def attribute_value_test(attribute_value):
node = make_node('Abs', ['X'], [], name='test_node', test_attribute=attribute_value)
model = make_model(make_graph([node], 'test_graph', [
make_tensor_value_info('X', onnx.TensorProto.FLOAT, [1, 2]),
], []), producer_name='ngraph')
wrapped_attribute = ModelWrapper(model).graph.node[0].get_attribute('test_attribute')
return wrapped_attribute.get_value()
def attribute_value_test(attribute_value):
node = make_node("Abs", ["X"], [], name="test_node", test_attribute=attribute_value)
model = make_model(make_graph([node], "test_graph", [
make_tensor_value_info("X", onnx.TensorProto.FLOAT, [1, 2])
], []), producer_name="ngraph")
wrapped_attribute = ModelWrapper(model).graph.node[0].get_attribute('test_attribute')
return wrapped_attribute.get_value()
def test_value_info_wrapper(onnx_model):
wrapped_model = ModelWrapper(onnx_model)
wrapped_value_info = wrapped_model.graph.input[0]
assert wrapped_value_info.get_dtype() == np.float32
assert wrapped_value_info.has_initializer
initializer = wrapped_value_info.get_initializer()
assert np.all(initializer.to_array() == np.array([[1., 1.]], dtype=np.float32))
axes = wrapped_value_info.get_ng_axes()
assert len(axes) == 2
assert axes[1].length == 2
placeholder = wrapped_value_info.get_ng_placeholder()
assert placeholder.__class__ == ng.op_graph.op_graph.AssignableTensorOp
assert placeholder.is_placeholder
assert placeholder.axes == axes
variable = wrapped_value_info.get_ng_variable()
assert variable.__class__ == ng.op_graph.op_graph.AssignableTensorOp
assert variable.is_trainable
assert variable.axes == axes
constant = wrapped_value_info.get_ng_constant()
assert constant.__class__ == ng.op_graph.op_graph.AssignableTensorOp
assert constant.is_constant
assert constant.axes == axes
ng_node = wrapped_value_info.get_ng_node()
assert ng_node == constant
def import_onnx_model(onnx_protobuf): # type: (ModelProto) -> List[Dict]
"""
Import an ONNX Protocol Buffers model (onnx_pb2.ModelProto) object
and convert it into a list of ngraph operations.
An ONNX model defines a set of output nodes. Each output node will be added to the
returned list as a dict with the following fields:
* 'name' - name of the output, as specified in the imported ONNX model
* 'inputs' - a list of ngraph placeholder ops, used to feed data into the model
* 'output' - ngraph Op representing the output of the model
Usage example:
>>> onnx_protobuf = onnx.load('y_equals_a_plus_b.onnx.pb')
>>> import_onnx_model(onnx_protobuf)
[{
'name': 'Y',
'inputs': [<AssignableTensorOp(placeholder):4552991464>,
<AssignableTensorOp(placeholder):4510192360>],
'output': <Add(Add_0):4552894504>
}]
>>> ng_model = import_onnx_model(model)[0]
>>> transformer = ng.transformers.make_transformer()
>>> computation = transformer.computation(ng_model['output'], *ng_model['inputs'])
>>> computation(4, 6)
array([ 10.], dtype=float32)
"""
model = ModelWrapper(onnx_protobuf)
return model.graph.get_ng_model()
def test_node_wrapper(onnx_model):
wrapped_model = ModelWrapper(onnx_model)
wrapped_node = wrapped_model.graph.node[0]
ng_inputs = wrapped_node.get_ng_inputs()
assert len(ng_inputs) == 2
assert ng_inputs[0].__class__ == ng.op_graph.op_graph.AssignableTensorOp
ng_outputs = wrapped_node.get_ng_nodes_dict()
assert len(ng_outputs) == 1
assert ng_outputs['Z'].__class__ == ng.op_graph.op_graph.Add
def test_graph_wrapper(onnx_model):
wrapped_model = ModelWrapper(onnx_model)
wrapped_graph = wrapped_model.graph
assert len(wrapped_graph.node) == 1
assert wrapped_graph.node[0].__class__ == NodeWrapper
assert len(wrapped_graph.input) == 2
assert wrapped_graph.input[0].__class__ == ValueInfoWrapper
assert len(wrapped_graph.output) == 1
assert wrapped_graph.output[0].__class__ == ValueInfoWrapper
assert len(wrapped_graph.initializer) == 1
assert wrapped_graph.initializer[0].__class__ == TensorWrapper
initializer = wrapped_graph.get_initializer('X')
assert np.all(initializer.to_array() == np.array([[1., 1.]], dtype=np.float32))
assert not wrapped_graph.get_initializer('Y')
ng_model = wrapped_graph.get_ng_model()[0]
assert ng_model['output'].__class__ == ng.op_graph.op_graph.Add
assert ng_model['inputs'][0].__class__ == ng.op_graph.op_graph.AssignableTensorOp
def test_model_wrapper(onnx_model):
wrapped_model = ModelWrapper(onnx_model)
assert wrapped_model.producer_name == 'ngraph ONNXImporter'
assert wrapped_model.graph.__class__ == GraphWrapper