def test_sub_output_double(self):
data = load_iris()
X = data.data
dec = DecorrelateTransformer2()
dec.fit(X)
update_registered_converter(DecorrelateTransformer2,
"SklearnDecorrelateTransformer2",
decorrelate_transformer_shape_calculator,
decorrelate_transformer_convertor2)
onx = to_onnx(dec, X.astype(np.float64), target_opset=TARGET_OPSET)
sess = InferenceSession(onx.SerializeToString())
exp = dec.transform(X.astype(np.float64))
got = sess.run(None, {'X': X.astype(np.float64)})[0]
assert_almost_equal(got, exp, decimal=4)
def test_decisiontree_regressor_decision_leaf(self):
model = DecisionTreeRegressor(max_depth=2)
X, y = make_classification(10, n_features=4, random_state=42)
X = X[:, :2]
model.fit(X, y)
initial_types = [('input', FloatTensorType((None, X.shape[1])))]
model_onnx = convert_sklearn(
model,
initial_types=initial_types,
options={id(model): {
'decision_leaf': True
}})
sess = InferenceSession(model_onnx.SerializeToString())
res = sess.run(None, {'input': X.astype(np.float32)})
pred = model.predict(X)
assert_almost_equal(pred, res[0].ravel())
dec = model.decision_path(X)
exp = path_to_leaf(model.tree_, dec.todense())
assert exp.tolist() == res[1].ravel().tolist()
def check_outputs(self,
model,
model_onnx,
Xtest,
predict_attributes,
decimal=5,
skip_if_float32=False,
disable_optimisation=True):
if "TransposeScaleMatMul" in str(model_onnx):
raise RuntimeError("This node must not be added.")
if predict_attributes is None:
predict_attributes = {}
exp = model.predict(Xtest, **predict_attributes)
if disable_optimisation and GraphOptimizationLevel is not None:
opts = SessionOptions()
opts.graph_optimization_level = (
GraphOptimizationLevel.ORT_DISABLE_ALL)
sess = InferenceSession(model_onnx.SerializeToString(),
sess_options=opts)
else:
sess = InferenceSession(model_onnx.SerializeToString())
got = sess.run(None, {'X': Xtest})
if isinstance(exp, tuple):
if len(exp) != len(got):
raise AssertionError("Mismatched number of outputs.")
for i, (e, g) in enumerate(zip(exp, got)):
if skip_if_float32 and g.dtype == np.float32:
continue
try:
assert_almost_equal(self.remove_dim1(e),
self.remove_dim1(g),
decimal=decimal)
except AssertionError as e: # noqa
raise AssertionError(
"Mismatch for output {} and attributes {}"
".".format(i, predict_attributes)) from e
else:
if skip_if_float32 and Xtest.dtype == np.float32:
return
assert_almost_equal(np.squeeze(exp),
np.squeeze(got),
decimal=decimal)
class OnnxWholeSession:
"""
Runs the prediction for a single :epkg:`ONNX`,
it lets the runtime handle the graph logic as well.
"""
def __init__(self, onnx_data, runtime):
"""
@param onnx_data :epkg:`ONNX` model or data
@param runtime runtime to be used,
mostly :epkg:`onnxruntime`
"""
if runtime != 'onnxruntime1':
raise NotImplementedError(
"runtime '{}' is not implemented.".format(runtime))
if hasattr(onnx_data, 'SerializeToString'):
onnx_data = onnx_data.SerializeToString()
self.runtime = runtime
sess_options = SessionOptions()
self.run_options = RunOptions()
try:
sess_options.session_log_severity_level = 3
# sess_options.sessions_log_verbosity_level = 0
except AttributeError:
# onnxruntime not recent enough.
pass
try:
self.run_options.run_log_severity_level = 3
# self.run_options.run_log_verbosity_level = 0
except AttributeError:
# onnxruntime not recent enough.
pass
self.sess = InferenceSession(onnx_data, sess_options=sess_options)
def run(self, inputs):
"""
Computes the predictions.
