def test_moving_params(self):
digits = datasets.load_digits()
digits_data = digits.data
digits_target = digits.target
p = Pipeline()
node_data = p.add(NumpyRead(digits_data))
node_target = p.add(NumpyRead(digits_target))
node_split = p.add(SplitTrainTest(2, random_state=0))
# parameters from
# http://scikit-learn.org/stable/auto_examples/plot_classifier_comparison.html
node_clf1 = p.add(
wrap_and_make_instance(RandomForestClassifier,
max_depth=5,
n_estimators=10,
max_features=1,
random_state=0))
node_clf2 = p.add(
wrap_and_make_instance(RandomForestClassifier,
max_depth=12,
n_estimators=100,
max_features=1000))
node_params_out_1 = p.add(
CSVWrite(self._tmp_files.get('out_params_1.csv')))
node_params_out_2 = p.add(
CSVWrite(self._tmp_files.get('out_params_2.csv')))
node_pred_out_1 = p.add(CSVWrite(
self._tmp_files.get('out_pred_1.csv')))
node_pred_out_2 = p.add(CSVWrite(
self._tmp_files.get('out_pred_2.csv')))
node_data['output'] > node_split['input0']
node_target['output'] > node_split['input1']
node_split['train0'] > node_clf1['X_train']
node_split['train1'] > node_clf1['y_train']
node_split['test0'] > node_clf1['X_test']
node_split['train0'] > node_clf2['X_train']
node_split['train1'] > node_clf2['y_train']
node_split['test0'] > node_clf2['X_test']
node_clf1['params_out'] > node_clf2['params_in']
node_clf1['params_out'] > node_params_out_1['input']
node_clf2['params_out'] > node_params_out_2['input']
node_clf1['y_pred'] > node_pred_out_1['input']
node_clf2['y_pred'] > node_pred_out_2['input']
self.run_pipeline(p)
params_1 = self._tmp_files.csv_read('out_params_1.csv')
params_2 = self._tmp_files.csv_read('out_params_2.csv')
self.assertTrue(np.array_equal(params_1, params_2))
y_pred_1 = self._tmp_files.csv_read('out_pred_1.csv')
y_pred_2 = self._tmp_files.csv_read('out_pred_2.csv')
self.assertTrue(np.array_equal(y_pred_1, y_pred_2))
def test_query_complex(self):
p = Pipeline()
csv_in = p.add(CSVRead(path_of_data('query.csv')))
q1_node = p.add(
Query("((id == value) and not (use_this_col == 'no'))"
"or name == 'fish'"))
csv_out = p.add(CSVWrite(self._tmp_files('out.csv')))
csv_comp = p.add(CSVWrite(self._tmp_files('out_comp.csv')))
csv_in['output'] > q1_node['input']
q1_node['output'] > csv_out['input']
q1_node['complement'] > csv_comp['input']
self.run_pipeline(p)
result = self._tmp_files.csv_read('out.csv')
ctrl = csv_read(path_of_data('query_ctrl.csv'))
self.assertTrue(np.array_equal(result, ctrl))
result = self._tmp_files.csv_read('out_comp.csv')
ctrl = csv_read(path_of_data('query_ctrl_comp.csv'))
self.assertTrue(np.array_equal(result, ctrl))
def test_cross_validation_score(self):
rows = 100
folds = 10
X = np.random.random((rows, 10))
y = np.random.randint(0, 2, (rows))
p = Pipeline()
np_in_X = p.add(NumpyRead(X))
np_in_y = p.add(NumpyRead(y))
cv_score = p.add(
CrossValidationScore(wrap(SVC), 'score', {}, folds,
random_state=0))
np_in_X['output'] > cv_score['X_train']
np_in_y['output'] > cv_score['y_train']
score_out = p.add(CSVWrite(self._tmp_files('out.csv')))
