def test_default_version_IR_V11_seperate_paths():
core = Core()
xml_path = "./serialized_function.xml"
bin_path = "./serialized_function.bin"
shape = [100, 100, 2]
parameter_a = ov.parameter(shape, dtype=np.float32, name="A")
parameter_b = ov.parameter(shape, dtype=np.float32, name="B")
parameter_c = ov.parameter(shape, dtype=np.float32, name="C")
parameter_d = ov.parameter(shape, dtype=np.float32, name="D")
model = ov.floor(
ov.minimum(ov.abs(parameter_a), ov.multiply(parameter_b, parameter_c)))
func = Model(model, [parameter_a, parameter_b, parameter_c], "Model")
pass_manager = Manager()
pass_manager.register_pass("Serialize",
xml_path=xml_path,
bin_path=bin_path,
version="IR_V11")
pass_manager.run_passes(func)
res_func = core.read_model(model=xml_path, weights=bin_path)
assert func.get_parameters() == res_func.get_parameters()
assert func.get_ordered_ops() == res_func.get_ordered_ops()
os.remove(xml_path)
os.remove(bin_path)
def test_low_latency2():
X = opset8.parameter(Shape([32, 40, 10]), np.float32, "X")
Y = opset8.parameter(Shape([32, 40, 10]), np.float32, "Y")
M = opset8.parameter(Shape([32, 2, 10]), np.float32, "M")
X_i = opset8.parameter(Shape([32, 2, 10]), np.float32, "X_i")
Y_i = opset8.parameter(Shape([32, 2, 10]), np.float32, "Y_i")
M_body = opset8.parameter(Shape([32, 2, 10]), np.float32, "M_body")
sum = opset8.add(X_i, Y_i)
Zo = opset8.multiply(sum, M_body)
body = Model([Zo], [X_i, Y_i, M_body], "body_function")
ti = opset8.tensor_iterator()
ti.set_body(body)
ti.set_sliced_input(X_i, X.output(0), 0, 2, 2, 39, 1)
ti.set_sliced_input(Y_i, Y.output(0), 0, 2, 2, -1, 1)
ti.set_invariant_input(M_body, M.output(0))
out0 = ti.get_iter_value(Zo.output(0), -1)
out1 = ti.get_concatenated_slices(Zo.output(0), 0, 2, 2, 39, 1)
result0 = opset8.result(out0)
result1 = opset8.result(out1)
model = Model([result0, result1], [X, Y, M])
m = Manager()
m.register_pass(LowLatency2())
m.run_passes(model)
# TODO: create TI which will be transformed by LowLatency2
assert count_ops(model, "TensorIterator") == [1]
def test_simple_pattern_replacement():
# Simple: for Extensions. Without any classes and inheritance.
def pattern_replacement():
param = WrapType("opset8.Parameter")
relu = WrapType("opset8.Relu", param.output(0))
def callback(m: Matcher) -> bool:
root = m.get_match_root()
