def test_multiprocessing_speedup(self, verbose=0):
if n_cpus > 1:
warnings.filterwarnings("ignore")
parallel.start_if_needed(n_cpus=n_cpus)
X, y = self.X_cls, self.y_cls = make_classification(n_classes=2, n_samples=500, n_features=40,
n_informative=20, random_state=test_seed)
mclf = MultichannelPipeline(n_channels=1)
mclf.add_layer(DummyClassifier(futile_cycles_fit=2000000, futile_cycles_pred=10))
# shut off warnings because ray and redis generate massive numbers
SETUP_CODE = '''
import pipecaster.cross_validation'''
TEST_CODE = '''
pipecaster.cross_validation.cross_val_score(mclf, [X], y, cv = 5, n_processes = 1)'''
t_serial = timeit.timeit(setup = SETUP_CODE,
stmt = TEST_CODE,
globals = locals(),
number = 5)
TEST_CODE = '''
pipecaster.cross_validation.cross_val_score(mclf, [X], y, cv = 5, n_processes = {})'''.format(n_cpus)
t_parallel = timeit.timeit(setup = SETUP_CODE,
stmt = TEST_CODE,
globals = locals(),
number = 5)
warnings.resetwarnings()
if verbose > 0:
print('serial run mean time = {} s'.format(t_serial))
print('parallel run mean time = {} s'.format(t_parallel))
if t_serial <= t_parallel:
warnings.warn('mulitple cpus detected, but parallel cross_val_score not faster than serial, possible problem with multiprocessing')
def test_multi_input_regression_parallel(self):
if n_cpus > 1:
warnings.filterwarnings("ignore")
parallel.start_if_needed(n_cpus=n_cpus)
mrgr = MultichannelPipeline(n_channels=1)
mrgr.add_layer(self.rgr)
pc_scores = pc_cross_validation.cross_val_score(mrgr, [self.X_rgr], self.y_rgr, scorer=explained_variance_score,
cv=self.cv, n_processes=n_cpus)
self.assertTrue(np.array_equal(self.rgr_scores, pc_scores), 'regressor scores from pipecaster.cross_validation.cross_val_score did not match sklearn control (multi input predictor)')
warnings.resetwarnings()
def test_multi_input_classification_parallel(self):
if n_cpus > 1:
warnings.filterwarnings("ignore")
parallel.start_if_needed()
mclf = MultichannelPipeline(n_channels=1)
mclf.add_layer(self.clf)
pc_scores = pc_cross_validation.cross_val_score(
mclf, [self.X_cls], self.y_cls, score_method='predict_proba',
scorer=roc_auc_score, cv=self.cv, n_processes=n_cpus)
self.assertTrue(np.array_equal(self.cls_scores, pc_scores), 'classifier scores from pipecaster.cross_validation.cross_val_score did not match sklearn control (multi input predictor)')
warnings.resetwarnings()
def test_multi_input_classification_parallel(self):
if n_cpus > 1:
warnings.filterwarnings("ignore")
parallel.start_if_needed(n_cpus=n_cpus)
mclf = MultichannelPipeline(n_channels=1)
mclf.add_layer(self.clf)
pc_predictions = pc_cross_validation.cross_val_predict(
mclf, [self.X_cls], self.y_cls, cv=self.cv, n_processes=n_cpus)
self.assertTrue(
np.array_equal(self.cls_predictions,
pc_predictions['predict']['y_pred']),
'pipecaster predictions did not match sklearn control')
warnings.resetwarnings()
def test_throttled_multiprocessing_speedup(self, verbose=0):
if n_cpus > 1:
warnings.filterwarnings("ignore")
parallel.start_if_needed(n_cpus=n_cpus)
X, y = self.X_cls, self.y_cls = make_classification(
n_classes=2,
n_samples=500,
n_features=40,
n_informative=20,
random_state=test_seed)
mclf = MultichannelPipeline(n_channels=1)
mclf.add_layer(
DummyClassifier(futile_cycles_fit=2000000,
futile_cycles_pred=10))
SETUP_CODE = '''
import pipecaster.cross_validation'''
TEST_CODE = '''
pipecaster.cross_validation.cross_val_predict(mclf, [X], y, cv = {}, n_processes = 1)'''.format(
n_cpus - 1)
t_serial = timeit.timeit(setup=SETUP_CODE,
stmt=TEST_CODE,
globals=locals(),
number=5)
TEST_CODE = '''
pipecaster.cross_validation.cross_val_predict(mclf, [X], y, cv = {}, n_processes = {})'''.format(
n_cpus - 1, n_cpus - 1)
t_parallel = timeit.timeit(setup=SETUP_CODE,
stmt=TEST_CODE,
globals=locals(),
number=5)
warnings.resetwarnings()
if verbose > 0:
print(
'number of CPUs detected and parallel jobs requested: {}'.
format(n_cpus))
print('duration of serial cross cross_val_predict task: {} s'.
format(t_serial))
print(
'duration of parallel cross cross_val_predict task (ray pool.starmap): {} s'
.format(t_parallel))
if t_serial <= t_parallel:
warnings.warn(
'multiple cpus detected, but parallel cross_val_predict not faster than serial using ray.multiprocessing.starmap(), possible problem with multiprocessing'
)
def test_multi_input_regression_parallel_starmap(self):
if n_cpus > 2:
warnings.filterwarnings("ignore")
parallel.start_if_needed(n_cpus=n_cpus)
mrgr = MultichannelPipeline(n_channels=1)
mrgr.add_layer(self.rgr)
pc_predictions = pc_cross_validation.cross_val_predict(
mrgr, [self.X_rgr],
self.y_rgr,
cv=self.cv,
n_processes=n_cpus - 1)
self.assertTrue(
np.array_equal(self.rgr_predictions,
pc_predictions['predict']['y_pred']),
'pipecaster predictions did not match sklearn '
'control')
warnings.resetwarnings()