cv_score['score'] > score_out['input']
self.run_pipeline(p)
result = self._tmp_files.csv_read('out.csv')['f0']
ctrl_kf = SKKFold(rows, folds, random_state=0)
ctrl = np.mean(cross_val_score(SVC(), X, y, cv=ctrl_kf))
self.assertTrue(np.allclose(ctrl, result))
def test_grid_search(self):
"""
Simulates behavior of example in:
http://scikit-learn.org/stable/modules/generated/sklearn.grid_search.GridSearchCV.html#sklearn.grid_search.GridSearchCV
"""
folds = 2
parameters = {
'kernel': (
'rbf',
'linear'),
'C': [
1,
10,
100],
'random_state': [0]}
iris = datasets.load_iris()
iris_data = iris.data
iris_target = iris.target
p = Pipeline()
node_data = p.add(NumpyRead(iris_data))
node_target = p.add(NumpyRead(iris_target))
node_split = p.add(SplitTrainTest(2, random_state=1))
node_search = p.add(GridSearch(
wrap(SVC),
parameters,
'score',
cv_stage_kwargs={'n_folds': folds}))
node_params_out = p.add(CSVWrite(self._tmp_files.get('out.csv')))
node_data['output'] > node_split['input0']
node_target['output'] > node_split['input1']
node_split['train0'] > node_search['X_train']
node_split['train1'] > node_search['y_train']
node_split['test0'] > node_search['X_test']
node_split['test1'] > node_search['y_test']
node_search['params_out'] > node_params_out['input']
self.run_pipeline(p)
result = self._tmp_files.csv_read('out.csv')
ctrl_X_train, _, ctrl_y_train, _ = train_test_split(
iris_data, iris_target, random_state=1)
ctrl_cv = SKKFold(ctrl_y_train.size, folds)
ctrl_search = grid_search.GridSearchCV(SVC(), parameters, cv=ctrl_cv)
ctrl_search.fit(ctrl_X_train, ctrl_y_train)
control = ctrl_search.best_params_
# TODO a number of configurations tie here, and sklearn picks a different
# best configuration than upsg does (although they have the same score)
# ideally, we want to find some parameters where there is a clear
# winner
control = {'C': 10, 'kernel': 'linear', 'random_state': 0}
self.assertEqual(np_sa_to_dict(np.array([result])), control)
def test_3_stage(self):
from sklearn.preprocessing import Imputer
infile_name = path_of_data('missing_vals.csv')
p = Pipeline()
csv_read_node = p.add(CSVRead(infile_name))
csv_write_node = p.add(CSVWrite(self._tmp_files.get('out.csv')))
impute_node = p.add(wrap_and_make_instance(Imputer))
csv_read_node['output'] > impute_node['X_train']
impute_node['X_new'] > csv_write_node['input']
self.run_pipeline(p)
ctrl_imputer = Imputer()
ctrl_X_sa = np.genfromtxt(infile_name,
dtype=None,
delimiter=",",
names=True)
num_type = ctrl_X_sa[0][0].dtype
ctrl_X_nd, ctrl_X_sa_type = np_sa_to_nd(ctrl_X_sa)
ctrl_X_new_nd = ctrl_imputer.fit_transform(ctrl_X_nd)
control = ctrl_X_new_nd
result = self._tmp_files.csv_read('out.csv', True)
self.assertTrue(np.allclose(result, control))
def test_lambda(self):
# Test output key generation
l1 = LambdaStage(lambda x, y: 0)
self.assertEqual(l1.input_keys, ['x', 'y'])
self.assertEqual(l1.output_keys, [
'output0',
])
l2 = LambdaStage(lambda: 0, n_outputs=3)
self.assertEqual(l2.input_keys, [])
self.assertEqual(l2.output_keys,
['output{}'.format(i) for i in xrange(3)])
