Problems with namespaces

[Hakan-Eyilmez]

Hello,

I am currently writing a custom adapter and have encountered an issue with namespaces. In multiple instances, the model ignores the specified namespace. Below is a simple example where "Central Park" is displayed outside its namespace. Is this behavior intentional? If so, would it be possible to include a feature to toggle this on or off?

[pkgoogle]

Hi @Hakan-Eyilmez, can you please provide your custom adapter and example model/graph so that we may reproduce this locally? Thanks.

[Hakan-Eyilmez]

Hello @pkgoogle, I am sorry for the inconvenience. Below I provide the code to an example adapter, which was based on your "my_adapter" adapter, with a few additions, and the graph, it produces.

[pkgoogle]

Hi @Hakan-Eyilmez, thanks for the extra information, it always helps to add as much context as possible. The problem appears to be with single nodes within a namespace. I can replicate your issue with your exact adapter, but I noticed "SanFrancisco" is a layer/namespace not a node. I am able to accomplish something similar to want you want if I add an extra node within the "Cities/NewYork" layer/namespace if I incorporate it as such... example (notice I don't have a "NewYork" node):

from typing import Dict

from model_explorer import (Adapter, AdapterMetadata, ModelExplorerGraphs,
                            graph_builder)


class MyAdapter(Adapter):
  """A simple adapter that returns a hard-coded graph.

  See more info at:
  https://github.com/google-ai-edge/model-explorer/wiki/6.-Develop-Adapter-Extension
  """

  metadata = AdapterMetadata(id='my-adapter',
                             name='My first adapter',
                             description='My first adapter!',
                             source_repo='https://github.com/user/my_adapter',
                             fileExts=['test'])

  # This is required.
  def __init__(self):
    super().__init__()

  def convert(self, model_path: str, settings: Dict) -> ModelExplorerGraphs:

    # Create a graph for my road trip.
    graph = graph_builder.Graph(id='road_trip')

    ###################
    # Add nodes.

    # Create start and end node.
    #
    # They are located at root level hence the empty `namespace` parameter.
    vancouver = graph_builder.GraphNode(
        id='vancouver', label='Vancouver', namespace='')
    la = graph_builder.GraphNode(id='la', label='Los Angeles', namespace='')

    # Create a node for Seattle and put it into a 'coastal drive' layer.
    seattle = graph_builder.GraphNode(
        id='seattle', label='Seatttttle', namespace='CoastalDrive')

    # Create San Franciso as a sublayer of the CoastalDrive layer and add some
    # tourist sites there.
    sf_golden_gate_bridge = graph_builder.GraphNode(
        id='sf_golden_gate_bridge', label='Golden gate bridge',
        namespace='CoastalDrive/SanFrancisco')
    sf_pier_39 = graph_builder.GraphNode(
        id='sf_pier_39', label='PIER 39',
        namespace='CoastalDrive/SanFrancisco')

    # Create another two cities and put them into an 'inland drive' layer.
    salt_lake_city = graph_builder.GraphNode(
        id='salt_lake_city', label='Salt lake city', namespace='InlandDrive')
    las_vegas = graph_builder.GraphNode(
        id='las_vegas', label='Las Vegas', namespace='InlandDrive')
    
    # Custom nodes from Hakan
    # new_york = graph_builder.GraphNode(
    #     id='new_york', label='New York', namespace='Cities')
    central_park = graph_builder.GraphNode(
        id='central_park', label='Central Park', namespace='Cities/NewYork')
    statue_of_liberty = graph_builder.GraphNode(
        id='statue_of_liberty', label='Statue of Liberty', namespace='Cities/NewYork')

    washington = graph_builder.GraphNode(
        id='washington', label='Washington DC', namespace='Cities')

    # Add all the nodes into graph.
    graph.nodes.extend([vancouver, la, seattle, sf_golden_gate_bridge,
                        sf_pier_39, salt_lake_city, las_vegas, central_park, statue_of_liberty, washington])

    ###################
    # Add edges.

    # Connect edges along the cities for coastal drive.
    la.incomingEdges.append(
        graph_builder.IncomingEdge(sourceNodeId='sf_pier_39'))
    sf_pier_39.incomingEdges.append(
        graph_builder.IncomingEdge(sourceNodeId='sf_golden_gate_bridge'))
    sf_golden_gate_bridge.incomingEdges.append(
        graph_builder.IncomingEdge(sourceNodeId='seattle'))
    seattle.incomingEdges.append(
        graph_builder.IncomingEdge(sourceNodeId='vancouver'))

    # Connect edges along the cities for inland drive.
    #
    # LA has two incoming edges from pier_39 and las_vegas.
    # We use targetNodeInputId to identify these two edges. pier_39 goes into
    # input id '0' (default), and las_vegas goes into input id '1'.
    la.incomingEdges.append(graph_builder.IncomingEdge(
        sourceNodeId='las_vegas', targetNodeInputId='1'))
    las_vegas.incomingEdges.append(
        graph_builder.IncomingEdge(sourceNodeId='salt_lake_city'))

