def test_adjust_weight(self):
backpropagator = Backpropagator()
output = ReceiveAllNeuron()
hidden_1 = ReceiveAllNeuron()
hidden_2 = ReceiveAllNeuron()
hidden_1.connect_to(output)
hidden_1.out_connections_list[0].weight = -0.75
hidden_2.connect_to(output)
hidden_2.out_connections_list[0].weight = -0.25
hidden_1.receive_signal(0.434719)
hidden_2.receive_signal(0.434719)
output.error = 0.09754925
hidden_1.error = -0.01745285
hidden_2.error = -0.00581762
connection_1 = hidden_1.out_connections_list[0]
backpropagator.adjust_weight(connection_1, 1.1)
# -0.75 + (1.1 * 0.09754925 * -0.45525)
self.assertAlmostEqual(connection_1.weight, -0.68486637)
connection_2 = hidden_2.out_connections_list[0]
backpropagator.adjust_weight(connection_2, 1.1)
# -0.25 + (1.1 * 0.09754925 * -0.15175)
self.assertAlmostEqual(connection_2.weight, -0.18486637)
def test_hiddenError(self):
backpropagator = Backpropagator()
output = ReceiveAllNeuron()
hidden_1 = ReceiveAllNeuron()
hidden_2 = ReceiveAllNeuron()
hidden_1.connect_to(output)
hidden_1.out_connections_list[0].weight = -0.75
hidden_2.connect_to(output)
hidden_2.out_connections_list[0].weight = -0.25
# sigmoid(0.434719) == 0.6070000
hidden_1.receive_signal(0.434719)
hidden_2.receive_signal(0.434719)
# sigmoid(0.434719) * -0.75 + sigmoid(0.434719) * -0.25 == -0.607
# sigmoid(-0.607) == 0.352744
# 0.78 - 0.352744 == 0.3527438
output.error = backpropagator.output_error(output, 0.78)
self.assertAlmostEqual(output.error, 0.09754925)
hidden_1.error = backpropagator.hidden_error(hidden_1)
hidden_2.error = backpropagator.hidden_error(hidden_2)
# 0.6070000 * (1 - 0.6070000)*(-0.75 * 0.09754925)
self.assertAlmostEqual(hidden_1.error, -0.01745285)
# 0.6070000 * (1 - 0.6070000)*(-0.25 * 0.09754925)
self.assertAlmostEqual(hidden_2.error, -0.00581762)
def test_teach_acceptable_error(self):
backpropagator = Backpropagator()
normalizer = Normalizer(in_min = -15, in_max = 15,
out_min = -30, out_max = 30,
norm_min = -2, norm_max = 2)
network = FeedForwardNN(normalizer, [1, 3, 1])
network.randomize_connection_weights(seed = 74)
expectations = [Expectation([i], [2*i])
for i in range(-5, 5)]
result = backpropagator.teach(network,
expectations,
learning_rate = 1.5,
max_iterations = 2000,
acceptable_error = .5)
self.assertLessEqual(result.error, .5)
errors = 0
for exp in expectations:
errors += backpropagator.calculate_error(
network.receive_inputs(exp.inputs), exp.outputs)
self.assertLessEqual(errors, .5)
def test_propagate_errors(self):
backpropagator = Backpropagator()
normalizer = Normalizer(in_max = 100, out_max = 200)
network = FeedForwardNN(normalizer, [1, 2, 1])
expectation = Expectation([50], [148])
result = network.receive_inputs(expectation.inputs)
backpropagator.learn(network, expectation, 1)
def test_teach_max_iterations(self):
backpropagator = Backpropagator()
normalizer = Normalizer(in_min = -15, in_max = 15,
out_min = -30, out_max = 30,
norm_min = -2, norm_max = 2)
network = FeedForwardNN(normalizer, [1, 2, 2, 1])
network.randomize_connection_weights(seed = 74)
expectations = [Expectation([i], [2*i])
for i in range(-5, 5)]
result = backpropagator.teach(network,
expectations,
learning_rate = 1.5,
max_iterations = 123,
acceptable_error = 0)
self.assertEqual(result.epochs, 123)
def learn_function(fn, in_min, in_max):
expectations = [Expectation([i], fn(i)) for i in np.linspace(in_min, in_max + 1, num=100)]
in_range = in_max - in_min
in_min = in_min - 0.1 * in_range
in_max = in_max + 0.1 * in_range
out_min = None
out_max = None
for e in expectations:
if out_min == None or e.outputs[0] < out_min:
out_min = e.outputs[0]
if out_max == None or e.outputs[0] > out_max:
out_max = e.outputs[0]
out_range = out_max - out_min
out_min = out_min - 0.1 * out_range
out_max = out_max + 0.1 * out_range
backpropagator = Backpropagator()
normalizer = Normalizer(in_min=in_min, in_max=in_max, out_min=out_min, out_max=out_max, norm_min=-10, norm_max=10)
network = FeedForwardNN(normalizer, [1, 3, 3, 1], is_regression=True)
network.randomize_connection_weights(min=-1, max=1)
show_fit_tracker(network, expectations)
backpropagator.teach(
network,
expectations,
learning_rate=0.008,
max_iterations=30000,
acceptable_error=1,
callback_func=update_fit_plot,
)
return network
def test_learn(self):
backpropagator = Backpropagator()
normalizer = Normalizer(in_min = 1, in_max = 15,
out_min = -1, out_max = 1,
norm_min = -3, norm_max = 3)
network = FeedForwardNN(normalizer, [1, 2, 2, 1])
network.randomize_connection_weights(seed = 74)
neurons = network.neurons
expectation = Expectation([1], [0.8415])
error = backpropagator.calculate_error(
network.receive_inputs(expectation.inputs),
expectation.outputs)
for i in range(20):
last_error = error
backpropagator.learn(network, expectation, 1.5)
actual = network.receive_inputs(expectation.inputs)
print(actual)
error = backpropagator.calculate_error(actual, expectation.outputs)
self.assertLess(error, last_error)
def test_calculate_error(self):
backpropagator = Backpropagator()
error = backpropagator.calculate_error([5.5, 1, -8], [5, -4, 2.5])
self.assertAlmostEqual(error, 135.5)
def test_outputError(self):
backpropagator = Backpropagator()
neuron = ReceiveAllNeuron()
neuron.receive_signal(-0.607)
error = backpropagator.output_error(neuron, 0.78)
self.assertAlmostEqual(error, 0.09754925)