# Test running in pipeline
in_data = np_nd_to_sa(np.random.random((100, 10)))
scale = np_nd_to_sa(np.array(3))
out_keys = ['augmented', 'log_col', 'sqrt_col', 'scale_col']
def log1_sqrt2_scale3(A, scale):
names = A.dtype.names
log_col = np.log(A[names[0]])
sqrt_col = np.sqrt(A[names[1]])
scale_col = A[names[2]] * scale[0][0]
return (append_fields(A, ['log1', 'sqrt2', 'scale3'],
(log_col, sqrt_col, scale_col)), log_col,
sqrt_col, scale_col)
p = Pipeline()
np_in = p.add(NumpyRead(in_data))
scale_in = p.add(NumpyRead(scale))
lambda_stage = p.add(LambdaStage(log1_sqrt2_scale3, out_keys))
np_in['output'] > lambda_stage['A']
scale_in['output'] > lambda_stage['scale']
csv_out_stages = []
for key in out_keys:
stage = p.add(CSVWrite(self._tmp_files('out_{}.csv'.format(key))))
csv_out_stages.append(stage)
lambda_stage[key] > stage['input']
self.run_pipeline(p)
controls = log1_sqrt2_scale3(in_data, scale)
for i, key in enumerate(out_keys):
control = controls[i]
if is_sa(control):
control = np_sa_to_nd(control)[0]
result = self._tmp_files.csv_read('out_{}.csv'.format(key),
as_nd=True)
self.assertTrue(np.allclose(control, result))
def test_sql(self):
# Make sure we don't accidentally corrupt our test database
db_path, db_file_name = self._tmp_files.tmp_copy(
path_of_data('small.db'))
db_url = 'sqlite:///{}'.format(db_path)
q_sel_employees = 'CREATE TABLE {tmp_emp} AS SELECT * FROM employees;'
# We have to be careful about the datetime type in sqlite3. It will
# forget if we don't keep reminding it, and if it forgets sqlalchemy
# will be unhappy. Hence, we can't use CREATE TABLE AS if our table
# has a DATETIME
q_sel_hours = ('CREATE TABLE {tmp_hrs} '
'(id INT, employee_id INT, time DATETIME, '
' event_type TEXT); '
'INSERT INTO {tmp_hrs} SELECT * FROM hours;')
q_join = ('CREATE TABLE {joined} '
'(id INT, last_name TEXT, salary REAL, time DATETIME, '
' event_type TEXT); '
'INSERT INTO {joined} '
'SELECT {tmp_emp}.id, last_name, salary, time, event_type '
'FROM {tmp_emp} JOIN {tmp_hrs} ON '
'{tmp_emp}.id = {tmp_hrs}.employee_id;')
p = Pipeline()
get_emp = p.add(RunSQL(db_url, q_sel_employees, [], ['tmp_emp'], {}))
get_hrs = p.add(RunSQL(db_url, q_sel_hours, [], ['tmp_hrs'], {}))
join = p.add(
RunSQL(db_url, q_join, ['tmp_emp', 'tmp_hrs'], ['joined'], {}))
csv_out = p.add(CSVWrite(self._tmp_files('out.csv')))
get_emp['tmp_emp'] > join['tmp_emp']
get_hrs['tmp_hrs'] > join['tmp_hrs']
join['joined'] > csv_out['input']
self.run_pipeline(p)
ctrl = csv_read(path_of_data('test_transform_test_sql_ctrl.csv'))
result = self._tmp_files.csv_read('out.csv')
# Because Numpy insists on printing times with local offsets, but
# not every computer has the same offset, we have to force it back
# into UTC
for i, dt in enumerate(result['time']):
# .item() makes a datetime, which we can format correctly later
# http://stackoverflow.com/questions/25134639/how-to-force-python-print-numpy-datetime64-with-specified-timezone
result['time'][i] = np.datetime64(dt).item().strftime(
'%Y-%m-%dT%H:%M:%S')
