def test_output1():
"""If a single-layer neural network has homogeneous inputs and a
homogeneous weight matrix, all the outputs should be the same."""
n_inputs = 10
n_outputs = 5
weights = np.ones((n_inputs + 1, n_outputs)) # Adding an extra weight for the bias
nn = neural_network.SimpleNeuralNetworkExecutable([ weights ], neural_network.sigmoid)
x = [1]*n_inputs
result = nn(x)
assert(len(result) == n_outputs)
for o in result:
assert(approx(result[0]) == o)
def test_output5():
"""If we give each ReLu neuron in a single-layer network a
negative input, their output should be zero."""
n_inputs = 10
n_outputs = 5
# Random weights all < 0, adding an extra weight for the bias
weights = -np.random.uniform(0, 1, (n_inputs + 1, n_outputs))
nn = neural_network.SimpleNeuralNetworkExecutable([ weights ], neural_network.relu)
x = [1]*n_inputs
result = nn(x)
assert(len(result) == n_outputs)
for o in result:
assert(approx(0.0) == o)
def test_output4():
"""If we give each sigmoid neuron in a single-layer network a very
negative input, their output should be approximately zero."""
n_inputs = 10
n_outputs = 5
# Random weights all < -100, adding an extra weight for the bias
weights = -100*np.random.uniform(1, 2, (n_inputs + 1, n_outputs))
nn = neural_network.SimpleNeuralNetworkExecutable([ weights ], neural_network.sigmoid)
x = [1]*n_inputs
result = nn(x)
assert(len(result) == n_outputs)
for o in result:
assert(approx(0.0) == o)
def test_output6():
"""If we give each ReLu neuron in a single-layer network a
positive input, with network inputs of all ones, their
output should equal the sum of all the weights."""
n_inputs = 10
n_outputs = 5
# Random weights all < 0, adding an extra weight for the bias
weights = np.random.uniform(0, 1, (n_inputs + 1, n_outputs))
nn = neural_network.SimpleNeuralNetworkExecutable([ weights ], neural_network.relu)
x = [1]*n_inputs
result = nn(x)
assert(len(result) == n_outputs)
for i, o in enumerate(result):
expected = np.sum(weights[:,i])
assert(approx(expected) == o)
def test_output2():
"""If a multi-layer neural network has homogeneous inputs and
homogeneous weight matrices, all the outputs should be the same."""
n_inputs = 10
n_hidden1 = 5
n_hidden2 = 3
n_outputs = 5
# Three layers, three matrices
weights = [ np.ones((n_inputs + 1, n_hidden1)), # Adding an extra weight for the bias
3*np.ones((n_hidden1 + 1, n_hidden2)),
2*np.ones((n_hidden2 + 1, n_outputs)) ]
nn = neural_network.SimpleNeuralNetworkExecutable(weights, neural_network.sigmoid)
x = [1]*n_inputs
result = nn(x)
assert(len(result) == n_outputs)
for o in result:
assert(approx(result[0]) == o)