def test_mi():
xs = [0, 0, 0, 0, 0, 0]
ys = [0, 0, 0, 0, 0, 0]
g = re.Remainder(xs, ys, k_max=2)
g.pz_xy = np.array([[[1, 0], [0, 0]], [[0, 0], [0, 1]]])
g.pxy = np.array([[0.5, 0], [0, 0.5]])
assert np.allclose(g.mi, np.log(2)), "%0.3f > log(2)" % g.mi
assert np.allclose(g.h, 0, atol=0.01), "%0.3f > 0" % g.h
def test_identity():
xs = np.array([0, 1, 0, 1])
ys = np.array([0, 0, 1, 1])
for _ in range(20):
g = re.Remainder(xs, ys, k_max=2)
zs = g.transform(xs, ys)
assert np.all(zs == xs)
predict_xs = g.predict(ys, zs)
assert np.all(xs == predict_xs), predict_xs
def test_perfect():
xs = np.repeat([0, 0, 1, 1, 1, 0], 5)
ys = np.repeat([0, 0, 0, 1, 1, 1], 5)
g = re.Remainder(xs, ys, k_max=2)
zs = g.transform(xs, ys)
print zip(xs, ys, zs)
assert np.allclose(g.mi, 0, atol=0.01), "%0.3f > 0" % g.mi
assert np.allclose(g.h, 0, atol=0.01), "%0.3f > 0" % g.h
predict_xs = g.predict(ys, zs)
assert np.all(xs == predict_xs), predict_xs
def test_null():
xs = [0, 0, 0, 0, 0, 0]
ys = [0, 0, 0, 0, 0, 0]
g = re.Remainder(xs, ys, k_max=2)
assert np.isclose(g.mi, 0), "No MI possible: %0.7f > 0" % g.mi
assert np.isclose(g.h, 0), "No uncertainty in x anyway: %0.3f > 0" % g.h
zs = g.transform(xs, ys)
print zip(xs, ys, zs)
assert np.all(zs == 0), "z=0 bc there is no signal."
predict_xs = g.predict(ys, zs)
assert np.all(xs == predict_xs), predict_xs
def test_deterministic():
xs = [0, 0, 1, 1]
ys = [0, 0, 1, 1]
g = re.Remainder(xs, ys, k_max=2)
zs = g.transform(xs, ys)
print zip(xs, ys, zs)
assert np.allclose(g.mi, 0, atol=0.01), "%0.3f > 0" % g.mi
assert np.allclose(g.h, 0, atol=0.01), "%0.3f > 0" % g.h
predict_xs = g.predict(ys, zs)
assert np.all(xs == predict_xs), predict_xs
assert np.all(g.transform([0] * 100, [0] * 100) == g.transform([0], [0]))
def test_probabilistic():
xs = [0, 0, 0, 1, 1, 0]
ys = [0, 0, 0, 1, 1, 1]
g = re.Remainder(xs, ys, k_max=3)
zs = g.transform(xs, ys)
print zip(xs, ys, zs)
print g.pz_xy
assert g.mi < 0.0001, "%0.5f" % g.mi
assert g.h < 0.0001, "%0.5f" % g.h
predict_xs = g.predict(ys, zs)
assert np.all(xs == predict_xs), predict_xs
def test_invertibility():
xs = np.random.randint(0, 5, 100)
ys = xs / 2 + np.random.randint(0, 2, 100)
g = re.Remainder(xs, ys, k_max=8)
zs = g.transform(xs, ys)
print 'mi, h', g.mi, g.h
print zip(xs, ys, zs)
print np.array_str(g.pz_xy, precision=2, suppress_small=True)
predict_xs = g.predict(ys, zs)
print zip(predict_xs, xs)
assert np.all(xs == predict_xs), predict_xs
assert g.h < 0.01, "%0.5f" % g.h
assert g.mi < 0.1, "%0.5f" % g.mi
def __init__(self, x, **kwargs):
k_max = kwargs.pop('k_max',
2) # Sets max cardinality for Remainder objects
self.verbose = kwargs.get('verbose', False)
self.corex = ce.Corex(**kwargs).fit(x)
self.labels = self.corex.labels
self.remainders = [
re.Remainder(xs[xs >= 0], self.labels[xs >= 0], k_max=k_max)
for xs in x.T
]
if self.verbose:
print 'z cardinalities', [
r.pz_xy.shape[0] for r in self.remainders
]