# Then we have to make the string field smaller
new_cols = []
for col in result.dtype.names:
new_cols.append(result[col].astype(ctrl.dtype[col]))
result = merge_arrays(new_cols, flatten=True)
result.dtype.names = ctrl.dtype.names
self.assertTrue(np.array_equal(result, ctrl))
def test_split_columns(self):
p = Pipeline()
csv_in = p.add(CSVRead(path_of_data('numbers.csv')))
split = p.add(SplitColumns(('F1', 'F3')))
csv_out_sel = p.add(CSVWrite(self._tmp_files('out_sel.csv')))
csv_out_rest = p.add(CSVWrite(self._tmp_files('out_rest.csv')))
csv_in['output'] > split['input']
split['output'] > csv_out_sel['input']
split['complement'] > csv_out_rest['input']
self.run_pipeline(p)
result = self._tmp_files.csv_read('out_sel.csv')
ctrl = csv_read(path_of_data('test_split_columns_ctrl_selected.csv'))
self.assertTrue(np.array_equal(result, ctrl))
result = self._tmp_files.csv_read('out_rest.csv')
ctrl = csv_read(path_of_data('test_split_columns_ctrl_rest.csv'))
self.assertTrue(np.array_equal(result, ctrl))
def test_fill_na(self):
p = Pipeline()
csv_in = p.add(CSVRead(path_of_data('missing_vals_mixed.csv')))
fill_na = p.add(FillNA(-1))
csv_out = p.add(CSVWrite(self._tmp_files('out.csv')))
csv_in['output'] > fill_na['input']
fill_na['output'] > csv_out['input']
self.run_pipeline(p)
result = self._tmp_files.csv_read('out.csv')
ctrl = csv_read(path_of_data('test_transform_test_fill_na_ctrl.csv'))
self.assertTrue(np.array_equal(result, ctrl))
def test_label_encode(self):
p = Pipeline()
csv_in = p.add(CSVRead(path_of_data('categories.csv')))
le = p.add(LabelEncode())
csv_out = p.add(CSVWrite(self._tmp_files('out.csv')))
csv_in['output'] > le['input']
le['output'] > csv_out['input']
self.run_pipeline(p)
result = self._tmp_files.csv_read('out.csv')
ctrl = csv_read(
path_of_data('test_transform_test_label_encode_ctrl.csv'))
self.assertTrue(np.array_equal(result, ctrl))
def test_rename_cols(self):
infile_name = path_of_data('mixed_csv.csv')
rename_dict = {'name': 'designation', 'height': 'tallness'}
p = Pipeline()
csv_read_node = p.add(CSVRead(infile_name))
trans_node = p.add(RenameCols(rename_dict))
csv_write_node = p.add(CSVWrite(self._tmp_files('out.csv')))
csv_read_node['output'] > trans_node['input']
trans_node['output'] > csv_write_node['input']
self.run_pipeline(p)
control = {'id', 'designation', 'tallness'}
result = set(self._tmp_files.csv_read('out.csv').dtype.names)
self.assertTrue(np.array_equal(result, control))
def test_rw(self):
infile_name = path_of_data('mixed_csv.csv')
p = Pipeline()
csv_read_node = p.add(CSVRead(infile_name))
csv_write_node = p.add(CSVWrite(self._tmp_files.get('out.csv')))
csv_read_node['output'] > csv_write_node['input']
self.run_pipeline(p)
control = np.genfromtxt(infile_name,
dtype=None,
delimiter=",",
names=True)
result = self._tmp_files.csv_read('out.csv')
self.assertTrue(np.array_equal(result, control))
def test_cross_validation_score(self):
rows = 100
folds = 10
X = np.random.random((rows, 10))
y = np.random.randint(0, 2, (rows))
trials = ((SKKFold,
