def test_preprocessing():
"""Tests feature preprocessing"""
base_prob = mock_problem()
base_prob.features.append('discrete_feat')
# Derive a problem with a single discrete feature perfectly correlated with the label
df = pd.DataFrame(base_prob.dataframe, copy=True)
df['discrete_feat'] = 'negative'
df['discrete_feat'].values[base_prob.y == 1] = 'positive'
# Verify that a manual upfront vectorize is equivalent to passing a vectorizer as the preprocess step
# to SelectAndClassify
prob = base_prob.set_data(df)
vectorized_prob = ProblemVectorizer().fit_apply(prob)
baseline_classifier = SelectAndClassify(SelectKBest(k='all'), LogisticRegression(), preprocess=None)
preprocess_classifier = SelectAndClassify(SelectKBest(k='all'), LogisticRegression(),
preprocess=ProblemVectorizer())
# First make sure that the baseline classifier cannot be fit on the unvectorized data
nose.tools.assert_raises(ValueError, lambda: baseline_classifier.fit_apply(prob))
baseline_scores = baseline_classifier.fit_apply(vectorized_prob)
preprocess_scores = preprocess_classifier.fit_apply(prob)
np.testing.assert_allclose(baseline_scores, preprocess_scores)
def test_null_feature_selector():
"""Validates that SelectAndClassify works with a null feature selector"""
def make_fixed_rs():
"""Utility: makes a fixed random state for use in this test"""
return np.random.RandomState(0xC0FFEE)
prob = mock_problem()
# selector=None and SelectKBest(k='all') should produce identical predictions
no_select_approach = SelectAndClassify(None, LogisticRegression(random_state=make_fixed_rs()),
classifier_grid={'C': [0.5, 1.0]},
random_state=make_fixed_rs()).fit(prob)
select_all_approach = SelectAndClassify(SelectKBest(k='all'),
LogisticRegression(random_state=make_fixed_rs()),
classifier_grid={'C': [0.5, 1.0]},
random_state=make_fixed_rs()).fit(prob)
# There should be no selection step in the underlying model
nose.tools.eq_(len(no_select_approach.model.steps), len(select_all_approach.model.steps) - 1)
# We should still be logging the right features
nose.tools.assert_list_equal(no_select_approach.selected_features, prob.features)
# Scores should be identical as k='all'
np.testing.assert_allclose(no_select_approach.apply(prob), select_all_approach.apply(prob))
def test_multiclass(working_dir):
""" Tests machine learning classification workfloor with multiclass for iris dataset
see http://scikit-learn.org/stable/modules/multiclass.html """
out_dir = os.path.join(working_dir, 'learn_output')
model_path = os.path.join(out_dir, 'model.txt')
iris = datasets.load_iris()
df = iris_to_df(iris)
features = [feat for feat in df.columns if feat not in ['Target']]
prob = Problem(df, features, "Target", positive_outcome=None)
rnd = np.random.RandomState(2016)
approach = SelectAndClassify(SelectKBest(score_func=f_pearson, k=3),
RandomForestClassifier(random_state=rnd))
learn_params = LearningParameters(metrics={
'auc':
roc_auc_score,
'accuracy':
accuracy_from_confusion_matrix
},
treat_as_binary=False)
cvg = CVSplitGenerator(prob,
n_folds=10,
n_repartitions=10,
random_state=rnd)
cv = CrossValidatedAnalysis(prob,
approach,
cv_generator=cvg,
runner=SerialRunner(),
params=learn_params)
results = cv.run()
renderer = ReportRenderer(out_dir)
ClassificationReport(renderer, False, prob.label_list).generate(results)
nose.tools.ok_(
os.path.exists(os.path.join(out_dir, 'sample_confusion_matrix.txt')))
average_accuracy = compute_average_accuracy(results)
nose.tools.assert_almost_equal(0.95, average_accuracy, delta=0.01)
classifier = SelectAndClassify(SelectKBest(score_func=f_pearson, k=3),
RandomForestClassifier(random_state=2016),
name='test multiclass model').fit(prob)
model = ClassificationModel(classifier, prob)
model.write(model_path)
read_model = ClassificationModel.read(model_path)
auc_average = read_model.training_auc
nose.tools.assert_almost_equal(1.0, auc_average, delta=1e-6)
def checkme(working_dir, n_samples, n_features, k, make_classifier,
test_vectorize):
"""Utility"""
assert n_samples % 4 == 0
model_path = os.path.join(working_dir, 'model.txt')
prob = mock_problem(n_samples=n_samples, n_features=n_features)
