def normalize_outputs(node: Node):
if node.out_port(0).disconnected():
output = Result(
node.graph, {
'name': node.name + '/Result_port_0/',
'keep_output_port':
node.has_and_set('remove_values_output')
}).create_node()
node.out_port(0).get_connection().set_destination(
output.in_port(0))
# we check port existing to support MaxPool_1 with only 1 output port and MaxPool_8 with 2 output ports
if node.has_port('out', 1) and node.out_port(1).disconnected():
output = Result(
node.graph, {
'name': node.name + '/Result_port_1/',
'keep_output_port':
node.has_and_set('remove_values_output')
}).create_node()
node.out_port(1).get_connection().set_destination(
output.in_port(0))
def merge_nodes(graph: Graph,
nodes_to_merge_names: list,
inputs_desc: list = None,
outputs_desc: list = None):
"""
Merges nodes specified in the set 'nodes_to_merge_names' into one mega-node, creating new edges between mega-node
and inputs/outputs nodes of the mega-node. The added edges contain name of input/output nodes which will be used for
generation of placeholders and will be saved to the IR xml so IE plug-in know how to map input/output data for the
layer. Also the function adds protobufs of the nodes of the sub-graph and 'Const' ops consumed by nodes in the
sub-graph to the node's attribute 'pbs'.
:param graph: the graph object to operate on.
:param nodes_to_merge_names: list of nodes names that should be merged into a single node.
:param inputs_desc: optional list describing input nodes order.
:param outputs_desc: optional list describing output nodes order.
"""
if not is_connected_component(graph, nodes_to_merge_names):
log.warning(
"The following nodes do not form connected sub-graph: {}".format(
nodes_to_merge_names))
# graph.dump_graph_for_graphviz(nodes_to_dump=nodes_to_merge_names)
new_node_name = graph.unique_id("TFSubgraphCall_")
log.info("Create new node with name '{}' for nodes '{}'".format(
new_node_name, ', '.join(nodes_to_merge_names)))
graph.add_node(new_node_name)
new_node_attrs = graph.node[new_node_name]
new_node_attrs['name'] = new_node_name
set_tf_custom_call_node_attrs(new_node_attrs)
new_node = Node(graph, new_node_name)
added_input_tensors_names = set(
) # set of tensors that are were added as input to the sub-graph
added_new_node_output_tensors = dict(
) # key - tensor name, value - out port
for node_name in nodes_to_merge_names:
node = Node(graph, node_name)
add_node_pb_if_not_yet_added(node, new_node)
# TODO: any improvements?
for in_node_name, edge_attrs in Node(graph, node_name).get_inputs():
in_node = Node(graph, in_node_name)
# internal edges between nodes of the sub-graph
if in_node_name in nodes_to_merge_names:
add_node_pb_if_not_yet_added(in_node, new_node)
continue
# edge outside of sub-graph into sub-graph
if in_node_name not in nodes_to_merge_names:
# we cannot use the 'in_node_name' as a protobuf operation name here
# because the 'in_node_name' could be a sub-graph matched before.
input_tensor_name = node.pb.input[edge_attrs['in']]
if input_tensor_name not in added_input_tensors_names:
if not new_node.has_port('in', edge_attrs['in']):
new_node.add_input_port(edge_attrs['in'])
graph.add_edge(
in_node_name, new_node_name,
**merge_edge_props(
{
'in':
find_input_port(new_node, inputs_desc,
node_name, edge_attrs['in']),
'out':
edge_attrs['out'],
'internal_input_node_name':
input_tensor_name,
'original_dst_node_name':
node_name,
'original_dst_port':
edge_attrs['in'],
'in_attrs': [
'in', 'internal_input_node_name',
'original_dst_node_name',
'original_dst_port', 'placeholder_name'
],
'out_attrs': ['out']
}, edge_attrs))
log.debug(
"Creating edge from outside of sub-graph to inside sub-graph: {} -> {}"
.format(in_node_name, new_node_name))
added_input_tensors_names.add(input_tensor_name)
# edge from inside sub-graph to outside sub-graph
for out_node_name, edge_attrs in Node(graph, node_name).get_outputs():
if out_node_name not in nodes_to_merge_names:
log.debug(
"Creating edge from inside of sub-graph to outside sub-graph: {} -> {}"
.format(new_node_name, out_node_name))
out_name = internal_output_name_for_node(
node_name, edge_attrs['out'])
if out_name not in added_new_node_output_tensors.keys():
added_new_node_output_tensors[out_name] = find_output_port(
new_node, outputs_desc, node_name, edge_attrs['out'])
if not new_node.has_port(
'out', added_new_node_output_tensors[out_name]):
new_node.add_output_port(
added_new_node_output_tensors[out_name])
graph.add_edge(
new_node_name, out_node_name,
**merge_edge_props(
{
'in': edge_attrs['in'],
'out': added_new_node_output_tensors[out_name],
'internal_output_node_name': out_name,
'in_attrs': ['in', 'internal_input_node_name'],
'out_attrs': ['out', 'internal_output_node_name']
}, edge_attrs))
new_node['output_tensors_names'] = [
val for val in
{v: k
for k, v in added_new_node_output_tensors.items()}.values()
]
# add nodes using the same order as in initial GraphDef so we can dump them to IR in "correct" order
new_node['nodes_order'] = [
node for node in graph.graph['initial_nodes_order']
if node in new_node['pbs'].keys()
]
for n in nodes_to_merge_names:
if graph.has_node(
n): # check if not deleted by another (similar) pattern
graph.remove_node(n)
return Node(graph, new_node_name)
def infer(node: Node):
num_of_inputs = len(node.in_ports())
opset = node.get_opset()
max_num_of_inputs = 6 if opset == 'opset5' else 5
input_msg_fmt = 'NonMaxSuppression node {} from {} must have from 2 to {} inputs'
node_name = node.soft_get('name', node.id)
inputs_msg = input_msg_fmt.format(node_name, opset, max_num_of_inputs)
assert 2 <= num_of_inputs <= max_num_of_inputs, inputs_msg
boxes_shape = node.in_port(0).data.get_shape()
assert boxes_shape is not None, 'The shape of tensor with boxes is not defined'
scores_shape = node.in_port(1).data.get_shape()
assert scores_shape is not None, 'The shape of tensor with scores is not defined'
assert len(boxes_shape
) == 3, 'Length of tensors with boxes must be equal to 3'
assert len(scores_shape
) == 3, 'Length of tensors with scores must be equal to 3'
# According to the specification of the operation NonMaxSuppression,
# the input 'max_output_boxes_per_class' (port 2) is optional, with default value 0.
if num_of_inputs >= 3:
max_output_boxes_per_class = node.in_port(2).data.get_value()
else:
max_output_boxes_per_class = 0
if not max_output_boxes_per_class:
log.info(
'Set default "max_output_boxes_per_class" for node {} to number of boxes'
.format(node.name))
max_output_boxes_per_class = boxes_shape[1]
# convert the np.array value to a scalar to avoid issue with ragged numpy array generation in the shape
# calculation formulas below
if isinstance(max_output_boxes_per_class, np.ndarray):
max_output_boxes_per_class = max_output_boxes_per_class.item()
num_classes = scores_shape[1]
num_input_boxes = boxes_shape[1]
assert scores_shape[2] is dynamic_dimension or scores_shape[2] == num_input_boxes or scores_shape[2] is None \
or num_input_boxes is None, 'Number of boxes mismatch for operation {}'.format(node_name)
if node.get_opset() in ['opset4', 'opset5']:
max_number_of_boxes = min(
num_input_boxes,
max_output_boxes_per_class) * boxes_shape[0] * num_classes
else:
max_number_of_boxes = min(
num_input_boxes,
boxes_shape[0] * max_output_boxes_per_class * num_classes)
node.out_port(0).data.set_shape(shape_array([max_number_of_boxes, 3]))
if opset == 'opset5':
node.out_port(0).data.set_shape(
shape_array([dynamic_dimension_value, 3]))
num_of_outputs = len([
port for port in node.out_ports().values()
if not port.disconnected()
])
if num_of_outputs >= 2 and node.has_port('out', 1):
node.out_port(1).data.set_shape(
shape_array([dynamic_dimension_value, 3]))
if num_of_outputs >= 3 and node.has_port('out', 2):
node.out_port(2).data.set_shape(shape_array([1]))