{'random_state': 0, 'n_folds': folds},
{'n': rows, 'n_folds': folds, 'random_state': 0}),
(StratifiedKFold,
{'random_state': 0, 'n_folds': folds},
{'y': y, 'n_folds': folds, 'random_state': 0}))
for PartIter, res_kwargs, ctrl_kwargs in trials:
p = Pipeline()
np_in_X = p.add(NumpyRead(X))
np_in_y = p.add(NumpyRead(y))
cv_score = p.add(CrossValidationScore(
wrap(SVC),
{},
'score',
wrap(PartIter),
res_kwargs))
np_in_X['output'] > cv_score['X_train']
np_in_y['output'] > cv_score['y_train']
score_out = p.add(CSVWrite(self._tmp_files('out.csv')))
cv_score['score'] > score_out['input']
self.run_pipeline(p)
result = self._tmp_files.csv_read('out.csv')['f0']
ctrl_kf = PartIter(**ctrl_kwargs)
ctrl = np.mean(cross_val_score(SVC(), X, y, cv=ctrl_kf))
self.assertTrue(np.allclose(ctrl, result))
def test_kfold(self):
folds = 3
rows = 6
X = np.random.randint(0, 1000, (rows, 3))
y = np.random.randint(0, 1000, (rows, 1))
p = Pipeline()
np_in_X = p.add(NumpyRead(X))
np_in_y = p.add(NumpyRead(y))
kfold = p.add(KFold(2, folds, random_state=0))
np_in_X['output'] > kfold['input0']
np_in_y['output'] > kfold['input1']
ctrl_kf = SKKFold(rows, n_folds=folds, random_state=0)
out_files = []
expected_folds = []
arrays = (X, y)
for fold_i, train_test_inds in enumerate(ctrl_kf):
for array_i, array in enumerate(arrays):
for select_i, selection in enumerate(('train', 'test')):
out_key = '{}{}_{}'.format(selection, array_i, fold_i)
out_file = out_key + '.csv'
out_files.append(out_file)
stage = p.add(CSVWrite(self._tmp_files(out_file)))
kfold[out_key] > stage['input']
slice_inds = train_test_inds[select_i]
expected_folds.append(
np_nd_to_sa(arrays[array_i][slice_inds]))
self.run_pipeline(p)
for out_file, expected_fold in zip(out_files, expected_folds):
self.assertTrue(
np.array_equal(self._tmp_files.csv_read(out_file),
expected_fold))
def test_tutorial(self):
"""
Verifies we can do what sklearn does here:
http://scikit-learn.org/stable/tutorial/basic/tutorial.html
"""
digits = datasets.load_digits()
digits_data = digits.data
# for now, we need a column vector rather than an array
digits_target = digits.target
p = Pipeline()
# load data from a numpy dataset
stage_data = NumpyRead(digits_data)
stage_target = NumpyRead(digits_target)
# train/test split
stage_split_data = SplitTrainTest(2, test_size=1, random_state=0)
# build a classifier
stage_clf = wrap_and_make_instance(SVC, gamma=0.001, C=100.)
# output to a csv
stage_csv = CSVWrite(self._tmp_files('out.csv'))
node_data, node_target, node_split, node_clf, node_csv = map(
p.add,
[stage_data, stage_target, stage_split_data, stage_clf, stage_csv])
# connect the pipeline stages together
node_data['output'] > node_split['input0']
node_target['output'] > node_split['input1']
node_split['train0'] > node_clf['X_train']
node_split['train1'] > node_clf['y_train']
node_split['test0'] > node_clf['X_test']
node_clf['y_pred'] > node_csv['input']
self.run_pipeline(p)
result = self._tmp_files.csv_read('out.csv', True)
# making sure we get the same result as sklearn
clf = SVC(gamma=0.001, C=100.)
