def _test_hidden_deltas():
mynet = Network(4, 2, 3)
mynet.create_connections()
mynet.feed_forward([1, 1, 1, 1])
print(mynet.output_ouputs)
mynet.matrix_who = np.matrix([[0.11, 0.12, 0.13], [0.21, 0.22, 0.23]])
mynet.output_ouputs = np.matrix([[0.1, 0.2, 0.3]])
target = [0.9, 0.8, 0.7]
ods = mynet.get_output_deltas(target)
mynet.matrix_who = np.matrix([[1, 2, 3], [4, 5, 6]])
mynet.hidden_outputs = np.matrix([[0.5, 0.6]])
ods = np.matrix([[0.05, 0.06, 0.07]])
hds = mynet.get_hidden_deltas(ods)
np.testing.assert_array_almost_equal(
np.matrix([mynet.gradient(0.5) * 0.38,
mynet.gradient(0.6) * 0.92]), hds)
def _test_back_propagate():
mynet = Network(4, 2, 3)
mynet.create_connections()
mynet.feed_forward([1, 1, 1, 1])
print(mynet.output_ouputs)
mynet.matrix_who = np.matrix([[0.11, 0.12, 0.13], [0.21, 0.22, 0.23]])
mynet.output_ouputs = np.matrix([[0.1, 0.2, 0.3]])
target = [0.9, 0.8, 0.7]
mynet.back_propagate(target)
def _test_ouput_deltas():
mynet = Network(4, 2, 3)
mynet.create_connections()
mynet.feed_forward([1, 1, 1, 1])
print(mynet.output_ouputs)
mynet.matrix_who = np.matrix([[0.11, 0.12, 0.13], [0.21, 0.22, 0.23]])
mynet.output_ouputs = np.matrix([[0.1, 0.2, 0.3]])
target = [0.9, 0.8, 0.7]
ods = mynet.get_output_deltas(target)
np.testing.assert_array_almost_equal(
np.matrix([(0.9 - 0.1) * mynet.gradient(0.1),
(0.8 - 0.2) * mynet.gradient(0.2),
(0.7 - 0.3) * mynet.gradient(0.3)]), ods)
def _test_update_who():
mynet = Network(3, 4, 3)
mynet.create_connections()
mynet.matrix_who = np.matrix([[-0.5910955, 0.75623487, -0.94522481],
[0.34093502, -0.1653904, 0.11737966],
[-0.71922612, -0.60379702, 0.60148914],
[0.93652315, -0.37315164, 0.38464523]])
d = [0.40795614, 0.62674606, 0.23841622, 0.49377636]
d = [1, 2, 3]
mynet.hidden_outputs = np.array(d, dtype=float)
output_deltas = np.array([0.10676222, -0.11685494, -0.12631629],
dtype=float)
# c = output_deltas * h
#result is addition of c to each row of matric_who
matrix_who_expected = np.matrix([[-0.56931835, 0.73239903, -0.97099057],
[0.37439142, -0.20200958, 0.07779554],
[-0.7064992, -0.61772708, 0.58643121],
[0.96288148, -0.40200175, 0.35345923]])
mynet.update_hidden_output_weight_new(output_deltas, 0.5)
assert mynet.matrix_who == matrix_who_expected