def test_kernel_gpu_cpu_shap(dtype, n_features, n_background, model):
X_train, X_test, y_train, y_test = create_synthetic_dataset(
n_samples=n_background + 3,
n_features=n_features,
test_size=3,
noise=0.1,
dtype=dtype)
mod = model().fit(X_train, y_train)
explainer, shap_values = get_shap_values(model=mod.predict,
background_dataset=X_train,
explained_dataset=X_test,
explainer=KernelExplainer)
exp_v = explainer.expected_value
fx = mod.predict(X_test)
for test_idx in range(3):
assert (np.sum(shap_values[test_idx]) -
abs(fx[test_idx] - exp_v)) <= 1e-5
if has_shap():
import shap
explainer = shap.KernelExplainer(mod.predict, cp.asnumpy(X_train))
cpu_shap_values = explainer.shap_values(cp.asnumpy(X_test))
assert np.allclose(shap_values,
cpu_shap_values,
rtol=1e-01,
atol=1e-01)
def test_kernel_shap_standalone(dtype, n_features, n_background, model):
X_train, X_test, y_train, y_test = create_synthetic_dataset(
n_samples=n_background + 3,
n_features=n_features,
test_size=3,
noise=0.1,
dtype=dtype)
mod = model(n_components=3).fit(X_train, y_train)
explainer, shap_values = get_shap_values(model=mod.transform,
background_dataset=X_train,
explained_dataset=X_test,
explainer=KernelExplainer)
exp_v = explainer.expected_value
# we have 3 lists of shap values, each corresponding to a component since
# transform gives back arrays of shape (nrows x ncomponents)
# we test that for each test row, for each component, the
# sum of the shap values is the same as the difference between the
# expected value for that component minus the value of the transform of
# the row.
for sv_idx in range(3):
# pca and tsvd transform give results back nested
fx = mod.transform(X_test[sv_idx].reshape(1, n_features))[0]
for comp_idx in range(3):
assert (np.sum(shap_values[comp_idx][sv_idx]) -
abs(fx[comp_idx] - exp_v[comp_idx])) <= 1e-5
def test_exact_regression_datasets(exact_shap_regression_dataset, model):
X_train, X_test, y_train, y_test = exact_shap_regression_dataset
models = []
models.append(model().fit(X_train, y_train))
models.append(cuml_skl_class_dict[model]().fit(X_train, y_train))
for mod in models:
explainer, shap_values = get_shap_values(model=mod.predict,
background_dataset=X_train,
explained_dataset=X_test,
explainer=KernelExplainer)
for i in range(3):
print(i)
assert_and_log(shap_values[i], golden_regression_results[model][i],
mod.predict(X_test[i].reshape(1, X_test.shape[1])),
explainer.expected_value)
def test_regression_datasets(exact_shap_regression_dataset, model):
X_train, X_test, y_train, y_test = exact_shap_regression_dataset
models = []
models.append(model().fit(X_train, y_train))
models.append(cuml_skl_class_dict[model]().fit(X_train, y_train))
for mod in models:
explainer, shap_values = get_shap_values(
model=mod.predict,
background_dataset=X_train,
explained_dataset=X_test,
explainer=PermutationExplainer)
fx = mod.predict(X_test)
exp_v = explainer.expected_value
for i in range(3):
assert (np.sum(cp.asnumpy(shap_values[i])) -
abs(fx[i] - exp_v)) <= 1e-5
def test_exact_classification_datasets(exact_shap_classification_dataset):
X_train, X_test, y_train, y_test = exact_shap_classification_dataset
models = []
models.append(cuml.SVC(probability=True).fit(X_train, y_train))
models.append(sklearn.svm.SVC(probability=True).fit(X_train, y_train))
for mod in models:
explainer, shap_values = get_shap_values(model=mod.predict_proba,
background_dataset=X_train,
explained_dataset=X_test,
explainer=KernelExplainer)
# Some values are very small, which mean our tolerance here needs to be
# a little looser to avoid false positives from comparisons like
# 0.00348627 - 0.00247397. The loose tolerance still tests that the
# distribution of the values matches.
for idx, svs in enumerate(shap_values):
assert_and_log(svs[0],
golden_classification_result[idx],
float(mod.predict_proba(X_test)[0][idx]),
explainer.expected_value[idx],
tolerance=1e-01)
def test_exact_classification_datasets(exact_shap_classification_dataset):
X_train, X_test, y_train, y_test = exact_shap_classification_dataset
models = []
models.append(cuml.SVC(probability=True).fit(X_train, y_train))
models.append(sklearn.svm.SVC(probability=True).fit(X_train, y_train))
for mod in models:
explainer, shap_values = get_shap_values(
model=mod.predict_proba,
background_dataset=X_train,
explained_dataset=X_test,
explainer=PermutationExplainer,
)
fx = mod.predict_proba(X_test)
exp_v = explainer.expected_value
for i in range(3):
print(i, fx[i][1], shap_values[1][i])
assert (np.sum(cp.asnumpy(shap_values[0][i])) -
abs(fx[i][0] - exp_v[0])) <= 1e-5
assert (np.sum(cp.asnumpy(shap_values[1][i])) -
abs(fx[i][1] - exp_v[1])) <= 1e-5