# The tutorial just splits using array slicing, but we need to make
# sure that both UPSG and sklearn are splitting the same way, so we
# do something more sophisticated
train_X, test_X, train_y, test_y = train_test_split(digits_data,
digits_target,
test_size=1,
random_state=0)
clf.fit(train_X, np.ravel(train_y))
control = clf.predict(test_X)[0]
self.assertAlmostEqual(result, control)
# model persistance
s = pickle.dumps(stage_clf)
stage_clf2 = pickle.loads(s)
self.assertEqual(stage_clf.get_params(), stage_clf2.get_params())
def __simple_pipeline(self,
sk_cls,
sk_method_name,
upsg_out_key,
init_kwargs={},
in_data=None):
X_in, y_in = self.__process_in_data(in_data)
ctrl_sk_inst = sk_cls(**init_kwargs)
est_params = ctrl_sk_inst.get_params()
try:
random_state = est_params['random_state']
if random_state is None:
# This has to be fixed. Set a state and try again
init_kwargs['random_state'] = 0
ctrl_sk_inst = sk_cls(**init_kwargs)
except KeyError:
pass
p = Pipeline()
sk_stage = p.add(wrap_and_make_instance(sk_cls, **init_kwargs))
X_in_stage = p.add(NumpyRead(X_in))
y_in_stage = p.add(NumpyRead(y_in))
if sk_method_name == 'predict':
train_test = p.add(SplitTrainTest(2, random_state=0))
X_in_stage['output'] > train_test['input0']
y_in_stage['output'] > train_test['input1']
input_keys = sk_stage.get_stage().input_keys
if 'X_train' in input_keys:
train_test['train0'] > sk_stage['X_train']
if 'X_test' in input_keys:
train_test['test0'] > sk_stage['X_test']
if 'y_train' in input_keys:
train_test['train1'] > sk_stage['y_train']
else:
X_in_stage['output'] > sk_stage['X_train']
y_in_stage['output'] > sk_stage['y_train']
csv_out = p.add(CSVWrite(self._tmp_files.get('out.csv')))
sk_stage[upsg_out_key] > csv_out['input']
self.run_pipeline(p)
if sk_method_name == 'predict':
ctrl_X_train, ctrl_X_test, ctrl_y_train, ctrl_y_test = (
train_test_split(X_in, y_in, random_state=0))
ctrl_sk_inst.fit(ctrl_X_train, ctrl_y_train)
control = ctrl_sk_inst.predict(ctrl_X_test)
else:
control = ctrl_sk_inst.fit_transform(X_in, y_in)
result = self._tmp_files.csv_read('out.csv', as_nd=True)
if result.ndim != control.ndim and result.ndim == 1:
result = result.reshape(result.size, 1)
self.assertTrue(result.shape == control.shape
and np.allclose(result, control))
def __metric_pipeline(self, metric, params={}, in_data=None):
X_in, y_in = self.__process_in_data(in_data)
metric_stage = wrap_and_make_instance(metric, **params)
in_keys = metric_stage.input_keys
out_keys = metric_stage.output_keys
p = Pipeline()
node_X_in = p.add(NumpyRead(X_in))
node_y_in = p.add(NumpyRead(y_in))
node_split = p.add(SplitTrainTest(2, random_state=0))
node_X_in['output'] > node_split['input0']
node_y_in['output'] > node_split['input1']
ctrl_X_train, ctrl_X_test, ctrl_y_train, ctrl_y_test = (
train_test_split(X_in, y_in, random_state=0))
node_clf = p.add(wrap_and_make_instance(SVC, random_state=0))
node_split['train0'] > node_clf['X_train']
node_split['train1'] > node_clf['y_train']
node_split['test0'] > node_clf['X_test']
ctrl_clf = SVC(random_state=0, probability=True)
ctrl_clf.fit(ctrl_X_train, ctrl_y_train)
node_proba_1 = p.add(SplitY(1))
node_clf['pred_proba'] > node_proba_1['input']
ctrl_y_score = ctrl_clf.predict_proba(ctrl_X_test)[:, 1]
node_metric = p.add(metric_stage)
ctrl_metric_args = {}
if 'y_true' in in_keys:
node_split['test1'] > node_metric['y_true']
ctrl_metric_args['y_true'] = ctrl_y_test
if 'y_score' in in_keys:
node_proba_1['y'] > node_metric['y_score']
ctrl_metric_args['y_score'] = ctrl_y_score
if 'probas_pred' in in_keys:
node_proba_1['y'] > node_metric['probas_pred']
ctrl_metric_args['probas_pred'] = ctrl_y_score
out_nodes = [
p.add(CSVWrite(self._tmp_files('out_{}.csv'.format(out_key))))
for out_key in out_keys
]
[
node_metric[out_key] > out_nodes[i]['input']
for i, out_key in enumerate(out_keys)
]
self.run_pipeline(p)
ctrl_returns = metric(**ctrl_metric_args)
if len(out_keys) == 1:
ctrl_returns = (ctrl_returns, )
for i, out_key in enumerate(out_keys):
control = ctrl_returns[i]
result = self._tmp_files.csv_read('out_{}.csv'.format(out_key),
as_nd=True)
self.assertTrue(result.shape == control.shape
and np.allclose(result, control))