if test_vectorize:
df = prob.dataframe
df['discrete_1'] = ['foo', 'bar'] * int(n_samples / 2)
df['discrete_2'] = ['foo', 'bar', 'baz',
float('nan')] * int(n_samples / 4)
df['continuous_with_missing'] = [0, 1, 2, float('nan')] * int(
n_samples / 4)
prob = Problem(
df, prob.features +
['discrete_1', 'discrete_2', 'continuous_with_missing'],
prob.outcome_column, prob.positive_outcome)
preprocess = ProblemVectorizer()
else:
preprocess = None
approach = SelectAndClassify(SelectKBest(k=k),
make_classifier(),
preprocess=preprocess).fit(prob)
model = ClassificationModel(approach, prob)
model.write(model_path)
reconstituted_model = ClassificationModel.read(model_path)
model.validate()
reconstituted_model.validate()
np.testing.assert_array_equal(model.approach.apply(prob),
reconstituted_model.approach.apply(prob))
if preprocess is not None:
approach_pipeline = ApproachPipeline([('preprocess', preprocess)])
approach_with_pipeline = SelectAndClassify(
SelectKBest(k=k),
make_classifier(),
preprocess=approach_pipeline).fit(prob)
# test approach serialization with Pipeline from learners.py
model_with_pipeline = ClassificationModel(approach_with_pipeline,
prob)
model_path2 = os.path.join(working_dir, 'model2.txt')
model_with_pipeline.write(model_path2)
reconstituted_model2 = ClassificationModel.read(model_path2)
reconstituted_model2.validate()
np.testing.assert_array_almost_equal(
model.approach.apply(prob),
reconstituted_model2.approach.apply(prob), 14)
def checkme(selector_grid, classifier_grid, expected_grid, expected_size):
"""Utility"""
learner = SelectAndClassify(SelectKBest(), LogisticRegression(), selector_grid=selector_grid,
classifier_grid=classifier_grid)
actual_grid, actual_grid_size = learner._get_grid() # pylint: disable=protected-access
nose.tools.eq_(actual_grid, expected_grid)
nose.tools.eq_(actual_grid_size, expected_size)
def test_predict_tool(working_dir):
"""Tests that the predict.py command line tool works as expected"""
out_dir = os.path.join(working_dir, 'learn_output')
model_path = os.path.join(out_dir, 'model.txt')
predictions_path = os.path.join(out_dir, 'predictions.txt')
# Mock up some input data
prob_path, prob = mock_input(working_dir)
os.mkdir(out_dir)
# Train a model and save it to a file
classifier = SelectAndClassify(SelectKBest(k=5),
GaussianNB(),
name='test model').fit(prob)
model = ClassificationModel(classifier, prob)
model.write(model_path)
# Run the predict tool with the model using the training data loaded from a file, and validate that
# the returned predictions match
predict.main(
[model_path, prob_path, predictions_path, '--index_col', 'sample_id'])
expected_predictions = pd.DataFrame({
'sample_id': prob.sample_ids,
'score': classifier.apply(prob)
})
actual_predictions = pd.read_csv(predictions_path, sep='\t')
np.testing.assert_allclose(actual_predictions['score'].values,
expected_predictions['score'].values)
def checkme(n_samples, n_features, n_features_to_select, theta):
"""Utility"""
prob = mock_problem(n_samples, n_features, theta)
train = prob[prob.dataframe['train_or_test'] == 'train']
test = prob[prob.dataframe['train_or_test'] == 'test']
approach = LoggingApproach(SelectAndClassify(SelectKBest(k=n_features_to_select), LogisticRegression()))
results = train_and_evaluate_model(approach, train, test, LearningParameters())
model = results['approach']['model']
# Simple sanity checks
nose.tools.assert_is_not_none(model)
nose.tools.eq_(results['train']['n_samples'], len(train))
nose.tools.eq_(results['test']['n_samples'], len(test))
nose.tools.eq_(len(results['approach']['selected_features']), n_features_to_select)
nose.tools.assert_greater(results['train']['metrics']['auc'], 0.75)
nose.tools.assert_greater(results['test']['metrics']['auc'], 0.75)
nose.tools.assert_greater(results['train']['metrics']['auc'], results['test']['metrics']['auc'])
for name, sub_prob in [('train', train), ('test', test)]:
nose.tools.assert_list_equal(results[name]['outcome'], list(sub_prob.dataframe[sub_prob.outcome_column]))
nose.tools.eq_(results[name]['positive_outcome'], 1)