# Just to check that capturing works and we can
# link pattern nodes with matched graph nodes.
assert relu in m.get_pattern_value_map()
new_relu = opset8.exp(
root.input_value(0)) # ot root.input(0).get_source_output()
replace_node(root, new_relu)
return True
return Matcher(relu, "SimpleReplacement"), callback
model = get_test_function()
m = Manager()
m.register_pass(MatcherPass(*pattern_replacement()))
m.run_passes(model)
assert count_ops(model, ("Relu", "Exp")) == [0, 1]
def test_convert_precision():
model = get_model()
m = Manager()
m.register_pass(ConvertFP32ToFP16())
m.run_passes(model)
assert model is not None
def test_constant_folding():
model = get_model()
m = Manager()
m.register_pass(ConstantFolding())
m.run_passes(model)
assert model is not None
assert count_ops(model, "ShapeOf") == [0]
def test_graph_rewrite():
model = get_test_function()
m = Manager()
# check that register pass returns pass instance
anchor = m.register_pass(GraphRewrite())
anchor.add_matcher(PatternReplacement())
m.run_passes(model)
assert count_ops(model, "Relu") == [2]
def test_matcher_pass():
model = get_test_function()
m = Manager()
# check that register pass returns pass instance
p = m.register_pass(PatternReplacement())
m.run_passes(model)
assert p.model_changed
assert count_ops(model, "Relu") == [2]
def test_make_stateful():
model = get_model()
m = Manager()
p = MakeStateful({"parameter": "result"})
m.register_pass(p)
m.run_passes(model)
assert model is not None
assert len(model.get_parameters()) == 0
assert len(model.get_results()) == 0
def test_register_new_node():
class InsertExp(MatcherPass):
def __init__(self):
MatcherPass.__init__(self)
self.model_changed = False
param = WrapType("opset8.Parameter")
def callback(m: Matcher) -> bool:
# Input->...->Result => Input->Exp->...->Result
root = m.get_match_value()
consumers = root.get_target_inputs()
exp = opset8.exp(root)
for consumer in consumers:
consumer.replace_source_output(exp.output(0))
# For testing purpose
self.model_changed = True
# Use new operation for additional matching
self.register_new_node(exp)
# Root node wasn't replaced or changed
return False
self.register_matcher(Matcher(param, "InsertExp"), callback)
class RemoveExp(MatcherPass):
def __init__(self):
MatcherPass.__init__(self)
self.model_changed = False
param = WrapType("opset8.Exp")
def callback(m: Matcher) -> bool:
root = m.get_match_root()
root.output(0).replace(root.input_value(0))
# For testing purpose
self.model_changed = True
return True
self.register_matcher(Matcher(param, "RemoveExp"), callback)
m = Manager()
ins = m.register_pass(InsertExp())
rem = m.register_pass(RemoveExp())
m.run_passes(get_test_function())
assert ins.model_changed
assert rem.model_changed
def test_convert_precision():
model = get_model()
assert model.get_parameters()[0].get_element_type().to_dtype(
) == np.float32
m = Manager()
m.register_pass(ConvertFP32ToFP16())
m.run_passes(model)
assert model is not None
assert model.get_parameters()[0].get_element_type().to_dtype(
) == np.float16
def test_serialize_pass():
core = Core()
xml_path = "serialized_function.xml"
bin_path = "serialized_function.bin"
func = get_test_function()
m = Manager()
m.register_pass(Serialize(xml_path, bin_path))
m.run_passes(func)
assert func is not None
res_func = core.read_model(model=xml_path, weights=bin_path)
assert func.get_parameters() == res_func.get_parameters()
assert func.get_ordered_ops() == res_func.get_ordered_ops()
os.remove(xml_path)
os.remove(bin_path)
def test_constant_folding():
node_constant = ov.constant(
np.array([[0.0, 0.1, -0.1], [-2.5, 2.5, 3.0]], dtype=np.float32))
node_ceil = ov.ceiling(node_constant)
func = Model(node_ceil, [], "TestFunction")
assert count_ops_of_type(func, node_ceil) == 1
assert count_ops_of_type(func, node_constant) == 1
pass_manager = Manager()
pass_manager.register_pass("ConstantFolding")
pass_manager.run_passes(func)
assert count_ops_of_type(func, node_ceil) == 0
