def test_group_pca_matches_sklearn_iris_results(n_components, exclude_groups):
X = iris.data
pca = PCA(n_components=n_components)
X_2 = np.hstack([X, X, X])
groups = [
np.arange(0, X.shape[1]),
np.arange(X.shape[1], 2 * X.shape[1]),
np.arange(2 * X.shape[1], 3 * X.shape[1]),
]
gpca = GroupPCA(n_components=n_components,
groups=groups,
exclude_groups=exclude_groups)
# _ = gpca.fit_transform(X_2)
X_r = pca.fit_transform(X)
X_r2 = gpca.fit_transform(X_2)
hstack_list = [X_r, X_r, X_r]
if exclude_groups is not None:
if isinstance(exclude_groups, int):
hstack_list[exclude_groups] = X
else:
for idx in exclude_groups:
hstack_list[idx] = X
assert_allclose(X_r2, np.hstack(hstack_list))
def test_theta_equals_zero_equals_unsupervised(binarize):
n_components = 3
diabetes = load_diabetes()
X = diabetes.data
y = diabetes.target
if binarize:
y = (y > np.median(y)).astype(int)
groups = [np.arange(0, X.shape[1]), np.arange(X.shape[1], 2 * X.shape[1])]
X = np.hstack([X, X])
gpca = GroupPCA(n_components=n_components, groups=groups)
sgpca = SupervisedGroupPCA(n_components=n_components,
groups=groups,
theta=0.0)
X_r = gpca.fit_transform(X, y)
X_rs = sgpca.fit_transform(X, y)
assert_allclose(X_r, X_rs)
def test_group_pca_inverse_matches_input():
X = iris.data
X_2 = np.hstack([X, X])
groups = [np.arange(0, X.shape[1]), np.arange(X.shape[1], 2 * X.shape[1])]
gpca = GroupPCA(groups=groups, group_names=["one", "two"]).fit(X_2)
Y = gpca.transform(X_2)
Y_inv = gpca.inverse_transform(Y)
assert_allclose(Y_inv, X_2)
assert gpca.get_feature_names() == (
[f"one_pc{idx}" for idx in range(Y.shape[1] // 2)] +
[f"two_pc{idx}" for idx in range(Y.shape[1] // 2)])
with pytest.raises(TypeError):
GroupPCA(groups=groups, exclude_groups="error").fit(X_2)
feature_names = afqdata.feature_names
group_names = afqdata.group_names
subjects = afqdata.subjects
# Here we reduce computation time by taking the first 10 principal components of each feature group and performing SGL logistic regression on those components.
# If you want to train an SGL model without group PCA, set ``do_group_pca = False``. This will increase the number of features by an order of magnitude and slow down execution time.
do_group_pca = True
if do_group_pca:
n_components = 10
# The next three lines retrieve the group structure of the group-wise PCA
# and store it in ``groups_pca``. We do not use the imputer or GroupPCA transformer
# for anything else
imputer = SimpleImputer(strategy="median")
gpca = GroupPCA(n_components=n_components, groups=groups)
groups_pca = gpca.fit(imputer.fit_transform(X)).groups_out_
transformer = GroupPCA
transformer_kwargs = {"groups": groups, "n_components": n_components}
else:
transformer = False
transformer_kwargs = None
pipe = make_afq_classifier_pipeline(
imputer_kwargs={"strategy": "median"}, # Use median imputation
use_cv_estimator=True, # Automatically determine the best hyperparameters
feature_transformer=transformer, # See note above about group PCA
feature_transformer_kwargs=transformer_kwargs,
scaler="standard", # Standard scale the features before regression
groups=groups_pca if do_group_pca else