# Make sure we trained on the right samples
nose.tools.eq_(len(model.fit_problems), 1) # fitting the same model multiple times might cause problems
np.testing.assert_array_equal(train.X, model.fit_problems[-1].X)
np.testing.assert_array_equal(train.y, model.fit_problems[-1].y)
# Make sure we applied the model to both training and test
for X in [train.X, test.X]:
nose.tools.ok_(any([np.array_equal(X, applied_prob.X) for applied_prob in model.apply_problems]))
def test_comparative_report():
"""Verifies that we correctly generate a classification report"""
features = ['f1', 'f2', 'f3', 'f4', 'f5']
cv_results_1 = mock_cv_results_list(
features, 'my test approach 1', [['f1', 'f2'], ['f2', 'f5']],
[[0.1, 0.2, 0.8, 1.0] * 25, [0.0, 0.1, 0.7, 0.9] * 25],
[[0.1, 0.2, 0.8, 1.0] * 25, [0.0, 0.1, 0.7, 0.9] * 25], [0.9, 1.0],
[0.66, 0.76], [0.05, 0.01])
cv_results_2 = mock_cv_results_list(
features, 'my test approach 1', [['f2', 'f3'], ['f3', 'f4']],
[[0.1, 0.2, 0.8, 1.0] * 25, [0.0, 0.1, 0.7, 0.9] * 25],
[[0.1, 0.2, 0.8, 1.0] * 25, [0.0, 0.1, 0.7, 0.9] * 25], [0.95, 0.97],
[0.85, 0.78], [0.05, 0.01])
approach1 = SelectAndClassify(SelectKBest(k=2),
LogisticRegression(),
name="logit")
approach2 = SelectAndClassify(SelectKBest(k=3), GaussianNB(), name="nb")
cv_results = {approach1: cv_results_1, approach2: cv_results_2}
renderer = LoggingRenderer()
report = ComparativeClassificationReport(renderer)
report.generate(cv_results)
for expected_plot in ['score_plots']:
nose.tools.assert_true(expected_plot in renderer.plots)
for expected_table in ['mean_scores', 'mean_metrics']:
nose.tools.assert_true(expected_table in renderer.tables)
# Should have the exact same number of entries for each approach
groups = dict(list(
renderer.tables[expected_table].groupby('approach')))
nose.tools.eq_(set(groups.keys()), {'logit', 'nb'})
assert_lengths_equal(*list(groups.values()))
metrics_df = renderer.tables["mean_metrics"]
nose.tools.assert_equal(
dict(list(zip(metrics_df["approach"], metrics_df["train_auc"]))), {
'logit': 0.95,
'nb': 0.96
})
nose.tools.assert_equal(
dict(list(zip(metrics_df["approach"], metrics_df["test_auc"]))), {
'logit': 0.71,
'nb': 0.815
})
def make_learning_approaches(self):
"""Creates LearningApproaches from the learner options"""
for k in self.n_features:
for cls in self.classifiers:
yield SelectAndClassify(SelectKBest(k=k),
cls,
name='SelectKBest(k={}) -> {}'.format(
k, cls.__class__.__name__),
preprocess=ProblemVectorizer())
def mock_model():
"""Creates a simple mock model for testing"""
prob = mock_problem()
logit = SelectAndClassify(selector=None,
classifier=LogisticRegression(),
preprocess=ProblemVectorizer(),
name="test model").fit(prob)
return ClassificationModel(logit, prob)
def test_overlapping_train_and_test():
"""Validates that we fail if samples overlap between training and test"""
prob = mock_problem()
train = prob[prob.dataframe['train_or_test'] == 'train']
test = prob[prob.dataframe['train_or_test'] == 'test']
approach = SelectAndClassify(SelectKBest(k='all'), LogisticRegression())
params = LearningParameters()
train_and_evaluate_model(approach, train, test, params) # no overlap -- should just work
nose.tools.assert_raises(ValueError, lambda: train_and_evaluate_model(approach, train, train, params)) # oops
nose.tools.assert_raises(ValueError, lambda: train_and_evaluate_model(approach, test, test, params)) # oops
def checkme(selector_grid, classifier_grid, optimal_params):
"""Utility: runs grid search and verifies that we selected the right parameters"""
prob = mock_problem()
learner = SelectAndClassify(SelectKBest(), LogisticRegression(), selector_grid=selector_grid,
classifier_grid=classifier_grid,
grid_search_scorer=make_test_grid_scorer(optimal_params),
grid_search_cv_folds=2, grid_search_cv_repartitions=1,
randomized_grid_size_cutoff=None)
model_params = learner.fit(prob).model.get_params()
params_to_check = sorted(optimal_params.keys())
nose.tools.assert_list_equal([(k, model_params[k]) for k in params_to_check],
[(k, optimal_params[k]) for k in params_to_check])
def test_feature_engineering():
"""Tests feature engineering"""