@param inputs dictionary *{variable, value}*
@return list of outputs
"""
return self.sess.run(None, inputs, self.run_options)
def test_onnx_init_sparse_coo(self):
row = np.array([0, 0, 1, 3, 1], dtype=np.float32)
col = np.array([0, 2, 1, 3, 1], dtype=np.float32)
data = np.array([1, 1, 1, 1, 1], dtype=np.float32)
X = coo_matrix((data, (row, col)), shape=(4, 4))
node = OnnxAdd('X',
X,
output_names=['Y'],
op_version=onnx.defs.onnx_opset_version())
model_def = node.to_onnx({'X': X}, outputs=[('Y', FloatTensorType())])
try:
sess = InferenceSession(model_def.SerializeToString())
except (RuntimeError, OrtInvalidArgument):
# Sparse tensor is not supported for constant.
return
res = sess.run(None, {'X': X})[0]
assert_almost_equal(X + X, res)
def test_randomforestclassifier_decision_path(self):
model = RandomForestClassifier(max_depth=2, n_estimators=2)
X, y = make_classification(3, n_features=4, random_state=42)
X = X[:, :2]
model.fit(X, y)
initial_types = [('input', FloatTensorType((None, X.shape[1])))]
model_onnx = convert_sklearn(
model, initial_types=initial_types,
options={id(model): {'decision_path': True, 'zipmap': False}},
target_opset=TARGET_OPSET)
sess = InferenceSession(model_onnx.SerializeToString())
res = sess.run(None, {'input': X.astype(numpy.float32)})
pred = model.predict(X)
assert_almost_equal(pred, res[0].ravel())
prob = model.predict_proba(X)
assert_almost_equal(prob, res[1])
dec = model.decision_path(X)
exp = binary_array_to_string(dec[0].todense())
got = numpy.array([''.join(row) for row in res[2]])
assert exp == got.ravel().tolist()
def test_multi_output_classifier_fallback(self):
X, y = make_multilabel_classification(n_classes=3, random_state=0)
X = X.astype(numpy.float32)
clf = MultiOutputClassifier(LogisticRegression()).fit(X, y)
del clf.classes_
onx = to_onnx(clf,
X[:1],
target_opset=TARGET_OPSET,
options={
'zipmap': False,
'output_class_labels': True
})
sess = InferenceSession(onx.SerializeToString())
res = sess.run(None, {'X': X})
exp_lab = clf.predict(X)
exp_prb = clf.predict_proba(X)
assert_almost_equal(exp_lab, res[0])
self.assertEqual(len(exp_prb), len(res[1]))
for e, g in zip(exp_prb, res[1]):
assert_almost_equal(e, g, decimal=5)
def verify(self, onnx_model_path):
"""
:param onnx_model_path: onnx model path.
"""
print("Checking ONNX model loading from: {}".format(onnx_model_path))
try:
onnx_options = SessionOptions()
sess = InferenceSession(onnx_model_path, onnx_options, providers=["CUDAExecutionProvider"])
print("Model correctly loaded")
if self.sample_inputs is not None:
inputs_onnx = {k: v.numpy() for k, v in self.sample_inputs.items()}
print(f"Model inputs name: {[tmp_obj.name for tmp_obj in sess.get_inputs()]}")
print(f"Model inputs shape: {[tmp_obj.shape for tmp_obj in sess.get_inputs()]}")
print(f"Model outputs name: {[tmp_obj.name for tmp_obj in sess.get_outputs()]}")
print(f"Model outputs shape: {[tmp_obj.shape for tmp_obj in sess.get_outputs()]}")
# Run the model (None = get all the outputs)
outputs_onnx = sess.run(None, inputs_onnx)
print("Model inference correctly")
except RuntimeException as re:
print("Error while loading the model: {}".format(re))
def test_pipeline_make_column_selector(self):
