def test_neuroquery_model():
x, y, voc = _dataset_and_voc()
vect = tokenization.TextVectorizer.from_vocabulary(voc)
reg = smoothed_regression.SmoothedRegression(n_components=10).fit(x, y)
encoder = encoding.NeuroQueryModel(vect,
reg,
mask_img=_mask_img(y.shape[1]))
text = "feature0 and feature8 compared to feature73"
res = encoder(text)
simil = res["similar_words"]
assert simil.loc["feature0"]["similarity"] != 0
assert simil.loc["feature0"]["weight_in_brain_map"] != 0
assert simil.loc["feature0"]["weight_in_query"] != 0
assert simil.loc["feature8"]["weight_in_query"] != 0
assert simil.loc["feature8"]["similarity"] != 0
assert simil.loc["feature8"]["weight_in_brain_map"] == pytest.approx(0)
assert simil.loc["feature18"]["weight_in_brain_map"] == pytest.approx(0)
assert simil.loc["feature18"]["weight_in_query"] == pytest.approx(0)
assert res["similar_documents"] is None
with tempfile.TemporaryDirectory() as tmp_dir:
encoder.to_data_dir(tmp_dir)
loaded = encoding.NeuroQueryModel.from_data_dir(tmp_dir)
assert not loaded.vectorizer.add_unigrams
encoded = loaded(text)["brain_map"].get_data()
assert np.allclose(encoded, res["brain_map"].get_data())
assert res["z_map"] is res["brain_map"]
n_docs = 4
tfidf = np.zeros((n_docs, x.shape[1]))
tfidf[:n_docs, :n_docs] = np.eye(n_docs)
metadata = pd.DataFrame.from_dict({"id": np.arange(n_docs)})
encoder = encoding.NeuroQueryModel(
vect,
reg,
mask_img=_mask_img(y.shape[1]),
corpus_info={
"tfidf": tfidf,
"metadata": metadata
},
)
for i in range(n_docs):
res = encoder(encoder.full_vocabulary()[i])
assert res["similar_documents"]["id"][0] == i
assert res["similar_words"]["n_documents"][0] == 1
with tempfile.TemporaryDirectory() as tmp_dir:
encoder.to_data_dir(tmp_dir)
loaded = encoding.NeuroQueryModel.from_data_dir(tmp_dir)
assert not loaded.vectorizer.add_unigrams
for i in range(n_docs):
res = encoder(encoder.full_vocabulary()[i])
assert res["similar_documents"].id[0] == i
def test_z_maps():
rng = np.random.RandomState(0)
X = rng.binomial(3, 0.3, size=(21, 9)).astype("float64")
Y = rng.randn(21, 11)
reg = smoothed_regression.SmoothedRegression(n_components=5).fit(X, Y)
z = reg.transform_to_z_maps(X)
assert z.shape == Y.shape
with tempfile.TemporaryDirectory() as tmp_dir:
reg.to_data_dir(tmp_dir)
loaded = smoothed_regression.SmoothedRegression.from_data_dir(tmp_dir)
assert np.allclose(loaded.transform_to_z_maps(X),
reg.transform_to_z_maps(X))
def test_predictions():
x, y = datasets.make_regression(
n_samples=102,
n_informative=5,
n_features=91,
n_targets=117,
effective_rank=9,
noise=0.5,
shuffle=False,
random_state=0,
)
x -= x.min() - 1
reg = smoothed_regression.SmoothedRegression(n_components=5,
smoothing_weight=1e-3)
print(reg.fit(x, y).predict(x))
sk_reg = RidgeCV()
score = cross_val_score(reg, x, y, cv=5)
sk_score = cross_val_score(sk_reg, x, y, cv=5)
assert score.mean() > sk_score.mean()
def test_text_to_brain():
x, y, voc = _dataset_and_voc()
vect = tokenization.TextVectorizer.from_vocabulary(voc)
reg = smoothed_regression.SmoothedRegression(n_components=10).fit(x, y)
encoder = text_to_brain.TextToBrain(vect,
reg,
mask_img=_mask_img(y.shape[1]))
text = "feature0 and feature8 but not feature73"
res = encoder(text)
simil = res["similar_words"]
assert simil.loc["feature0"]["similarity"] != 0
assert simil.loc["feature0"]["weight_in_brain_map"] != 0
assert simil.loc["feature0"]["weight_in_query"] != 0
assert simil.loc["feature8"]["weight_in_query"] != 0
assert simil.loc["feature8"]["similarity"] != 0
assert simil.loc["feature8"]["weight_in_brain_map"] == pytest.approx(0)
assert simil.loc["feature18"]["weight_in_brain_map"] == pytest.approx(0)
assert simil.loc["feature18"]["weight_in_query"] == pytest.approx(0)
with tempfile.TemporaryDirectory() as tmp_dir:
encoder.to_data_dir(tmp_dir)
loaded = text_to_brain.TextToBrain.from_data_dir(tmp_dir)
encoded = loaded(text)["z_map"].get_data()
assert np.allclose(encoded, res["z_map"].get_data())