def test_init_net_simple(self):
"""
Tests the creation of a neural net with 2 inputs, 1 hidden layer
with 2 neurons, 1 output
"""
net = ecn.NeuralNet(2, (2, ), 1)
self.assertEqual(2, len(net.weights.keys()))
self.assertEqual((2, 3), np.shape(net.weights['h0']))
self.assertEqual((1, 3), np.shape(net.weights['y']))
print('Finished testing simple neural net init\n')
def test_init_net_complex(self):
"""
Tests the creation of a neural net with 10 inputs, 2 hidden layers,
one with 5 neurons, the other with 3 neurons, 2 output
"""
net = ecn.NeuralNet(10, (5, 3), 2)
self.assertEqual(3, len(net.weights.keys()))
self.assertEqual((5, 11), np.shape(net.weights['h0']))
self.assertEqual((3, 6), np.shape(net.weights['h1']))
self.assertEqual((2, 4), np.shape(net.weights['y']))
print('Finished testing complex neural net init\n')
def test_simple_net_forward(self):
"""
Tests the feed forward process in a simple network. The weights
are initalized to a fixed value in order to do that.
"""
net = ecn.NeuralNet(2, (2, ), 1)
net.weights = self._set_initial_weights()
dataset = [[1, 1]]
targets = [[0]]
net.train(dataset, targets, 0.5, 1)
self.assertTrue(net.fit_values[0] == [0.3, 1.4])
self.assertTrue(net.outputs[0] == [0.5744, 0.8022])
self.assertTrue(net.fit_values[1] == [0.1922])
self.assertTrue(net.outputs[1] == [0.5479])
print('Finished testing simple neural net forward\n')
def test_net_backpropagation_two_inputs(self):
"""
Tests the backpropagation process in a simple network with two inputs.
The weights are initalized to a fixed value in order to do that.
"""
net = ecn.NeuralNet(2, (2, ), 1)
net.weights = self._set_initial_weights()
dataset = [[1, 1], [0, 0]]
targets = [[0], [0]]
net.train(dataset, targets, 0.5, 1)
self.assertTrue((net.weights['h0'][0] == [-0.5876, 0.4066,
0.5066]).all())
self.assertTrue((net.weights['h0'][1] == [-0.2218, 0.7903,
0.7903]).all())
self.assertTrue((net.weights['y'][0] == [-0.4256, -0.4595,
0.8198]).all())
print('Finished testing backpropagation two inputs\n')
def test_net_backpropagation_one_input(self):
"""
Tests the backpropagation process in a simple network with one input.
The weights are initalized to a fixed value in order to do that.
"""
net = ecn.NeuralNet(2, (2, ), 1)
net.weights = self._set_initial_weights()
dataset = [[1, 1]]
targets = [[0]]
net.train(dataset, targets, 0.5, 1)
self.assertTrue((net.weights['h0'][0] == [-0.5934, 0.4066,
0.5066]).all())
self.assertTrue((net.weights['h0'][1] == [-0.2097, 0.7903,
0.7903]).all())
self.assertTrue((net.weights['y'][0] == [-0.3679, -0.4390,
0.8456]).all())
print('Finished testing backpropagation one input\n')
def test_net_backpropagation_four_inputs(self):
"""
Tests the backpropagation process in a simple network with three inputs.
The weights are initalized to a fixed value in order to do that.
"""
net = ecn.NeuralNet(2, (2, ), 1)
net.weights = self._set_initial_weights()
dataset = [[1, 1], [0, 0], [0, 1], [1, 0]]
targets = [[0], [0], [1], [1]]
net.train(dataset, targets, 0.5, 1)
self.assertTrue((net.weights['h0'][0] == [-0.6018, 0.400,
0.499]).all())
self.assertTrue((net.weights['h0'][1] == [-0.1969, 0.8027,
0.8028]).all())
self.assertTrue((net.weights['y'][0] == [-0.2970, -0.3995,
0.9021]).all())
print('Finished testing backpropagation four inputs\n')
N = 10
n_epochs = 50
learning_rate = 0.5
id_matrix = ecn.IDData(
N).ID_MATRIX # creates a template with an identity 10x10 matrix
error = 1
model = None
plt.subplot(1, 2, 1)
for test_idx in range(10):
test_dataset = id_matrix[test_idx]
train_dataset = [
inst for idx, inst in enumerate(id_matrix) if idx != test_idx
]
net = ecn.NeuralNet(N, (round(log(N, 2)), ), N)
net.train(train_dataset, train_dataset, learning_rate, n_epochs)
if net.errors[-1] < error:
model = net
error = net.errors[-1]
plt.plot(range(0, n_epochs), model.errors)
plt.xlabel('Epochs')
plt.ylabel('Root Mean Square Error')
plt.title("Root Mean Squared Error vs. Epochs")
plt.tight_layout()
plt.grid()
plt.axis([0, n_epochs, 0, 1])