def test_naive_multiclass_ova_vs_binary_asgd():
rstate = RandomState(42)
n_classes = 3
Xtrn, ytrn = get_fake_multiclass_data(N_POINTS, N_FEATURES, n_classes,
rstate)
Xtst, ytst = get_fake_multiclass_data(N_POINTS, N_FEATURES, n_classes,
rstate)
# -- ground truth 'gt': emulate OVA with binary asgd classifiers
# binary assignments
ytrns = ytrn[np.newaxis, :] == np.arange(n_classes)[:, np.newaxis]
# transform into -1 / +1
ytrns = 2 * ytrns - 1
# get individual predictions
basgds = [
BinaryASGD(*DEFAULT_ARGS, rstate=RandomState(999),
**DEFAULT_KWARGS).partial_fit(Xtrn, ysel) for ysel in ytrns
]
ytsts = [basgd.decision_function(Xtst) for basgd in basgds]
# convert to array of shape (n_classes, n_points)
gt = np.array(ytsts).T
# -- given 'gv': use the OVA class
clf = OVAASGD(*((n_classes, ) + DEFAULT_ARGS),
rstate=RandomState(999),
**DEFAULT_KWARGS)
clf.partial_fit(Xtrn, ytrn)
gv = clf.decision_function(Xtst)
# -- test
assert_allclose(gv, gt, rtol=RTOL, atol=ATOL)
def test_naive_multiclass_ova_vs_binary_asgd():
rstate = RandomState(42)
n_classes = 3
Xtrn, ytrn = get_fake_multiclass_data(
N_POINTS, N_FEATURES, n_classes, rstate)
Xtst, ytst = get_fake_multiclass_data(
N_POINTS, N_FEATURES, n_classes, rstate)
# -- ground truth 'gt': emulate OVA with binary asgd classifiers
# binary assignments
ytrns = ytrn[np.newaxis, :] == np.arange(n_classes)[:, np.newaxis]
# transform into -1 / +1
ytrns = 2 * ytrns - 1
# get individual predictions
basgds = [
BinaryASGD(
*DEFAULT_ARGS,
rstate=RandomState(999),
**DEFAULT_KWARGS).partial_fit(Xtrn, ysel)
for ysel in ytrns
]
ytsts = [basgd.decision_function(Xtst) for basgd in basgds]
# convert to array of shape (n_classes, n_points)
gt = np.array(ytsts).T
# -- given 'gv': use the OVA class
clf = OVAASGD(*((n_classes,) + DEFAULT_ARGS),
rstate=RandomState(999), **DEFAULT_KWARGS)
clf.partial_fit(Xtrn, ytrn)
gv = clf.decision_function(Xtst)
# -- test
assert_allclose(gv, gt, rtol=RTOL, atol=ATOL)
