def test_hjmi(library, madelon):
pymit._set_library(library)
data, labels = madelon
bins = 10
expected_features = [
241, 338, 378, 105, 472
] #, 475, 433, 64, 128, 442, 453, 336, 48, 493, 281, 318, 153, 28, 451, 455]
[num_examples, num_features] = data.shape
data_discrete = np.zeros([num_examples, num_features])
for i in range(num_features):
_, bin_edges = pymit._lib.histogram(data[:, i], bins=bins)
data_discrete[:, i] = pymit._lib.digitize(data[:, i],
bin_edges,
right=False)
max_features = len(expected_features)
selected_features = []
j_h = 0
hjmi = None
for i in range(0, max_features):
jmi = np.zeros([num_features], dtype=np.float)
for X_k in range(num_features):
if X_k in selected_features:
continue
jmi_1 = pymit.I(data_discrete[:, X_k], labels, bins=[bins, 2])
jmi_2 = 0
for X_j in selected_features:
tmp1 = pymit.I(data_discrete[:, X_k],
data_discrete[:, X_j],
bins=[bins, bins])
tmp2 = pymit.I_cond(data_discrete[:, X_k],
data_discrete[:, X_j],
labels,
bins=[bins, bins, 2])
jmi_2 += tmp1 - tmp2
if len(selected_features) == 0:
jmi[X_k] = j_h + jmi_1
else:
jmi[X_k] = j_h + jmi_1 - jmi_2 / len(selected_features)
f = jmi.argmax()
j_h = jmi[f]
if hjmi is None or (j_h - hjmi) / hjmi > 0.03:
hjmi = j_h
selected_features.append(f)
else:
break
assert np.array_equal(expected_features, selected_features)
def test_jmi(library, madelon):
pymit._set_library(library)
data, labels = madelon
bins = 10
expected_features = [
241, 338, 378, 105, 472
] #, 475, 433, 64, 128, 442, 453, 336, 48, 493, 281, 318, 153, 28, 451, 455]
[num_examples, num_features] = data.shape
data_discrete = np.zeros([num_examples, num_features])
for i in range(num_features):
_, bin_edges = pymit._lib.histogram(data[:, i], bins=bins)
data_discrete[:, i] = pymit._lib.digitize(data[:, i],
bin_edges,
right=False)
max_features = len(expected_features)
selected_features = []
mi = np.zeros([num_features], dtype=np.float)
for i in range(num_features):
mi[i] = pymit.I(data_discrete[:, i], labels, bins=[bins, 2])
f = mi.argmax()
selected_features.append(f)
for i in range(1, max_features):
jmi = np.zeros([num_features], dtype=np.float)
for X_k in range(num_features):
if X_k in selected_features:
continue
for X_j in selected_features:
sum1 = pymit.I(data_discrete[:, X_j], labels, bins=[bins, 2])
sum2 = pymit.I_cond(data_discrete[:, X_k],
labels,
data_discrete[:, X_j],
bins=[bins, 2, bins])
jmi[X_k] += sum1 + sum2
f = jmi.argmax()
selected_features.append(f)
assert np.array_equal(expected_features, selected_features)
def calculate_jmi(X, Y, features, selected_features):
JMI = numpy.full([len(features)], numpy.nan, dtype=numpy.float)
for i,X_k in enumerate(features):
if X_k in selected_features:
continue
jmi = 0
for X_j in selected_features:
sum1 = pymit.I(X[:, X_j], Y, bins=[bins, 2])
sum2 = pymit.I_cond(X[:, X_k], Y, X[:, X_j], bins=[bins, 2, bins])
jmi += sum1 + sum2
JMI[i] = jmi
return [JMI]
def calculate_mi(X, Y, features):
MI = numpy.full([len(features)], numpy.nan, dtype=numpy.float)
for i,X_i in enumerate(features):
MI[i] = pymit.I(X[:, X_i], Y , bins=[bins, 2])
return [MI]
Y = labels
bins = 10
[tmp, features] = X.shape
D = numpy.zeros([tmp, features])
for i in range(features):
N, E = numpy.histogram(X[:, i], bins=bins)
D[:, i] = numpy.digitize(X[:, i], E, right=False)
max_features = 20
selected_features = []
MI = numpy.full([features], numpy.nan, dtype=numpy.float)
for i in range(features):
MI[i] = pymit.I(D[:, i], Y, bins=[bins, 2])
f = MI.argmax()
selected_features.append(f)
print("001 {:0>3d} {}".format(f, MI[f]))
for i in range(1, max_features):
JMI = numpy.zeros([features], dtype=numpy.float)
for X_k in range(features):
if X_k in selected_features:
continue
for X_j in selected_features:
sum1 = pymit.I(D[:, X_j], Y, bins=[bins, 2])
sum2 = pymit.I_cond(D[:, X_k], Y, D[:, X_j], bins=[bins, 2, bins])
for i in range(features):
N, E = numpy.histogram(X[:, i], bins=bins)
D[:, i] = numpy.digitize(X[:, i], E, right=False)
max_features = 200
selected_features = []
j_h = 0
hjmi = None
for i in range(0, max_features):
JMI = numpy.zeros([features], dtype=numpy.float)
for X_k in range(features):
if X_k in selected_features:
continue
jmi_1 = pymit.I(D[:, X_k], Y, bins=[bins, 2])
jmi_2 = 0
for X_j in selected_features:
tmp1 = pymit.I(D[:, X_k], D[:, X_j], bins=[bins, bins])
tmp2 = pymit.I_cond(D[:, X_k], D[:, X_j], Y, bins=[bins, bins, 2])
jmi_2 += tmp1 - tmp2
if len(selected_features) == 0:
JMI[X_k] += j_h + jmi_1
else:
JMI[X_k] += j_h + jmi_1 - jmi_2 / len(selected_features)
f = JMI.argmax()
j_h = JMI[f]
if (hjmi == None) or ((j_h - hjmi) / hjmi > 0.03):
hjmi = j_h
selected_features.append(f)
print("{:0>3d} {:>3d} {}".format(len(selected_features), f, j_h))