X = pandas.DataFrame({
'city': ['London', 'London', 'Paris', 'Sallisaw'],
'rating': [5, 3, 4, 5]
})
X['rating'] = X['rating'].astype(numpy.float32)
ct = make_column_transformer(
(StandardScaler(),
make_column_selector(dtype_include=numpy.number)),
(OneHotEncoder(), make_column_selector(dtype_include=object)))
expected = ct.fit_transform(X)
onx = to_onnx(ct, X, target_opset=TARGET_OPSET)
sess = InferenceSession(onx.SerializeToString())
names = [i.name for i in sess.get_inputs()]
got = sess.run(
None, {
names[0]: X[names[0]].values.reshape((-1, 1)),
names[1]: X[names[1]].values.reshape((-1, 1))
})
assert_almost_equal(expected, got[0])
def test_local_outlier_factor_p3(self):
lof = LocalOutlierFactor(n_neighbors=2, novelty=True, p=3)
data = np.array([[-1.1, -1.2], [0.3, 0.2], [0.5, 0.4], [100., 99.]],
dtype=np.float32)
model = lof.fit(data)
model_onnx = to_onnx(model, data, target_opset=TARGET_OPSET)
self.assertNotIn('CDist', str(model_onnx))
data = data.copy()
data[:, 0] += 0.1
sess = InferenceSession(model_onnx.SerializeToString())
names = [o.name for o in sess.get_outputs()]
self.assertEqual(names, ['label', 'scores'])
got = sess.run(None, {'X': data})
self.assertEqual(len(got), 2)
expected_label = lof.predict(data)
expected_decif = lof.decision_function(data)
assert_almost_equal(expected_label, got[0].ravel())
assert_almost_equal(expected_decif, got[1].ravel(), decimal=5)
def test_container_init(self):
onx = OnnxReshapeApi13(OnnxReshapeApi13('X',
np.array([1, -1],
dtype=np.int64),
op_version=TARGET_OPSET),
np.array([1, -1], dtype=np.int64),
output_names=['Y'],
op_version=TARGET_OPSET)
X = np.array([[1, 2], [3, 4]], dtype=np.float32)
model_def = onx.to_onnx({'X': X},
outputs=[('Y', FloatTensorType([None, 2]))],
target_opset=TARGET_OPSET)
sess = InferenceSession(model_def.SerializeToString())
got = sess.run(None, {'X': X})[0]
assert_almost_equal(X.reshape((1, -1)), got)
inits = [
row for row in str(model_def).split('\n')
if row.startswith(" initializer {")
]
self.assertEqual(len(inits), 1)
def test_isolation_forest_score_samples(self):
isol = IsolationForest(n_estimators=3, random_state=0)
data = np.array([[-1.1, -1.2], [0.3, 0.2], [0.5, 0.4], [100., 99.]],
dtype=np.float32)
model = isol.fit(data)
model_onnx = to_onnx(model,
data,
target_opset=TARGET_OPSET,
options={'score_samples': True})
sess = InferenceSession(model_onnx.SerializeToString())
names = [o.name for o in sess.get_outputs()]
self.assertEqual(names, ['label', 'scores', 'score_samples'])
got = sess.run(None, {'X': data})
self.assertEqual(len(got), 3)
expected_label = isol.predict(data)
expected_decif = isol.decision_function(data)
expected_score = isol.score_samples(data)
assert_almost_equal(expected_label, got[0].ravel())
assert_almost_equal(expected_decif, got[1].ravel())
assert_almost_equal(expected_score, got[2].ravel())
def test_rf_regressor_decision_leaf(self):
model = RandomForestRegressor(n_estimators=2, max_depth=3)
X, y = make_regression(10, n_features=4, random_state=42)
X = X[:, :2]
model.fit(X, y)
initial_types = [('input', FloatTensorType((None, X.shape[1])))]
model_onnx = convert_sklearn(
model,
initial_types=initial_types,
options={id(model): {
