def sine_training():
n = NeuralNet([1, 5, 1])
sines = get_sines(10)
testing_data = next(sines)
for i in range(1000):
x, y = next(sines)
n.train(x, y, 10, 1, testing_data=testing_data)
# validation
x, y = next(sines)
hx = n.predict(x)
errors = []
print('doing validation---------------------------')
for x, y, y_predicted in zip(x, y, hx):
errors.append(abs(y[0] - y_predicted[0]) / y[0])
print(x[0], y[0], y_predicted[0])
print(sum(errors) / len(errors))
x = np.reshape([2 * math.pi * i/1000 for i in range(1000)], (1000, 1))
y = [(i[0] - 0.5)*2 for i in n.predict(x)]
y_actual = [i[0] for i in np.sin(x)]
plt.plot(x, y_actual)
plt.hold(True)
plt.plot(x, y)
plt.show()
def test_random_initialize_weights(self):
for i, layers in enumerate(self.valid_layers):
net = NeuralNet(layers)
net.random_initialize_weights()
# check shape of weights
self.assertEqual([weight.shape for weight in net.weights], self.weights_for_valid_layers[i])
def test_instance_creation(self):
for layers in self.valid_layers:
self.assertIsInstance(NeuralNet(layers), NeuralNet)
for layers in self.invalid_layers:
with self.assertRaises(NeuralNetError):
NeuralNet(layers)
def run_training(self, metadata, predict_test_data=False):
network = NeuralNet(layers=metadata['layers'],
lmda=metadata['lambda'])
generator = self.reader.get(metadata['batch_size'], 'training')
for i in range(metadata['epochs']):
try:
x, y = next(generator)
except StopIteration:
return network
network.train(x, y, iterations=metadata['iterations'],
learning_rate=metadata['learning_rate'])
return network
def validate(reader, weights):
data_stream = reader.get(30, mode='validation')
n = NeuralNet(metadata['layers'])
n.weights = weights
counter = 0
matches = 0
for x, y in data_stream:
counter += x.shape[0]
predicted = n.predict(x)
p_digits = np.argmax(predicted, axis=1)
y_digits = np.argmax(y, axis=1)
accuracy = sum([1 if i == j else 0 for i, j in zip(p_digits, y_digits)]) / len(p_digits)
print 'predicted-\n', p_digits
print 'actual-\n', y_digits
print('accuracy-{0}%'.format(round(accuracy * 100, 2)))
return None # Let's just run only one loop for the time being
def sine_training():
n = NeuralNet([1, 5, 1])
sines = get_sines(10)
testing_data = next(sines)
for i in range(1000):
x, y = next(sines)
n.train(x, y, 10, 1, testing_data=testing_data)
# validation
x, y = next(sines)
hx = n.predict(x)
errors = []
print('doing validation---------------------------')
for x, y, y_predicted in zip(x, y, hx):
errors.append(abs(y[0] - y_predicted[0]) / y[0])
print(x[0], y[0], y_predicted[0])
print(sum(errors) / len(errors))
x = np.reshape([2 * math.pi * i / 1000 for i in range(1000)], (1000, 1))
y = [(i[0] - 0.5) * 2 for i in n.predict(x)]
y_actual = [i[0] for i in np.sin(x)]
plt.plot(x, y_actual)
plt.hold(True)
plt.plot(x, y)
plt.show()
