch %s' % (nnet.name, i)
NNUtils.xy_prediction_plot(actual, estimated, filepath=filepath,
title=title,
x1_label='Actual Values',
x2_label='Predicted Values',
x1_axis_label='months into the future',
x2_axis_label='monthly sunspot count',
y_minmax=(0,90))
i += 1
def test():
nnet = NNUtils.load_nn_obj_from_file('./tmpnetwork.pkl')
scale_min, scale_max = nnet.get_scale_minmax()
x1 = []
x2 = []
for sample_n in xrange(TEST_SAMPLE_N):
inputs = nnet.get_inputs()
outputs = nnet.get_outputs()
x1.extend(NNUtils.denormalize(outputs, scale_min, scale_max))
validations = nnet.get_validation_vectors()[sample_n][1]
x2.extend(NNUtils.denormalize(validations, scale_min, scale_max))
actual_value = NNUtils.denormalize(
nnet.get_validation_vectors()[sample_n][1],
scale_min, scale_max)[0]
#predicted_value = NNUtils.denormalize(outputs, scale_min, scale_max)[0]
#print '%s,%s' % (predicted_value, actual_value,)
inputs = inputs[len(outputs):]
inputs.extend(nnet.get_outputs())
nnet.set_inputs(inputs)
nnet.feed_forward()
return (x1, x2)
def main():
path = 'input/training.txt'
# create a simple 2x2x1 network.
nnet = NN.NeuralNetwork(name='Example XOR Network')
nnet.set_learning_rate(0.4)
nnet.create_network_architecture(n_inputs=2, n_hiddens=2, n_outputs=1)
# train the network
(vectors, scale_min, scale_max) = NNUtils.parse_training_file(path, 2, 1)
nnet.set_training_vectors(vectors, scale_min, scale_max)
nnet.training_loop(tsse_threshold=.005)
#nnet.training_loop(epochs_threshold=9999)
print '-'*80
# test the network
# - input values that are NOT xor'ed should put the output value 'low'
# - input values that ARE xor'ed should put the output value 'high'
for vector in [[0,0], [0,1], [1,0], [1,1]]:
nnet.set_inputs([vector[0], vector[1]])
nnet.feed_forward()
print 'NN test:', vector, '==', nnet.get_outputs()[0]
nnet.print_network_state()
print '-'*80
print nnet
# plot the activation due to a bunch of numbers 0-1
# 1) uncomment save_network_state and save_network_state_iteration_modulo
# in the above constructor.
# 2) uncomment the following 5 lines of code
"""
def main():
#N_INPUTS = 365
#N_HIDDENS = 365
N_INPUTS = 12
N_HIDDENS = 12
N_OUTPUTS = 1
LEARNING_RATE = 0.1
NAME = 'Example Sunspot Prediction Network'
nnet = NN.NeuralNetwork(name=NAME)
nnet.set_learning_rate(LEARNING_RATE)
(all_vectors, scale_min, scale_max) = NNUtils.parse_windowed_training_file(
TRAININGPATH,
N_INPUTS,
N_OUTPUTS,
1)
# take TEST_SAMPLE_N samples from the training set and use this set to
# measure the neural network's predictions against factual values.
cut_index = len(all_vectors) - TEST_SAMPLE_N
training_vectors = all_vectors[:cut_index]
validation_vectors = all_vectors[cut_index:]
nnet.set_validation_vectors(validation_vectors)
nnet.set_training_vectors(training_vectors, scale_min, scale_max)
nnet.create_network_architecture(n_inputs=N_INPUTS, n_hiddens=N_HIDDENS,
n_outputs=N_OUTPUTS)
train(nnet, MAX_EPOCHS, plot=PLOTTING_ENABLED, test_epoch_modulo=24)