'decision_leaf': True
}},
target_opset=TARGET_OPSET)
sess = InferenceSession(model_onnx.SerializeToString())
res = sess.run(None, {'input': X.astype(numpy.float32)})
pred = model.predict(X)
assert_almost_equal(pred, res[0].ravel(), decimal=4)
dec = model.decision_path(X)
exp = path_to_leaf(model.estimators_, dec[0].todense(), dec[1])
assert exp.tolist() == res[1].tolist()
def test_onnx_subgraphs1(self):
x = numpy.array([1, 2, 4, 5, 5, 4]).astype(numpy.float32).reshape(
(3, 2))
cop = OnnxAdd(OnnxIdentity('input', op_version=TARGET_OPSET),
'input',
op_version=TARGET_OPSET)
cdist = onnx_squareform_pdist(cop,
dtype=numpy.float32,
op_version=TARGET_OPSET)
cop2 = OnnxIdentity(cdist,
output_names=['cdist'],
op_version=TARGET_OPSET)
model_def = cop2.to_onnx({'input': FloatTensorType([None, None])},
outputs=[('cdist',
FloatTensorType([None, None]))],
target_opset=TARGET_OPSET)
sess = InferenceSession(model_def.SerializeToString())
res = sess.run(None, {'input': x})
self.assertEqual(len(res), 1)
def test_extratreesregressor_decision_path(self):
model = ExtraTreesRegressor(max_depth=2, n_estimators=2)
X, y = make_classification(10, n_features=4, random_state=42)
X = X[:, :2]
model.fit(X, y)
initial_types = [('input', FloatTensorType((None, X.shape[1])))]
model_onnx = convert_sklearn(
model,
initial_types=initial_types,
options={id(model): {
'decision_path': True
}})
sess = InferenceSession(model_onnx.SerializeToString())
res = sess.run(None, {'input': X.astype(numpy.float32)})
pred = model.predict(X)
assert_almost_equal(pred, res[0].ravel())
dec = model.decision_path(X)
exp = binary_array_to_string(dec[0].todense())
got = numpy.array([''.join(row) for row in res[1]])
assert exp == got.ravel().tolist()
def test_onnx_if_algebra_direct(self):
opv = TARGET_OPSET
x1 = np.array([[0, 3], [7, 0]], dtype=np.float32)
x2 = np.array([[1, 0], [2, 0]], dtype=np.float32)
node = OnnxAdd('x1', 'x2', output_names=['absxythen'], op_version=opv)
then_body = node.to_onnx({
'x1': x1,
'x2': x2
},
target_opset=opv,
outputs=[('absxythen', FloatTensorType())])
node = OnnxSub('x1', 'x2', output_names=['absxyelse'], op_version=opv)
else_body = node.to_onnx({
'x1': x1,
'x2': x2
},
target_opset=opv,
outputs=[('absxyelse', FloatTensorType())])
del else_body.graph.input[:]
del then_body.graph.input[:]
cond = OnnxGreater(OnnxReduceSum('x1', op_version=opv),
OnnxReduceSum('x2', op_version=opv),
op_version=opv)
ifnode = OnnxIf(cond,
then_branch=then_body.graph,
else_branch=else_body.graph,
op_version=opv,
output_names=['y'])
model_def = ifnode.to_onnx({
'x1': x1,
'x2': x2
},
target_opset=opv,
outputs=[('y', FloatTensorType())])
sess = InferenceSession(model_def.SerializeToString())
res = sess.run(None, {'x1': x1, 'x2': x2})
assert_almost_equal(x1 + x2, res[0])
def test_kernel_cosine_double(self):
ker = PairwiseKernel(metric='cosine')
onx = convert_kernel(ker, 'X', output_names=['Y'], dtype=np.float64,
op_version=_TARGET_OPSET_)
model_onnx = onx.to_onnx(
inputs=[('X', DoubleTensorType([None, None]))],
target_opset=TARGET_OPSET)
x = np.random.randn(4, 3)
x[0, 0] = x[1, 1] = x[2, 2] = 10.
x[3, 2] = 5.
try:
sess = InferenceSession(model_onnx.SerializeToString())
except NotImplemented:
# Failed to find kernel for FusedMatMul(1).
return
res = sess.run(None, {'X': x.astype(np.float64)})[0]
m1 = res
m2 = ker(x)
assert_almost_equal(m1, m2, decimal=5)
def test_model_tfidf_vectorizer_nan(self):
corpus = numpy.array([
"This is the first document.",
"This document is the second document.",
"And this is the third one.",
"Is this the first document?",
]).reshape((4, 1))
vect = TfidfVectorizer(ngram_range=(1, 1), norm=None)
vect.fit(corpus.ravel())
options = copy.deepcopy(self.get_options())
options[TfidfVectorizer]['nan'] = True
model_onnx = convert_sklearn(vect,
"TfidfVectorizer",
[("input", StringTensorType())],
options=options,
target_opset=TARGET_OPSET)
sess = InferenceSession(model_onnx.SerializeToString())
res = sess.run(None, {'input': corpus.ravel()})[0]
assert res.shape == (4, 9)
assert numpy.isnan(res[0, 0])
def test_custom_ordinal_woe(self):
update_registered_converter(OrdinalWOETransformer,
"OrdinalWOETransformer",
ordwoe_encoder_shape_calculator,
ordwoe_encoder_converter,
parser=ordwoe_encoder_parser)
data = load_iris()
X, y = data.data, data.target
X = X.astype(np.int64)[:, :2]
y = (y == 2).astype(np.int64)
ordwoe = OrdinalWOETransformer()
ordwoe.fit(X, y)
expected = ordwoe.transform(X)
onx = to_onnx(ordwoe, X, target_opset=TARGET_OPSET)
sess = InferenceSession(onx.SerializeToString())
got = sess.run(None, {'X': X})[0]
assert_almost_equal(expected, got)
def common_check_alpha(self, name, fct):
onx = function_onnx_graph(name,
target_opset=get_max_opset(),
dtype=numpy.float32)
x1 = numpy.random.randn(10, 1).astype(numpy.float32)
x2 = numpy.random.randn(10, 1).astype(numpy.float32)
alpha = numpy.random.randn(1).astype(numpy.float32)
fin = fct(x1, x2, alpha)
oinf = OnnxInference(onx)
got = oinf.run({'X1': x1, 'X2': x2, 'alpha': alpha})
self.assertEqualArray(fin, got['Y'], decimal=5)
providers = device_to_providers('cpu')
so = SessionOptions()
so.log_severity_level = 4
sess = InferenceSession(onx.SerializeToString(),
so,
providers=providers)
got = sess.run(None, {'X1': x1, 'X2': x2, 'alpha': alpha})
self.assertEqualArray(fin, got[0], decimal=5)
def test_onnx_test_knn_transform(self):
iris = datasets.load_iris()
X, _ = iris.data, iris.target
X_train, X_test = train_test_split(X, random_state=11)
clr = NearestNeighbors(n_neighbors=3, radius=None)
clr.fit(X_train)
for to in (9, 10, 11):
if to > onnx_opset_version():
break
model_def = to_onnx(clr, X_train.astype(numpy.float32),
target_opset=to)
oinf = InferenceSession(model_def.SerializeToString())
X_test = X_test[:3]
y = oinf.run(None, {'X': X_test.astype(numpy.float32)})
dist, ind = clr.kneighbors(X_test)
assert_almost_equal(dist, DataFrame(y[1]).values, decimal=5)
assert_almost_equal(ind, y[0])
def test_sub_estimator(self):
data = load_iris()
X, y = data.data, data.target
X_train, X_test, y_train, y_test = train_test_split(X, y)
model = ValidatorClassifier()
model.fit(X_train, y_train)
update_registered_converter(ValidatorClassifier,
'CustomValidatorClassifier',
validator_classifier_shape_calculator,
validator_classifier_converter,
parser=validator_classifier_parser)
X32 = X_test[:5].astype(np.float32)
model_onnx = to_onnx(model, X32, target_opset=TARGET_OPSET)
sess = InferenceSession(model_onnx.SerializeToString())
res = sess.run(None, {'X': X32})
assert_almost_equal(model.predict(X32), res[0])
assert_almost_equal(model.predict_proba(X32), res[1], decimal=4)
assert_almost_equal(model.validate(X32), res[2])
def test_gradient_boosting_regressor_learning_rate(self):
X, y = make_classification(n_features=100,
n_samples=1000,
n_classes=2,
n_informative=8)
X_train, X_test, y_train, _ = train_test_split(X,
y,
test_size=0.5,
random_state=42)
model = GradientBoostingClassifier().fit(X_train, y_train)
onnx_model = convert_sklearn(
model, 'lr2', [('input', FloatTensorType(X_test.shape))])
