plt.grid(True)
labels = map(str, range(2,11))
xlocations = na.array(range(len(points))) + .5
width = .5
plt.bar(xlocations, points, width=width)
plt.xticks(xlocations + width/2, labels)
plt.savefig(file_name)
plt.clf()
log_file = open('backpropagation.log', 'w')
for training_set, ts_name in training_sets:
training_set = load_training_set(training_set)
print '.. begining %s' % ts_name
log_file.write('.. begining %s\n' % ts_name)
lowest_error = maxint
total_error_log = []
for hs in range(2, 11):
print ' %s hidden neurons' % hs
nn = NeuralNetwork(2, hs, 1)
error_log = training(nn, training_set, max_iterations=2000)
test_log, total_error = test(nn, get_random_set(10000))
log_file.write('%s neurons: %s errors \nhidden: %s\noutputs: %s\n\n' % (hs, total_error, [hn.weights for hn in nn.hidden], [on.weights for on in nn.outputs]))
total_error_log.append((hs, total_error, error_log, test_log))
best = min(total_error_log, key=itemgetter(1))
print ' lowest error on %s hidden neurons network: %s failures' % (best[0], best[1])
plot(best[2], best[3], save='media/%s-%s-best-performance.png' % (ts_name, best[0]))
bar_plot(map(itemgetter(1), total_error_log), 'media/error-derivation-%s' % ts_name)
log_file.close()
if __name__ == "__main__":
batch, output = 10, 3
train_percent = 80
episodes = int(1e3)
learning_rate = 1e-2
encoding = [("setosa", 0), ("versicolor", 1), ("virginica", 2)]
X, y = network.extract("iris.csv",
encoding=encoding,
output=output,
label="species")
dimension = len(X[0])
neurons = [dimension, 10, 20, 20, 5, output]
trainer, tester, validater = network.partition(X, y, output, batch,
train_percent)
activity = ["relu", "relu", "relu", "relu", ""]
cost = "crossentropy"
optimizer = "adam"
model, error, accuracy = network.learn(trainer, neurons, activity,
learning_rate, episodes, cost,
optimizer)
network.plot(
error, "forestgreen",
"accumulated_errors_over_each_epoch_multilabel_classification",
"Episode", "Error")
network.plot(accuracy, "mediumvioletred",
"accuracy_over_each_epoch_multilabel_classification",
"Episode", "Learning accuracy")
network.test(model, tester, output)
log_file = open('backpropagation.log', 'w')
for training_set, ts_name in training_sets:
training_set = load_training_set(training_set)
print '.. begining %s' % ts_name
log_file.write('.. begining %s\n' % ts_name)
lowest_error = maxint
total_error_log = []
for hs in range(2, 11):
print ' %s hidden neurons' % hs
nn = NeuralNetwork(2, hs, 1)
error_log = training(nn, training_set, max_iterations=2000)
test_log, total_error = test(nn, get_random_set(10000))
log_file.write('%s neurons: %s errors \nhidden: %s\noutputs: %s\n\n' %
(hs, total_error, [hn.weights for hn in nn.hidden
], [on.weights
for on in nn.outputs]))
total_error_log.append((hs, total_error, error_log, test_log))
best = min(total_error_log, key=itemgetter(1))
print ' lowest error on %s hidden neurons network: %s failures' % (
best[0], best[1])
plot(best[2],
best[3],
save='media/%s-%s-best-performance.png' % (ts_name, best[0]))
bar_plot(map(itemgetter(1), total_error_log),
'media/error-derivation-%s' % ts_name)
log_file.close()