def test_simple_column_transformer(self):
if ColumnTransformer is None:
return
data = numpy.array([[0, 0], [0, 0], [2, 1], [2, 1]],
dtype=numpy.float32)
model = ColumnTransformer([("scaler1", StandardScaler(), [0]),
("scaler2", RobustScaler(), [1])])
model.fit(data)
all_models = list(enumerate_pipeline_models(model))
steps = collect_intermediate_steps(model, "coulmn transformer",
[("input",
FloatTensorType([None, 2]))])
assert len(steps) == 2
assert len(all_models) == 3
model.transform(data)
for step in steps:
onnx_step = step['onnx_step']
sess = onnxruntime.InferenceSession(onnx_step.SerializeToString())
onnx_outputs = sess.run(None, {'input': data})
onnx_output = onnx_outputs[0]
skl_outputs = step['model']._debug.outputs['transform']
assert_almost_equal(onnx_output, skl_outputs)
compare_objects(onnx_output.tolist(), skl_outputs.tolist())
def test_simple_pipeline_predict_proba(self):
data = load_iris()
X, y = data.data, data.target
model = Pipeline([("scaler1", StandardScaler()),
("lr", LogisticRegression())])
model.fit(X, y)
all_models = list(enumerate_pipeline_models(model))
steps = collect_intermediate_steps(
model, "pipeline",
[("input", FloatTensorType([None, X.shape[1]]))])
assert len(steps) == 2
assert len(all_models) == 3
model.predict_proba(X)
for step in steps:
onnx_step = step['onnx_step']
sess = onnxruntime.InferenceSession(onnx_step.SerializeToString())
onnx_outputs = sess.run(None, {'input': X.astype(numpy.float32)})
dbg_outputs = step['model']._debug.outputs
if 'transform' in dbg_outputs:
onnx_output = onnx_outputs[0]
skl_outputs = dbg_outputs['transform']
else:
onnx_output = onnx_outputs[1]
skl_outputs = dbg_outputs['predict_proba']
assert_almost_equal(onnx_output, skl_outputs, decimal=6)
compare_objects(onnx_output, skl_outputs)
def test_simple_pipeline(self):
data = numpy.array([[0, 0], [0, 0], [2, 1], [2, 1]],
dtype=numpy.float32)
model = Pipeline([("scaler1", StandardScaler()),
("scaler2", StandardScaler())])
model.fit(data)
all_models = list(enumerate_pipeline_models(model))
steps = collect_intermediate_steps(model, "pipeline",
[("input",
FloatTensorType([None, 2]))])
assert len(steps) == 2
assert len(all_models) == 3
model.transform(data)
for step in steps:
onnx_step = step['onnx_step']
sess = onnxruntime.InferenceSession(onnx_step.SerializeToString())
onnx_outputs = sess.run(None, {'input': data})
onnx_output = onnx_outputs[0]
skl_outputs = step['model']._debug.outputs['transform']
assert str(step['model']._debug) is not None
sdt = step['model']._debug.display(data, 5)
assert 'shape' in sdt
assert_almost_equal(onnx_output, skl_outputs)
compare_objects(onnx_output, skl_outputs)
def test_missing_converter(self):
data = numpy.array([[0, 0], [0, 0], [2, 1], [2, 1]],
dtype=numpy.float32)
model = Pipeline([("scaler1", StandardScaler()),
("scaler2", StandardScaler()),
("scaler3", MyScaler())])
model.fit(data)
all_models = list(enumerate_pipeline_models(model))
try:
collect_intermediate_steps(model,
"pipeline",
[("input", FloatTensorType([None, 2]))],
target_opset=TARGET_OPSET)
except MissingShapeCalculator as e:
assert "MyScaler" in str(e)
assert "gallery" in str(e)
_alter_model_for_debugging(model, recursive=True)
model.transform(data)
all_models = list(enumerate_pipeline_models(model))
for ind, step, last in all_models:
if ind == (0, ):
# whole pipeline
continue
step_model = step
data_in = step_model._debug.inputs['transform']
t = guess_data_type(data_in)
try:
onnx_step = convert_sklearn(step_model,
initial_types=t,
target_opset=TARGET_OPSET)
except MissingShapeCalculator as e:
if "MyScaler" in str(e):
continue
raise
sess = onnxruntime.InferenceSession(onnx_step.SerializeToString())
onnx_outputs = sess.run(None, {'input': data_in})
onnx_output = onnx_outputs[0]
skl_outputs = step_model._debug.outputs['transform']
assert_almost_equal(onnx_output, skl_outputs)
compare_objects(onnx_output, skl_outputs)
def test_simple_pipeline(self):
for opset in (11, TARGET_OPSET):
if opset > TARGET_OPSET:
continue
data = numpy.array([[0, 0], [0, 0], [2, 1], [2, 1]],
dtype=numpy.float32)
model = Pipeline([("scaler1", StandardScaler()),
("scaler2", StandardScaler())])
model.fit(data)
all_models = list(enumerate_pipeline_models(model))
steps = collect_intermediate_steps(
model,
"pipeline", [("input", FloatTensorType([None, 2]))],
target_opset=opset)
assert len(steps) == 2
assert len(all_models) == 3
expected = 'version:%d}' % opset
expected1 = 'version:1}'