sess = InferenceSession(onnx_model.SerializeToString())
res = sess.run(None, input_feed={'input': X_test.astype(np.float32)})
r1 = np.mean(
np.isclose(model.predict_proba(X_test),
list(map(lambda x: list(map(lambda y: x[y], x)),
res[1])),
atol=1e-4))
r2 = np.mean(res[0] == model.predict(X_test))
assert r1 == r2
def test_model_bayesian_ridge_return_std_normalize_double(self):
model, X = fit_regression_model(
linear_model.BayesianRidge(normalize=True),
n_features=2,
n_samples=50)
model_onnx = convert_sklearn(
model,
"bayesian ridge",
[("input", DoubleTensorType([None, X.shape[1]]))],
options={linear_model.BayesianRidge: {
'return_std': True
}},
target_opset=TARGET_OPSET)
self.assertIsNotNone(model_onnx)
X = X.astype(numpy.float64)
sess = InferenceSession(model_onnx.SerializeToString())
outputs = sess.run(None, {'input': X})
pred, std = model.predict(X, return_std=True)
assert_almost_equal(pred, outputs[0].ravel())
assert_almost_equal(std, outputs[1].ravel(), decimal=4)
def common_test_gpc(self, dtype=np.float32, n_classes=2):
gp = GaussianProcessClassifier()
gp, X = self.fit_classification_model(gp, n_classes=n_classes)
# return_cov=False, return_std=False
if dtype == np.float32:
cls = FloatTensorType
else:
cls = DoubleTensorType
model_onnx = to_onnx(
gp, initial_types=[('X', cls([None, None]))],
target_opset=TARGET_OPSET,
options={GaussianProcessClassifier: {
'zipmap': False, 'optim': 'cdist'}})
self.assertTrue(model_onnx is not None)
try:
sess = InferenceSession(model_onnx.SerializeToString())
except OrtFail:
if not hasattr(self, 'path'):
return
suffix = 'Double' if dtype == np.float64 else 'Float'
# Operator Solve is missing
model_onnx = change_onnx_domain(
model_onnx, {'Solve': ('Solve%s' % suffix, 'ai.onnx.contrib')})
so = SessionOptions()
so.register_custom_ops_library(self.path)
sess = InferenceSession(model_onnx.SerializeToString(), so)
res = sess.run(None, {'X': X.astype(dtype)})
assert_almost_equal(res[0].ravel(), gp.predict(X).ravel())
assert_almost_equal(res[1], gp.predict_proba(X),
decimal=3)
return
dt = 32 if dtype == np.float32 else 64
dump_data_and_model(
X.astype(dtype), gp, model_onnx, verbose=False,
basename="SklearnGaussianProcessRBFT%d%d" % (n_classes, dt))
def test_sub_div(self):
class CustomOpTransformer(BaseEstimator, TransformerMixin):
def __init__(self):
pass
def fit(self, X, y=None):
self.W = np.mean(X, axis=0)
self.S = np.std(X, axis=0)
return self
def transform(self, X):
return (X - self.W) / self.S
mat = np.array([[0., 1.], [0., 1.], [2., 2.]])
tr = CustomOpTransformer()
tr.fit(mat)
z = tr.transform(mat)
def conv(scope, operator, container):
W = operator.raw_operator.W
S = operator.raw_operator.S
X = operator.inputs[0]
out = operator.outputs
op = OnnxDiv(OnnxSub(X, W), S, output_names=out)
op.add_to(scope, container)
def shape(operator):
N = operator.inputs[0].type.shape[0]
W = operator.raw_operator.W
operator.outputs[0].type.shape = [N, W.shape[0]]
model_onnx = convert_sklearn(
tr,
'a-sub-div', [('input', FloatTensorType([1, 2]))],
custom_shape_calculators={CustomOpTransformer: shape},
custom_conversion_functions={CustomOpTransformer: conv})
sess = InferenceSession(model_onnx.SerializeToString())
z2 = sess.run(None, {'input': mat.astype(np.float32)})[0]
assert_almost_equal(z, z2)
def test_pipeline_voting_tfidf_svc(self):
pipe1 = Pipeline([('tfidf1', TfidfVectorizer()),
('svc', SVC(probability=True, kernel='linear'))])
pipe2 = Pipeline([('tfidf2', TfidfVectorizer(norm='l2',
use_idf=False)),
('sgd',
SGDClassifier(alpha=0.0001,
penalty='l2',
loss='modified_huber'))])
pipe3 = Pipeline([('tfidf3', TfidfVectorizer()),
('mnb', MultinomialNB())])
voting = VotingClassifier([('p1', pipe1), ('p2', pipe2),
('p3', pipe3)],
voting='soft',
flatten_transform=False)
data = numpy.array([
