def test_disassemble_params(self):
net = FeedforwardNetwork(ninput=3,
noutput=2,
nhidden=[2, 3],
activ_func=Logistic())
params = [
1111, 1112, 1113, 1121, 1122, 1123, 121, 122, 2111, 2112, 2121,
2122, 2131, 2132, 221, 222, 223, 3111, 3112, 3113, 3121, 3122,
3123, 321, 322
]
w, b = net.disassemble_params(params)
W0 = [[1111, 1112, 1113], [1121, 1122, 1123]]
W1 = [[2111, 2112], [2121, 2122], [2131, 2132]]
W2 = [[3111, 3112, 3113], [3121, 3122, 3123]]
B0 = [[121], [122]]
B1 = [[221], [222], [223]]
B2 = [[321], [322]]
assert_array_almost_equal(w[0], W0)
assert_array_almost_equal(w[1], W1)
assert_array_almost_equal(w[2], W2)
assert_array_almost_equal(b[0], B0)
assert_array_almost_equal(b[1], B1)
assert_array_almost_equal(b[2], B2)
def test_two_hiddenlayer_three_input_two_output(self):
net = FeedforwardNetwork(ninput=3,
noutput=2,
nhidden=[2, 3],
activ_func=Logistic())
params = np.array([
1.111, 1.112, 11.13, 1.121, 0.1122, 1123, 1.21, 1.22, 0.2111,
0.2112, 2.121, 2.122, 0.2131, 0.2132, 2.21, 2.22, 0.223, 3.111,
0.3112, 3.113, 0.3121, 0.3122, 0.3123, 0.321, 3.22
])
x = [0.1, 2.2, 0.34]
def call_x(y):
return net(y, params)
def call_params(theta):
return net(x, theta)
dx_numeric = nd.Jacobian(call_x)(x)
dparams_numeric = nd.Jacobian(call_params)(params)
dx, dparams = net.derivatives(x, params)
assert_array_almost_equal(dx, dx_numeric)
assert_array_almost_equal(dparams, dparams_numeric)
def test_three_hiddenlayer_three_input_two_output(self):
net = FeedforwardNetwork(ninput=3,
noutput=2,
nhidden=[2, 3, 4],
activ_func=[
Logistic(),
HyperbolicTangent(),
Identity(),
Logistic()
])
for i in range(5):
params = np.random.normal(size=(net.nparams, ))
x = np.random.normal(size=(net.ninput, ))
def call_x(y):
return net(y, params)
def call_params(theta):
return net(x, theta)
dx_numeric = nd.Jacobian(call_x)(x)
dparams_numeric = nd.Jacobian(call_params)(params)
dx, dparams = net.derivatives(x, params)
assert_array_almost_equal(dx, dx_numeric)
assert_array_almost_equal(dparams, dparams_numeric)
def test_two_hiddenlayer_three_input_two_output(self):
l = Logistic()
i = Identity()
net = FeedforwardNetwork(ninput=3,
noutput=2,
nhidden=[2, 3],
activ_func=Logistic())
params = np.array([
1.111, 1.112, 11.13, 112.1, 0.1122, 1123, 12.1, 1.22, 0.2111,
0.2112, 2.121, 2.122, 0.2131, 0.2132, 2.21, 2.22, 0.223, 3.111,
0.3112, 3.113, 0.3121, 0.3122, 0.3123, 0.321, 3.22
])
w, b = net.disassemble_params(params)
innet = np.asarray([1, 2, 3]).reshape(3, 1)
expected = i(
np.dot(w[2], l(np.dot(w[1], l(np.dot(w[0], innet) + b[0])) +
b[1])) + b[2]).flatten()
assert_array_almost_equal(net(innet, params), expected, decimal=3)
innet = np.asarray([1.78, -2, 3]).reshape(3, 1)
expected = i(
np.dot(w[2], l(np.dot(w[1], l(np.dot(w[0], innet) + b[0])) +
b[1])) + b[2]).flatten()
assert_array_almost_equal(net(innet, params), expected, decimal=3)
innet = np.asarray([1, 25, 3]).reshape(3, 1)
expected = i(
np.dot(w[2], l(np.dot(w[1], l(np.dot(w[0], innet) + b[0])) +
b[1])) + b[2]).flatten()
assert_array_almost_equal(net(innet, params), expected, decimal=3)
def test_assemble_params_raise_exception(self):
net = FeedforwardNetwork(ninput=3,
noutput=2,
nhidden=[2, 3],
activ_func=Logistic())
expected_params = [
1111, 1112, 1113, 1121, 1122, 1123, 121, 122, 2111, 2112, 2121,
2122, 2131, 2132, 221, 222, 223, 3111, 3112, 3113, 3121, 3122,
3123, 321, 322
]
W0 = [[1111, 1112, 1113], [1121, 1122, 1123]]
W1 = [[2111, 2112], [2121, 2122], [2131, 2132]]
W2 = [[3111, 3112, 3113], [3121, 3122, 3123]]
B0 = [[121], [122]]
B1 = [[221], [222], [223]]
B2 = [[321], [322]]
assert_raises(ValueError, net.assemble_params, [W0, W1], [B0, B1, B2])
assert_raises(ValueError, net.assemble_params, [W0, W1, W2], [B1, B2])
assert_raises(ValueError, net.assemble_params, [W0, W1, W2],
