nGraph is the OpenVINO™ graph manipulation library, used to represent neural network models in the form of a computational graph. With nGraph Python APIs, you can create, inspect, and modify computational graphs.
nGraph is able to represent a large set of mathematical operations, chosen to cover most operations used in popular neural network models. The ngraph module is the basic namespace used to expose commonly used API methods. You can inspect the ngraph module or call help on any method for usage information.
import ngraph as ng
dir(ng)
help(ng.space_to_batch)You can use nGraph's Python API to create computational models. For example, to create a model that represents the formula y=ReLU(x + b), you can use the following code to create parameters x and b and create the computational graph:
import numpy as np
import ngraph as ng
parameter_x = ng.parameter(shape=[2, 2], dtype=np.float32, name="x")
parameter_b = ng.parameter(shape=[2, 2], dtype=np.float32, name="b")
graph = ng.relu(parameter_x + parameter_b, name="y")nGraph operations are organized into sets, which are released with new versions of OpenVINO. As operation signatures can change between operation sets, it is a good practice to use operations from a specific operation set in your scripts. This ensures that your code is compatible with future versions of OpenVINO.
The example, the code above can be rewritten like this to use only operations from opset 4:
import ngraph.opset4 as ops
parameter_x = ops.parameter(shape=[2, 2], dtype=np.float32, name="x")
parameter_b = ops.parameter(shape=[2, 2], dtype=np.float32, name="b")
graph = ops.relu(parameter_x + parameter_b, name="y")You can combine a graph together with its input parameters into an nGraph function object. An nGraph function can be passed to other parts of OpenVINO to perform analysis and inference.
>>> function = ng.Function(graph, [parameter_x, parameter_b], "TestFunction")
<Function: 'TestFunction' ({2,2})>In order to run inference on an nGraph function, convert it to an Inference Engine network and call its infer method.
>>> from openvino.inference_engine import IECore
>>>
>>> ie_network = ng.function_to_cnn(function)
>>>
>>> ie = IECore()
>>> executable_network = ie.load_network(ie_network, 'CPU')
>>>
>>> output = executable_network.infer({
>>> 'x': np.array([[-1, 0], [1, 2]]),
>>> 'b': np.array([[-1, 0], [0, 1]]),
>>> })
>>> output['y']
array([[-0., 0.],
[ 1., 3.]], dtype=float32)You can load a model from a file using the Inference Engine read_network method:
network = ie_core.read_network(model=model_path, weights=weights_path)
An nGraph function can be extracted from an Inference Engine network using the ngraph.function_from_cnn method.
The following example shows how an ONNX model can be loaded to an nGraph Function:
from openvino.inference_engine import IECore
ie = IECore()
model_path='/path/to/model.onnx'
network = ie.read_network(model=model_path)
import ngraph as ng
function = ng.function_from_cnn(network)You can use the nGraph function's get_ordered_ops method to get a topologically sorted list of its graph Nodes.
>>> function.get_ordered_ops()
[<Parameter: 'b' ({2,2}, float)>,
<Parameter: 'x' ({2,2}, float)>,
<Add: 'Add_2' ({2,2})>,
<Relu: 'y' ({2,2})>,
<Result: 'Result_4' ({2,2})>]Each Node has a unique name property, assigned at creation time. User-provided names can be retrieved using the get_friendly_name method. get_type_name returns the operation type of the node and the shape property returns the shape of the node's output tensor.
>>> for node in function.get_ordered_ops():
>>> print('Node name: {:15} Friendly name: {:10} Op: {:10} {}'.format(
>>> node.name, node.get_friendly_name(), node.get_type_name(), node.shape))
Node name: Parameter_1 Friendly name: b Op: Parameter Shape{2, 2}
Node name: Parameter_0 Friendly name: x Op: Parameter Shape{2, 2}
Node name: Add_2 Friendly name: Add_2 Op: Add Shape{2, 2}
Node name: Relu_3 Friendly name: y Op: Relu Shape{2, 2}
Node name: Result_4 Friendly name: Result_4 Op: Result Shape{2, 2}Node's parents can be retrieved using the inputs() method, which returns Input objects. Each Input is a connection to an upstream Output, which is connected to a Node. The upstream node can be retrieved using input.get_source_output().get_node() methods.