model.transform(data)
for step in steps:
onnx_step = step['onnx_step']
text = str(onnx_step).replace('\n', ' ').replace(' ', '')
if expected not in text and expected1 not in text:
raise AssertionError("Unable to find '{}'\n'{}'\n".format(
expected, text))
sess = onnxruntime.InferenceSession(
onnx_step.SerializeToString())
onnx_outputs = sess.run(None, {'input': data})
onnx_output = onnx_outputs[0]
skl_outputs = step['model']._debug.outputs['transform']
assert str(step['model']._debug) is not None
sdt = step['model']._debug.display(data, 5)
assert 'shape' in sdt
assert_almost_equal(onnx_output, skl_outputs)
compare_objects(onnx_output, skl_outputs)
def test_simple_feature_union(self):
data = numpy.array([[0, 0], [0, 0], [2, 1], [2, 1]],
dtype=numpy.float32)
model = FeatureUnion([("scaler1", StandardScaler()),
("scaler2", RobustScaler())])
model.fit(data)
all_models = list(enumerate_pipeline_models(model))
steps = collect_intermediate_steps(
model,
"feature union", [("input", FloatTensorType([None, 2]))],
target_opset=TARGET_OPSET)
assert len(steps) == 2
assert len(all_models) == 3
model.transform(data)
for step in steps:
onnx_step = step['onnx_step']
sess = onnxruntime.InferenceSession(onnx_step.SerializeToString())
onnx_outputs = sess.run(None, {'input': data})
onnx_output = onnx_outputs[0]
skl_outputs = step['model']._debug.outputs['transform']
assert_almost_equal(onnx_output, skl_outputs)
compare_objects(onnx_output, skl_outputs)
print("onnxruntime")
print(timeit("sess.run(None, {'input': X_digits[:1].astype(np.float32)})[1]",
number=10000, globals=globals()))
###############################################
# Intermediate steps
# ++++++++++++++++++
#
# Let's imagine the final output is wrong and we need
# to look into each component of the pipeline which one
# is failing. The following method modifies the scikit-learn
# pipeline to steal the intermediate outputs and produces
# an smaller ONNX graph for every operator.
steps = collect_intermediate_steps(pipe, "pipeline",
initial_types)
assert len(steps) == 2
pipe.predict_proba(X_digits[:2])
for i, step in enumerate(steps):
onnx_step = step['onnx_step']
sess = rt.InferenceSession(onnx_step.SerializeToString())
onnx_outputs = sess.run(None, {'input': X_digits[:2].astype(np.float32)})
skl_outputs = step['model']._debug.outputs
if 'transform' in skl_outputs:
compare_objects(skl_outputs['transform'], onnx_outputs[0])
print("benchmark", step['model'].__class__)
print("scikit-learn")
print(timeit("step['model'].transform(X_digits[:1])",
from skl2onnx.common.data_types import FloatTensorType
###########################
# The pipeline.
data = load_iris()
X = data.data
pipe = Pipeline(steps=[('std', StandardScaler()), ('km', KMeans(3))])
pipe.fit(X)
#################################
# The function goes through every step,
# overloads the methods *transform* and
# returns an ONNX graph for every step.
steps = collect_intermediate_steps(
pipe, "pipeline", [("X", FloatTensorType([None, X.shape[1]]))])
#####################################
# We call method transform to population the
# cache the overloaded methods *transform* keeps.
pipe.transform(X)
#######################################
# We compute every step and compare
# ONNX and scikit-learn outputs.
for step in steps:
print('----------------------------')
print(step['model'])
onnx_step = step['onnx_step']
sess = InferenceSession(onnx_step.SerializeToString())
timeit("sess.run(None, {'input': X_digits[:1].astype(np.float32)})[1]",
number=10000,
globals=globals()))
###############################################
# Intermediate steps
# ++++++++++++++++++
#
# Let's imagine the final output is wrong and we need
# to look into each component of the pipeline which one
# is failing. The following method modifies the scikit-learn
# pipeline to steal the intermediate outputs and produces
# an smaller ONNX graph for every operator.
steps = collect_intermediate_steps(pipe,
"pipeline",
initial_types,
target_opset=11)
assert len(steps) == 2
pipe.predict_proba(X_digits[:2])
for i, step in enumerate(steps):
onnx_step = step['onnx_step']
sess = rt.InferenceSession(onnx_step.SerializeToString())
onnx_outputs = sess.run(None, {'input': X_digits[:2].astype(np.float32)})
skl_outputs = step['model']._debug.outputs
if 'transform' in skl_outputs:
compare_objects(skl_outputs['transform'], onnx_outputs[0])
print("benchmark", step['model'].__class__)
print("scikit-learn")