def test_subspace_alignment():
"""Test the alignment between datasets."""
X = rnd.randn(100, 10)
Z = np.dot(rnd.randn(100, 10), np.diag(np.arange(1, 11)))
clf = SubspaceAlignedClassifier()
V, CX, CZ = clf.subspace_alignment(X, Z, num_components=3)
assert not np.any(np.isnan(V))
assert CX.shape[1] == 3
assert CZ.shape[1] == 3
def prepare(self, initial_target_labels=False, X_t_init=[], y_t_init=[]):
np.random.seed(self.seed)
self.initial_target_labels = initial_target_labels
suba = SubspaceAlignedClassifier()
if not self.initial_target_labels:
self.X_t_init = np.empty([0, self.X_t_ori.shape[1]])
self.y_t_init = np.array([])
if self.domain_adaptation:
V, CX, self.CZ = suba.subspace_alignment(
self.X_s_ori,
self.X_t_ori,
num_components=self.num_components)
else:
assert X_t_init.shape[
0] > 0, 'Initial target data must not be empty'
self.X_t_init = X_t_init
self.y_t_init = y_t_init
if self.domain_adaptation:
V, CX, self.CZ = suba.subspace_alignment(
self.X_s_ori,
np.vstack([X_t_init, self.X_t_ori]),
num_components=self.num_components)
if self.domain_adaptation:
self.X_s = self.X_s_ori @ CX # map to principal component
self.X_s = self.X_s @ V # align to subspace
self.X_t = self.X_t_ori @ self.CZ
else:
self.X_s = self.X_s_ori
self.X_t = self.X_t_ori
if self.initial_target_labels and self.domain_adaptation:
self.X_t_init = self.X_t_init @ self.CZ
if self.use_expert:
assert initial_target_labels, "to use expert classifier, initial target labels must be true"
self.expert = RandomForestClassifier(n_estimators=64).fit(
self.X_t_init, self.y_t_init)
self.model.fit(self.X_s,
self.X_t_init,
self.y_s,
self.y_t_init,
warm_start=False,
max_iter=self.max_iter,
use_dropout=True,
desc='Preparing',
regularize=True)
## TRANSDUCTION THROUGH SOURCE-SPECIFIC NET
pred_proba_f = self.model.predict_proba(self.X_t, 1).T
pred_proba = (pred_proba_f)
if self.initial_target_labels:
alpha = self.alpha
pred_proba_g = self.model.predict_proba(self.X_t, 2).T
pred_proba = alpha * pred_proba_f + (1 - alpha) * pred_proba_g
if self.use_expert:
pred_proba_expert = self.expert.predict_proba(self.X_t)
pred_proba = ((1 - self.beta) * pred_proba +
(self.beta) * pred_proba_expert)
self.p = pred_proba
# max confidence of prediction on each instance
proba_max = pred_proba.max(axis=1)
idx_gt_threshold = np.where(proba_max > self.min_confidence)
proba_gt_threshold = proba_max[idx_gt_threshold]
self.initial_selected_num = len(proba_gt_threshold)
# Evaluate 1st phase transduction
pred = (pred_proba_f).argmax(axis=1)
acc = accuracy_score(pred, self.y_t)
acc_sel = accuracy_score(pred[idx_gt_threshold],
self.y_t[idx_gt_threshold])
if self.verbosity > 1:
print('selected : ', self.initial_selected_num)
print('trans acc :', acc)
print('trans sel acc :', acc_sel)
# Select label with high confidence
if self.initial_target_labels:
self.X_trans = np.vstack(
[self.X_t_init, self.X_t[idx_gt_threshold]])
self.y_trans = np.concatenate(
[self.y_t_init, pred[idx_gt_threshold]])
else:
self.X_trans = self.X_t[idx_gt_threshold]
self.y_trans = pred[idx_gt_threshold]
self.trained = True