def test_scheduler(self):
scheduler = ReduceOnWorsening()
clf = algorithms.GRAM(max_iter=10, learning_rate=.01, scheduler=scheduler)\
.fit(self.KLtr, self.Ytr)
scheduler = ReduceOnWorsening(multiplier=.6, min_lr=1e-4)
clf = algorithms.GRAM(max_iter=10, learning_rate=.01, scheduler=scheduler)\
.fit(self.KLtr, self.Ytr)
def test_multiclass_ova(self):
mkl = algorithms.GRAM(multiclass_strategy='ova',
learner=SVC(),
max_iter=2).fit(self.KLtr, self.Ytr)
clf = multiclass.OneVsRestMKLClassifier(mkl).fit(self.KLtr, self.Ytr)
classes = np.unique(self.Ytr)
self.assertEqual(len(mkl.solution), len(classes))
for c in classes:
self.assertListEqual(clf.solution[c].weights.tolist(),
mkl.solution[c].weights.tolist())
def test_callbacks(self):
earlystop_auc = callbacks.EarlyStopping(self.KLte,
self.Yte,
patience=5,
cooldown=1,
metric='auc')
earlystop_acc = callbacks.EarlyStopping(self.KLte,
self.Yte,
patience=3,
cooldown=2,
metric='accuracy')
cbks = [earlystop_auc, earlystop_acc]
clf = algorithms.GRAM(max_iter=100, learning_rate=.01, callbacks=cbks)
clf = clf.fit(self.KLtr, self.Ytr)
def test_callbacks(self):
earlystop_auc = callbacks.EarlyStopping(self.KLte,
self.Yte,
patience=30,
cooldown=2,
metric='roc_auc')
earlystop_acc = callbacks.EarlyStopping(self.KLte,
self.Yte,
patience=30,
cooldown=2,
metric='accuracy')
monitor = callbacks.Monitor(
metrics=[metrics.radius, metrics.margin, metrics.frobenius])
cbks = [earlystop_auc, earlystop_acc, monitor]
clf = algorithms.GRAM(max_iter=60, learning_rate=.001, callbacks=cbks)
clf = clf.fit(self.KLtr, self.Ytr)
self.assertEqual(len(monitor.history), 3)
print(monitor.objective)
self.assertEqual(len(monitor.objective), 60)
def test_GRAM(self):
self.base_evaluation(algorithms.GRAM(max_iter=10))
self.base_evaluation(algorithms.GRAM(max_iter=10, learner=SVC(C=10)))
self.base_evaluation(
algorithms.GRAM(max_iter=10, learner=algorithms.KOMD(lam=1)))
