def test_save_and_load_meta(tmp_path):
classifier = SemanticEquivalenceMetaClassifier(n_numerical_features=1,
pretrained="bert",
batch_size=2,
eval_batch_size=1)
# Save and load for Meta Models only accepts strings (not PosixPath)
classifier._initialise_models()
classifier.save(str(tmp_path.absolute()) + '.h5')
config_1 = classifier.config
classifier = SemanticEquivalenceMetaClassifier(n_numerical_features=1)
classifier.load(str(tmp_path.absolute()) + '.h5')
config_2 = classifier.config
assert config_1 == config_2
def test_semantic_meta_fit():
classifier = SemanticEquivalenceMetaClassifier(n_numerical_features=2,
pretrained="scibert",
batch_size=6,
eval_batch_size=4,
dropout_rate=0.1,
batch_norm=True)
X = [[
'This sentence has context_1', 'This one also has context_1', 0.1, 0.2
], [
'This sentence has context_2', 'This one also has context_2', 0.2, 0.2
],
[
'This sentence is about something else', 'God save the queen',
-0.5, -0.5
]] * 5
y = [1, 1, 0] * 5
classifier.fit(X, y, epochs=5)
# loss_initial = classifier.history['loss'][0]
scores = classifier.predict_proba(X)
# Assert it returns a vector of correct length (15 training points) and 15*2 scores
assert len(classifier.predict(X)) == 15
assert (scores > 0).sum() == 15 * 2
assert (scores < 1).sum() == 15 * 2
def test_semantic_meta_fit():
classifier = SemanticEquivalenceMetaClassifier(n_numerical_features=2,
pretrained="scibert",
batch_size=2,
eval_batch_size=1,
dropout_rate=0.1,
batch_norm=True)
X = [[
'This sentence has context_1', 'This one also has context_1', 0.1, 0.2
], [
'This sentence has context_2', 'This one also has context_2', 0.2, 0.2
],
[
'This sentence is about something else', 'God save the queen',
-0.5, -0.5
]]
y = [1, 1, 0]
classifier.fit(X, y, epochs=3)
loss_initial = classifier.history['loss'][0]
scores = classifier.score(X)
assert len(classifier.predict(X)) == 3
assert (scores > 0).sum() == 6
assert (scores < 1).sum() == 6
# Fits two extra epochs
classifier.fit(X, y, epochs=2)
# Asserts that the classifier model is adding to the history, and still
# not re-training from scratch
assert len(classifier.history['loss']) == 5
loss_final = classifier.history['loss'][4]
# Asserts loss is decreasing
assert loss_final < loss_initial