"first sentance", "second sentence", "many sentances",
"dummy sentance", "no sentance at all"
])
y = numpy.array([0, 0, 1, 0, 1])
voting.fit(data, y)
expected_label = voting.predict(data)
expected_proba = voting.predict_proba(data)
df = pandas.DataFrame(data)
df.columns = ['text']
model_onnx = convert_sklearn(voting,
initial_types=[
('text', StringTensorType([None, 1]))
],
target_opset=TARGET_OPSET,
options={id(voting): {
'zipmap': False
}})
# with open("debug.onnx", "wb") as f:
# f.write(model_onnx.SerializeToString())
sess = InferenceSession(model_onnx.SerializeToString())
got = sess.run(None, {'text': data.reshape((-1, 1))})
assert_almost_equal(expected_proba, got[1], decimal=5)
assert_almost_equal(expected_label, got[0])
class Predict:
def __init__(self, model_path, bert_path):
self.processor = TextProcessor()
self.sess_options = SessionOptions()
# self.sess_options.intra_op_num_threads = 1
self.sess_options.intra_op_num_threads = psutil.cpu_count(logical=True)
# self.sess_options.graph_optimization_level = GraphOptimizationLevel.ORT_ENABLE_ALL
self.session = InferenceSession(model_path, self.sess_options)
self.use_gpu = torch.cuda.is_available()
self.device = torch.device("cuda:7" if use_gpu else "cpu")
self.tokenizer = BertTokenizer.from_pretrained(self.bert_path)
def to_numpy(self, tensor):
return tensor.detach().cpu().numpy(
) if tensor.requires_grad else tensor.cpu().numpy()
def run(self, record):
text_a, text_b = record[0], record[1]
example = self.processor._create_single_example(text_a, text_b)
feature = convert_single_example(example, self.max_seq_length,
self.tokenizer)
input_ids = torch.tensor(feature.input_ids,
dtype=torch.long).unsqueeze(0)
input_mask = torch.tensor(feature.input_mask,
dtype=torch.long).unsqueeze(0)
segment_ids = torch.tensor(feature.segment_ids,
dtype=torch.long).unsqueeze(0)
ort_inputs = {
'input_ids': input_ids,
'input_mask': input_mask,
'segment_ids': segment_ids
}
ort_outputs = self.session.run(None, ort_inputs)
print(ort_outputs)
print(type(ort_outputs))
def infer(self, data_path):
pass
def test_onnx_example_algebra(self):
initial = np.array([0, 0]).astype(np.float32).reshape((2, ))
x = np.array([1, 2, 3, 4, 5, 6]).astype(np.float32).reshape((3, 2))
opv = _TARGET_OPSET_
add_node = OnnxAdd('sum_in',
'next',
output_names=['sum_out'],
op_version=opv)
id_node = OnnxIdentity(add_node,
output_names=['scan_out'],
op_version=opv)
scan_body = id_node.to_onnx({
'sum_in': initial,
'next': initial
},
outputs=[('sum_out', FloatTensorType()),
('scan_out', FloatTensorType())])
node = OnnxScan('initial',
'x',
output_names=['y', 'z'],
num_scan_inputs=1,
body=scan_body.graph,
op_version=opv)
model_def = node.to_onnx({
'initial': initial,
'x': x
},
outputs=[('y', FloatTensorType()),
('z', FloatTensorType())])
sess = InferenceSession(model_def.SerializeToString())
res = sess.run(None, {'initial': initial, 'x': x})
y = np.array([9, 12]).astype(np.float32).reshape((2, ))
z = np.array([1, 2, 4, 6, 9, 12]).astype(np.float32).reshape((3, 2))
assert_almost_equal(y, res[0])
assert_almost_equal(z, res[1])
def test_model_classifier_multi_class_string_zipmap_columns(self):
model, X = fit_classification_model(linear_model.LogisticRegression(),
3,
n_features=4,
label_string=False)
model_onnx = convert_sklearn(
model,
"multi-class ridge classifier",
[("input", FloatTensorType([None, X.shape[1]]))],
options={linear_model.LogisticRegression: {
'zipmap': 'columns'
}},
target_opset=TARGET_OPSET)
self.assertIsNotNone(model_onnx)
sess = InferenceSession(model_onnx.SerializeToString())
names = [_.name for _ in sess.get_outputs()]
self.assertEqual(['output_label', 'i0', 'i1', 'i2'], names)
xt = X[:10].astype(np.float32)
got = sess.run(None, {'input': xt})
prob = model.predict_proba(xt)
for i in range(prob.shape[1]):
assert_almost_equal(prob[:, i], got[i + 1])