>>> for node_input in node.inputs():
>>> parent = node_input.get_source_output().get_node()
>>> print('{} node has parent {}'.format(node.get_friendly_name(), parent.get_friendly_name()))
Add_2 node has parent x
Add_2 node has parent bBy analogy, Node's children can be retrieved using the outputs() method, which returns a list of Output objects. By contrast to inputs, each output can be connected to multiple downstream nodes.
>>> for node_output in node.outputs():
>>> children = [target.get_node() for target in node_output.get_target_inputs()]
>>> print('{} node has children {}'.format(node.get_friendly_name(), ','.join([c.get_friendly_name() for c in children])))
>>>
Add_2 node has children yEach Input and Output has an associated tensor shape, which describes what data it can accept or produce. Shapes can be dynamic, which means that they are only partially known. You can get information about this shape by calling the get_partial_shape method. If the returned PartialShape object's is_dynamic property returns True, then the shape is known partially. Otherwise, you can convert the shape to a fully known static shape (list of integers).
partial_shape = node_input.get_partial_shape()
if partial_shape.is_dynamic:
print('{} node has a dynamic shape {}'.format(node.get_friendly_name(), partial_shape))
else:
shape = list(partial_shape.to_shape())
print('{} node has a static shape {}'.format(node.get_friendly_name(), shape))Nodes may also have additional attributes. For example, a Softmax node has an axis attribute. Each attribute may be retrieved using a dedicated method, for example, node.get_axis().
>>> node = ng.softmax(data=[[1, 2], [3, 4]], axis=1)
>>> node.get_axis()
1You can also set attribute values using corresponding setter methods, for example:
node.set_axis(0)Currently, you can get all attributes of a node using the _get_attributes method. Please note that this is an internal API method and may change in future versions of OpenVINO.
The following code displays all attributes for all nodes in a function:
for node in function.get_ordered_ops():
attributes = node._get_attributes()
if(attributes):
print('Operation {} of type {} has attributes:'.format(node.get_friendly_name(), node.get_type_name()))
for attr, value in attributes.items():
print(" * {}: {}".format(attr, value))
else:
print('Operation {} of type {} has no attributes.'.format(node.get_friendly_name(), node.get_type_name()))Attributes are properties of nodes in the computational graph, which will be stored when the model is serialized to a file. However, there can be additional properties of nodes, which are only important during the execution of a graph. An example of such a property is affinity, which determines which operation will be executed on which hardware in a heterogeneous environment. You can get and set runtime information by using the get_rt_info method of a node.
NOTE: If you set affinity manually, be careful at the current moment Inference Engine plugins don't support constant (
Constant->Result) and empty (Parameter->Result) networks. Please avoid such subgraphs when you set affinity manually.
rt_info = node.get_rt_info()
rt_info["affinity"] = "test_affinity"Please note that the rt_info API may change in future versions of OpenVINO.
You can retrieve the data tensor associated with a Constant node by calling its get_data method:
if node.get_type_name() == 'Constant':
data = node.get_data()nGraph supports a growing number of transformation passes, which can be applied to a graph. For example, to perform constant folding of applicable subgraphs, you can use the ConstantFolding pass. To do this, create an optimization pass Manager, register the passes you want to execute, and feed an nGraph function to the configured pass manager.
The following code example illustrates the process:
from ngraph.impl.passes import Manager
pass_manager = Manager()
pass_manager.register_pass('ConstantFolding')
pass_manager.run_passes(function)Please note that the list of available transformations may change in future versions of OpenVINO.