    # Custom connections of City nodes
    washington.incomingEdges.append(
        graph_builder.IncomingEdge(sourceNodeId='statue_of_liberty'))
    central_park.incomingEdges.append(
        graph_builder.IncomingEdge(sourceNodeId='vancouver'))
    statue_of_liberty.incomingEdges.append(
        graph_builder.IncomingEdge(sourceNodeId='central_park'))
    
    la.incomingEdges.append(graph_builder.IncomingEdge(
        sourceNodeId='washington', targetNodeInputId='2'))
    

    # Vancouver has two outgoing edges to seattle and salt_lake_city.
    # We use sourceNodeOutputId to identify these two edges. Vancouver's output
    # id '0' (default) goes to seattle, and its output id '1' goes to salt_lake_city.
    salt_lake_city.incomingEdges.append(
        graph_builder.IncomingEdge(sourceNodeId='vancouver',
                                   sourceNodeOutputId='1'))

    #######################
    # Add node attributes.

    temperatures = ['52F', '74F', '55F', '64F', '65F', '62F', '90F', '88F', '82F', '108F']
    for i, node in enumerate(graph.nodes):
      node.attrs.append(graph_builder.KeyValue(
          key="temperature", value=temperatures[i]))

    #######################
    # Add outputs metadata.

    # This is the edge from vancouver to seattle.
    vancouver.outputsMetadata.append(graph_builder.MetadataItem(
        # This identifies which output id the metadata attached to.
        #
        # From the "add edges" section we know that output id "0" connects to
        # seattle.
        id='0',
        attrs=[graph_builder.KeyValue(key='distance', value='230 km'),
               # "__tensor_tag" is a special metadata key whose value will be
               # used as output name in side panel.
               graph_builder.KeyValue(key='__tensor_tag', value='coastal')]))

    # This is the edge from vancouver to salt_lake_city.
    vancouver.outputsMetadata.append(graph_builder.MetadataItem(
        # From the "add edges" section we know that output id "1" connects to
        # salt_lake_city.
        id='1',
        attrs=[graph_builder.KeyValue(key='distance', value='1554 km'),
               graph_builder.KeyValue(key='__tensor_tag', value='inland')]))
    
    vancouver.outputsMetadata.append(graph_builder.MetadataItem(
        # From the "add edges" section we know that output id "2" connects to
        # new_york
        id='2',
        attrs=[graph_builder.KeyValue(key='distance', value='154 km'),
               graph_builder.KeyValue(key='__tensor_tag', value='cities')]))

    # Add other distances
    def add_distance_output_metadata(from_node: graph_builder.GraphNode, distance: str):
      from_node.outputsMetadata.append(graph_builder.MetadataItem(
          id='0',
          attrs=[graph_builder.KeyValue(key='distance', value=distance)]))

    add_distance_output_metadata(salt_lake_city, '677 km')
    add_distance_output_metadata(las_vegas, '439 km')
    add_distance_output_metadata(seattle, '1310 km')
    add_distance_output_metadata(sf_golden_gate_bridge, '10 km')
    add_distance_output_metadata(sf_pier_39, '613 km')
    add_distance_output_metadata(central_park, '1490 km')
    add_distance_output_metadata(washington, '405 km')

    #######################
    # Add inputs metadata.

    la.inputsMetadata.append(graph_builder.MetadataItem(
        id='0', attrs=[graph_builder.KeyValue(key='__tensor_tag', value='coastal')]))
    la.inputsMetadata.append(graph_builder.MetadataItem(
        id='1', attrs=[graph_builder.KeyValue(key='__tensor_tag', value='inland')]))
    la.inputsMetadata.append(graph_builder.MetadataItem(
        id='2', attrs=[graph_builder.KeyValue(key='__tensor_tag', value='cities')]))

    return {'graphs': [graph]}

If I remove the extra node ("statue_of_liberty") It reverts to what you say:

    # Custom nodes from Hakan
    # new_york = graph_builder.GraphNode(
    #     id='new_york', label='New York', namespace='Cities')
    central_park = graph_builder.GraphNode(
        id='central_park', label='Central Park', namespace='Cities/NewYork')
    # statue_of_liberty = graph_builder.GraphNode(
    #     id='statue_of_liberty', label='Statue of Liberty', namespace='Cities/NewYork')
    ....
    # Add all the nodes into graph.
    graph.nodes.extend([vancouver, la, seattle, sf_golden_gate_bridge,
                        sf_pier_39, salt_lake_city, las_vegas, central_park, washington])
    ....
    # Custom connections of City nodes
    washington.incomingEdges.append(
        graph_builder.IncomingEdge(sourceNodeId='central_park'))
    central_park.incomingEdges.append(
        graph_builder.IncomingEdge(sourceNodeId='vancouver'))
    # statue_of_liberty.incomingEdges.append(
    #     graph_builder.IncomingEdge(sourceNodeId='central_park'))
    
    la.incomingEdges.append(graph_builder.IncomingEdge(
        sourceNodeId='washington', targetNodeInputId='2'))

@jinjingforever can you please take a look at the single node issue? Thanks.

[jinjingforever]

Hi @Hakan-Eyilmez sorry for the confusion! Model explorer would automatically remove the group/layer node if it only has one single op-node child. This is mainly for better readability. There is a setting to turn this off:

I will add this to the user guide and the adapter development guide. Thank you!

[jinjingforever]

The setting is here: