def test_theanets_partial_fit():
clf_complete = TheanetsClassifier(layers=[2],
trainers=[{
'algo': 'rmsprop',
'learning_rate': 0.1
}, {
'algo': 'rprop',
'learning_rate': 0.1
}])
clf_partial = TheanetsClassifier(layers=[2],
trainers=[{
'algo': 'rmsprop',
'learning_rate': 0.1
}])
X, y, sample_weight = generate_classification_data()
clf_complete.fit(X, y)
clf_partial.fit(X, y)
clf_partial.partial_fit(X, y, algo='rprop', learning_rate=0.1)
assert clf_complete.trainers == clf_partial.trainers, 'trainers not saved in partial fit'
auc_complete = roc_auc_score(y, clf_complete.predict_proba(X)[:, 1])
auc_partial = roc_auc_score(y, clf_partial.predict_proba(X)[:, 1])
# Known fail of theanets
assert auc_complete == auc_partial, 'same networks return different results'
def test_theanets_reproducibility():
clf = TheanetsClassifier()
X, y, sample_weight = generate_classification_data()
clf.fit(X, y)
auc = roc_auc_score(y, clf.predict_proba(X)[:, 1])
for i in range(2):
clf.fit(X, y)
curr_auc = roc_auc_score(y, clf.predict_proba(X)[:, 1])
assert auc == curr_auc, 'running a network twice produces different results'
cloned_clf = clone(clf)
cloned_clf.fit(X, y)
cloned_auc = roc_auc_score(y, cloned_clf.predict_proba(X)[:, 1])
assert cloned_auc == auc, 'cloned network produces different result'
def test_theanets_partial_fit():
clf_complete = TheanetsClassifier(trainers=[{'optimize': 'rmsprop', 'patience': 0}, {'optimize': 'rprop'}])
clf_partial = TheanetsClassifier(trainers=[{'optimize': 'rmsprop', 'patience': 0}])
X, y, sample_weight = generate_classification_data()
clf_complete.fit(X, y)
clf_partial.fit(X, y)
clf_partial.partial_fit(X, y, optimize='rprop')
assert clf_complete.trainers == clf_partial.trainers, 'trainers not saved in partial fit'
auc_complete = roc_auc_score(y, clf_complete.predict_proba(X)[:, 1])
auc_partial = roc_auc_score(y, clf_partial.predict_proba(X)[:, 1])
assert auc_complete == auc_partial, 'same networks return different results'
def test_theanets_reproducibility():
clf = TheanetsClassifier()
X, y, sample_weight = generate_classification_data()
clf.fit(X, y)
auc = roc_auc_score(y, clf.predict_proba(X)[:, 1])
for i in range(2):
clf.fit(X, y)
curr_auc = roc_auc_score(y, clf.predict_proba(X)[:, 1])
assert auc == curr_auc, 'running a network twice produces different results'
cloned_clf = clone(clf)
cloned_clf.fit(X, y)
cloned_auc = roc_auc_score(y, cloned_clf.predict_proba(X)[:, 1])
assert cloned_auc == auc, 'cloned network produces different result'
def test_theanets_partial_fit():
clf_complete = TheanetsClassifier(layers=[2], trainers=[{'algo': 'rmsprop', 'learning_rate': 0.1},
{'algo': 'rprop', 'learning_rate': 0.1}])
clf_partial = TheanetsClassifier(layers=[2], trainers=[{'algo': 'rmsprop', 'learning_rate': 0.1}])
X, y, sample_weight = generate_classification_data()
clf_complete.fit(X, y)
clf_partial.fit(X, y)
clf_partial.partial_fit(X, y, algo='rprop', learning_rate=0.1)
assert clf_complete.trainers == clf_partial.trainers, 'trainers not saved in partial fit'
auc_complete = roc_auc_score(y, clf_complete.predict_proba(X)[:, 1])
auc_partial = roc_auc_score(y, clf_partial.predict_proba(X)[:, 1])
# Known fail of theanets
assert auc_complete == auc_partial, 'same networks return different results'
def test_theanets_partial_fit():
clf_complete = TheanetsClassifier(trainers=[{
'optimize': 'rmsprop'
}, {
'optimize': 'rprop'
}])
clf_partial = TheanetsClassifier(trainers=[{'optimize': 'rmsprop'}])
X, y, sample_weight = generate_classification_data()
clf_complete.fit(X, y)
clf_partial.fit(X, y)
clf_partial.partial_fit(X, y, optimize='rprop')
assert clf_complete.trainers == clf_partial.trainers, 'trainers not saved in partial fit'
auc_complete = roc_auc_score(y, clf_complete.predict_proba(X)[:, 1])
auc_partial = roc_auc_score(y, clf_partial.predict_proba(X)[:, 1])
assert auc_complete == auc_partial, 'same networks return different results'