prob = mock_problem()
classifier = SelectAndClassify(SelectKBest(k='all'), LogisticRegression(), feature_engineering=PCA(n_components=2))
model = classifier.fit(prob).model
steps = dict(model.steps)
nose.tools.ok_('PCA' in str(classifier))
nose.tools.ok_('feature_engineering' in steps)
nose.tools.assert_is_not_none(steps['feature_engineering'].components_)
# Check that classifier.apply() works
nose.tools.eq_(len(classifier.apply(prob)), prob.n_samples)
# Test that SelectAndClassify still works without feature engineering
classifier = SelectAndClassify(SelectKBest(k='all'), LogisticRegression())
model = classifier.fit(prob).model
steps = dict(model.steps)
nose.tools.ok_('PCA' not in str(classifier))
nose.tools.ok_('feature_engineering' not in steps)
def test_multiclass_auc():
""" Tests auc value for multiclass problem"""
data = []
class_values = ['A', 'B', 'C', 'D']
for index_class in range(4):
data, _ = mock_coords_data(data, index_class,
class_values[index_class], None, True)
df = pd.DataFrame(columns=['coord0', 'coord1', 'class'], data=data)
prob = Problem(df, ['coord0', 'coord1'], 'class', None)
classifier = SelectAndClassify(SelectKBest(k='all'),
LogisticRegression(),
name='test multiclass model').fit(prob)
model = ClassificationModel(classifier, prob)
auc_average = model.training_auc
nose.tools.assert_almost_equal(0.853333333, auc_average, delta=1e-6)
prob_binary = Problem(df, ['coord0', 'coord1'], 'class', 'A')
classifier_binary = SelectAndClassify(SelectKBest(k='all'),
LogisticRegression(),
name='binary model').fit(prob_binary)
model_binary = ClassificationModel(classifier_binary, prob_binary)
auc_binary = model_binary.training_auc
nose.tools.assert_almost_equal(auc_binary, auc_average, delta=1e-6)
def checkme(optimal_params):
"""Utility: runs grid search and verifies that we selected (approximately) the right parameters"""
np.random.seed(0xC0FFEE)
prob = mock_problem(n_samples=100)
learner = SelectAndClassify(SelectKBest(), LogisticRegression(), selector_grid={'k': [10, 20]},
classifier_grid={'C': np.linspace(0.5, 1.0, 1000)},
grid_search_scorer=make_test_grid_scorer(optimal_params),
grid_search_cv_folds=2, grid_search_cv_repartitions=1,
randomized_grid_size_cutoff=100)
model_params = learner.fit(prob).model.get_params()
for param_name in sorted(optimal_params.keys()):
# Might not be exactly optimal, but should be close
tolerance = 0.05 * abs(optimal_params[param_name])
nose.tools.assert_almost_equal(model_params[param_name], optimal_params[param_name],
delta=tolerance)
error_log.append(make_test_grid_scorer(optimal_params)(learner.model, prob.X, prob.y))
def cv_analysis_run_one(task_runner=None):
"""\
Does one CV analysis run, then validates and returns the results
:param task_runner: Task runner to use for the CV analysis mockup
:return: CV analysis results
"""
prob = mock_problem(n_samples=1000, n_features=100)
approach = SelectAndClassify(SelectKBest(k=17), LogisticRegression(random_state=np.random.RandomState(0xC0FFEE)))
cv_generator = CVSplitGenerator(prob, n_folds=10, n_repartitions=2, random_state=np.random.RandomState(0xC0FFEE))
analysis = CrossValidatedAnalysis(prob, approach, cv_generator=cv_generator, runner=task_runner)
results = analysis.run()
nose.tools.eq_(len(results), cv_generator.n_total_splits) # One per CV split
for field_name in ['metrics', 'n_samples']:
nose.tools.ok_(all([field_name in r['train'] for r in results]))
nose.tools.ok_(all([field_name in r['test'] for r in results]))
return results
def learning_curves_run_one(fractions=None):
"""Runs a single LearningCurves analysis and validates the output"""
prob = mock_problem(n_samples=1000, n_features=100)
approach = SelectAndClassify(SelectKBest(k=17), LogisticRegression(random_state=np.random.RandomState(0xC0FFEE)))
cv_generator = CVSplitGenerator(prob, n_folds=10, n_repartitions=2, random_state=np.random.RandomState(0xC0FFEE))
analysis = LearningCurveAnalysis(prob, approach, cv_generator=cv_generator, fractions=fractions,
runner=SerialRunner())
results = analysis.run()
nose.tools.eq_(len(results), len(fractions)) # One per fraction