[[1, 2, 3], B1, B2])
assert_raises(ValueError, net.assemble_params, [[1, 2, 3], W1, W2],
[B0, B1, B2])
def test_uniform_distribution_std_check(self):
nexp = 1000
net = FeedforwardNetwork(ninput=3,
noutput=2,
nhidden=[2, 3],
activ_func=Logistic())
mean_per_layer = np.zeros(net.nlayers)
variance_per_layer = np.zeros(net.nlayers)
for i in range(nexp):
np.random.seed(i * 2)
weights, bias = LeCunInitializer(distribution='uniform')(net)
for j in range(net.nlayers):
w = weights[j]
we = w.shape[0] * w.shape[1]
mean_per_layer[j] += 1 / (nexp * we) * np.sum(w)
variance_per_layer[j] += 1 / (nexp * we) * np.sum(w**2)
std_per_layer = np.sqrt(variance_per_layer)
assert_array_almost_equal(
std_per_layer, [1 / np.sqrt(3), 1 / np.sqrt(2), 1 / np.sqrt(3)],
decimal=2)
assert_array_almost_equal(mean_per_layer, [0, 0, 0], decimal=2)
def test_disassemble_params_raise_exception(self):
net = FeedforwardNetwork(ninput=3,
noutput=2,
nhidden=[2, 3],
activ_func=Logistic())
params = [1111]
assert_raises(ValueError, net.disassemble_params, params)
def test_assemble_params(self):
net = FeedforwardNetwork(ninput=3,
noutput=2,
nhidden=[2, 3],
activ_func=Logistic())
expected_params = [
1111, 1112, 1113, 1121, 1122, 1123, 121, 122, 2111, 2112, 2121,
2122, 2131, 2132, 221, 222, 223, 3111, 3112, 3113, 3121, 3122,
3123, 321, 322
]
W0 = [[1111, 1112, 1113], [1121, 1122, 1123]]
W1 = [[2111, 2112], [2121, 2122], [2131, 2132]]
W2 = [[3111, 3112, 3113], [3121, 3122, 3123]]
B0 = [[121], [122]]
B1 = [[221], [222], [223]]
B2 = [[321], [322]]
params = net.assemble_params([W0, W1, W2], [B0, B1, B2])
assert_array_almost_equal(params, expected_params)
def test_shape_check(self):
net = FeedforwardNetwork(ninput=3,
noutput=2,
nhidden=[2, 3],
activ_func=Logistic())
weights, bias = LeCunInitializer()(net)
# Check shape
assert_equal(weights[0].shape, (2, 3))
assert_equal(weights[1].shape, (3, 2))
assert_equal(weights[2].shape, (2, 3))
assert_equal(bias[0].shape, (2, 1))
assert_equal(bias[1].shape, (3, 1))
assert_equal(bias[2].shape, (2, 1))
def test_raise_exception(self):
net = FeedforwardNetwork(ninput=3,
noutput=2,
nhidden=[2, 3],
activ_func=Logistic())
params = np.array([
1.111, 1.112, 11.13, 112.1, 0.1122, 1123, 12.1, 1.22, 0.2111,
0.2112, 2.121, 2.122, 0.2131, 0.2132, 2.21, 2.22, 0.223, 3.111,
0.3112, 3.113, 0.3121, 0.3122, 0.3123, 0.321, 3.22
])
assert_raises(ValueError, net.derivatives, [1, 2, 3, 4], params)
assert_raises(ValueError, net.derivatives, [1, 2, 3], params[:-1])
def test_single_activation_function(self):
activ_func = Logistic()
net = FeedforwardNetwork(ninput=3,
noutput=2,
nhidden=[2, 3],
activ_func=activ_func)
assert_equal(net.ninput, 3)
assert_equal(net.noutput, 2)
assert_equal(net.nhidden, [2, 3])
assert_equal(net.weights_shape, [(2, 3), (3, 2), (2, 3)])
assert_equal(net.bias_shape, [(2, 1), (3, 1), (2, 1)])
assert_equal(net.nparams, 25)
assert_equal(net.nlayers, 3)
assert_equal(isinstance(net.activ_func[0], Logistic), True)
assert_equal(isinstance(net.activ_func[1], Logistic), True)
assert_equal(isinstance(net.activ_func[2], Identity), True)
def test_listof_activation_function(self):
net = FeedforwardNetwork(ninput=3,
noutput=2,
nhidden=[2, 3, 4],
activ_func=[
Logistic(),
HyperbolicTangent(),
Identity(),
Logistic()
])
assert_equal(net.ninput, 3)
assert_equal(net.noutput, 2)
assert_equal(net.nhidden, [2, 3, 4])
assert_equal(net.weights_shape, [(2, 3), (3, 2), (4, 3), (2, 4)])
assert_equal(net.bias_shape, [(2, 1), (3, 1), (4, 1), (2, 1)])
assert_equal(net.nlayers, 4)
assert_equal(net.nparams, 43)
assert_equal(isinstance(net.activ_func[0], Logistic), True)
assert_equal(isinstance(net.activ_func[1], HyperbolicTangent), True)
assert_equal(isinstance(net.activ_func[2], Identity), True)
assert_equal(isinstance(net.activ_func[3], Logistic), True)