assert count_ops_of_type(func, node_constant) == 1
new_const = func.get_results()[0].input(0).get_source_output().get_node()
values_out = new_const.get_vector()
values_expected = [0.0, 1.0, 0.0, -2.0, 3.0, 3.0]
assert np.allclose(values_out, values_expected)
def test_serialize_pass_mixed_args_kwargs():
core = Core()
xml_path = "serialized_function.xml"
bin_path = "serialized_function.bin"
shape = [3, 2]
parameter_a = ov.parameter(shape, dtype=np.float32, name="A")
parameter_b = ov.parameter(shape, dtype=np.float32, name="B")
model = parameter_a - parameter_b
func = Model(model, [parameter_a, parameter_b], "Model")
pass_manager = Manager()
pass_manager.register_pass("Serialize", xml_path, bin_path=bin_path)
pass_manager.run_passes(func)
res_func = core.read_model(model=xml_path, weights=bin_path)
assert func.get_parameters() == res_func.get_parameters()
assert func.get_ordered_ops() == res_func.get_ordered_ops()
os.remove(xml_path)
os.remove(bin_path)
def test_serialize_results():
core = Core()
node_constant = ov.constant(
np.array([[0.0, 0.1, -0.1], [-2.5, 2.5, 3.0]], dtype=np.float32))
node_ceil = ov.ceiling(node_constant)
func = Model(node_ceil, [], "Model")
xml_path = "serialized_function.xml"
bin_path = "serialized_function.bin"
pass_manager = Manager()
pass_manager.register_pass("Serialize",
xml_path=xml_path,
bin_path=bin_path)
pass_manager.run_passes(func)
res_func = core.read_model(model=xml_path, weights=bin_path)
const = func.get_results()[0].input(0).get_source_output().get_node()
new_const = res_func.get_results()[0].input(
0).get_source_output().get_node()
assert const == new_const
os.remove(xml_path)
os.remove(bin_path)
def test_serialize_seperate_paths_args():
core = Core()
xml_path = "serialized_function.xml"
bin_path = "serialized_function.bin"
shape = [2, 2]
parameter_a = ov.parameter(shape, dtype=np.float32, name="A")
parameter_b = ov.parameter(shape, dtype=np.float32, name="B")
parameter_c = ov.parameter(shape, dtype=np.float32, name="C")
parameter_d = ov.parameter(shape, dtype=np.float32, name="D")
model = ((parameter_a + parameter_b) * parameter_c) / parameter_d
func = Model(model, [parameter_a, parameter_b, parameter_c, parameter_d],
"Model")
pass_manager = Manager()
pass_manager.register_pass("Serialize", xml_path, bin_path)
pass_manager.run_passes(func)
res_func = core.read_model(model=xml_path, weights=bin_path)
assert func.get_parameters() == res_func.get_parameters()
assert func.get_ordered_ops() == res_func.get_ordered_ops()
os.remove(xml_path)
os.remove(bin_path)
def dump_exec_graph(compiled_model, model_path):
weight_path = model_path[:model_path.find(".xml")] + ".bin"
pass_manager = Manager()
pass_manager.register_pass("Serialize", model_path, weight_path)
pass_manager.run_passes(compiled_model.get_runtime_model())
def test_model_pass():
m = Manager()
p = m.register_pass(MyModelPass())
m.run_passes(get_test_function())
assert p.model_changed
.convert_element_type() \
.convert_color(ColorFormat.RGB) \
.resize(ResizeAlgorithm.RESIZE_LINEAR) \
.mean([123.675, 116.28, 103.53]) \
.scale([58.624, 57.12, 57.375])
# Dump preprocessor
print(f'Dump preprocessor: {ppp}')
model = ppp.build()
# ======== Step 2: Change batch size ================
# In this example we also want to change batch size to increase throughput
set_batch(model, 2)
# ======== Step 3: Save the model ================
pass_manager = Manager()
pass_manager.register_pass(pass_name="Serialize",
xml_path='/path/to/some_model_saved.xml',
bin_path='/path/to/some_model_saved.bin')
pass_manager.run_passes(model)
# ! [ov:preprocess:save]
# ! [ov:preprocess:save_load]
core = Core()
core.set_property({'CACHE_DIR': '/path/to/cache/dir'})
# In case that no preprocessing is needed anymore, we can load model on target device directly
# With cached model available, it will also save some time on reading original model
compiled_model = core.compile_model('/path/to/some_model_saved.xml', 'CPU')
# ! [ov:preprocess:save_load]