for fraction in sorted(results):
nose.tools.eq_(len(results[fraction]), cv_generator.n_total_splits)
for field_name in ['metrics', 'n_samples']:
nose.tools.ok_(all([field_name in r['train'] for r in results[fraction]]))
nose.tools.ok_(all([field_name in r['test'] for r in results[fraction]]))
seen_test_samples = Counter()
for split_results in results[fraction]:
# No train/test overlap
nose.tools.eq_(set(split_results['train']['sample']) & set(split_results['test']['sample']), set())
# Size of the test set should be 10% of problem size
nose.tools.eq_(len(split_results['test']['sample']), 0.1 * prob.n_samples)
# Size of the training set should be 90% of problem size * fraction
nose.tools.assert_almost_equal(len(split_results['train']['sample']), 0.9 * fraction * prob.n_samples,
delta=1)
# Record test samples
seen_test_samples.update(split_results['test']['sample'])
# Must have seen all test samples, all of them the same number of times
nose.tools.eq_(set(seen_test_samples.keys()), set(prob.sample_ids))
nose.tools.eq_(set(seen_test_samples.values()), set([cv_generator.n_repartitions]))
# Test sets should be identical across fractions. Otherwise difference between fractions will be a product
# of both the training set size and different CV splits, but we only really care about the former.
test_sets_by_fraction = {fraction: tuple([tuple(sorted(set(split_results['test']['sample'])))
for split_results in results[fraction]])
for fraction in sorted(results.keys())}
nose.tools.eq_(len(set(test_sets_by_fraction.values())), 1)
return results
def test_binary_report_with_score_vector():
" Test that in binary case score as vector contains same data as with positive outcome only"
data = []
class_values = ['A', 'B']
for index_class in range(4):
data = mock_coords_data(data,
index_class,
class_values[index_class % 2],
data2=None,
append_missed=False)[0]
df = pd.DataFrame(columns=['coord0', 'coord1', 'class'], data=data)
prob = Problem(df, ['coord0', 'coord1'], 'class', 'B')
classifier = SelectAndClassify(
SelectKBest(k='all'),
LogisticRegression(),
name='test binary with score vector').fit(prob)
y_score_positive = classifier.apply(prob)
y_score_all = classifier.apply(prob, False)
nose.tools.ok_(np.allclose(y_score_positive, y_score_all[:, 1]))
def test_model_validation(working_dir):
"""Validates that we fail if a model has been corrupted or otherwise produces bad output"""
model_path = os.path.join(working_dir, 'model.txt')
prob = mock_problem()
approach = SelectAndClassify(SelectKBest(k=7),
LogisticRegression()).fit(prob)
model = ClassificationModel(approach, prob)
model.write(model_path)
# Change an expected score for a sample -- this should cause model loading to fail because actual
# classifier output will no longer match the expected output
with open(model_path, 'r') as f:
model_string = '\n'.join(f.readlines())
nose.tools.ok_(str(model.expected_scores[17]) in model_string)
bad_model_string = model_string.replace(
str(model.expected_scores[17]),
str(model.expected_scores[17] + 0.5))
with open(model_path, 'w') as f:
f.write(bad_model_string)
nose.tools.assert_raises(ValueError,
lambda: ClassificationModel.read(model_path))
def test_multiclass_label_subset():
""" Tests y_score for multiclass problem with training set
having subset of possible classes """
data = []
data2 = []
class_values = ['A', 'B', 'C', 'D']
for index_class in range(4):
data, data2 = mock_coords_data(data, index_class,
class_values[index_class], data2, True)
df = pd.DataFrame(columns=['coord0', 'coord1', 'class'], data=data)
prob = Problem(df, ['coord0', 'coord1'], 'class', None)
df2 = pd.DataFrame(columns=['coord0', 'coord1', 'class'], data=data2)
prob2 = Problem(df2, ['coord0', 'coord1'], 'class', None, prob.label_list)
classifier = SelectAndClassify(SelectKBest(k='all'),
LogisticRegression(),
name='test multiclass model').fit(prob2)
y_pred = classifier.predict(prob2)
y_score = classifier.prediction_probabilities(prob2)
# check that "C" class has probabilities 0
for i_row in range(y_pred.shape[0]):
nose.tools.assert_almost_equal(0.0, y_score[i_row, 2], delta=1e-6)