Missing Pieces

Due to the compact nature of the Python function definitions and the decorator syntax there are a number of features that are supported by the .ogn file that are not supported directly by the decorator syntax. In some cases the Python API is able to fill in the gaps if you really need this extra functionality, albeit in a much less intuitive way than you get if you implement the .ogn/.py file pairs instead. This is intentional as the goal of AutoNode is rapid node creation so tradeoffs were made accordingly.

Structural Pieces

These definitions are typically found in the .ogn file and describe part of the attribute or node type’s structural characteristics.

Node Type Version Number

Every .ogn file is required to supply a node version number using the “version”: N keyword. With an AutoNode definition there is no version number, partly due to the fact that it is less useful for this type of node type definition. In particular for an AutoNode definition:

  • the node types are not serialized so the version number would not persist

  • there is no mechanism for updating a node from an earlier version number, if one was requested

Node Types Not Implemented In Python

It may seem obvious but is still worth mentioning that all AutoNode definitions create node type definitions for node types that are implemented in Python. The decorator is in Python and the function it decorates is in Python so it is natural that the node type stays in Python. This does not preclude using Python libraries such as Warp to move the Python code to other languages such as Cuda or C++, however that is something the user would have to do.

Generation Of Documentation, Tests, And USD Example Files

One of the side benefits of having a definition in a .ogn file is that the code generator can take that definition and create other useful outputs from that definition. In particular it is typical to have node type documentation automatically generated, which is a restructuredText file containing a nicely formatted page with the node type metadata and all of the attribute information.

Similarly a set of standardized tests can be created that test basic node operation, including reading the node type definition from a .usda file to confirm that things like default values behave correctly. With an AutoNode definition this is less useful to have since the whole point is rapid development and extra elements such as documentation and tests are more suited to more permanent node type definitions, where they might belong to an extension or be indexed in a library for easy discovery.

Attribute Documentation

While the node type documentation can be pulled from the decorated function’s docstring there is no similar location to attach such information to the input or output attributes. As a result the attributes will be undocumented and when users are interacting with them they will not have that information available.

An easy workaround to this shortcoming is to embed the attribute information in the function docstring, much in the same way as you might see in regular coding documentation styles. For example, here’s a function that uses the Google developer documentation style.

import numpy as np
import omni.graph.core as og
import omni.graph.core.types as ot

def autonode_documented(a: ot.point3d, b: ot.point3d) -> ot.vector3d:
    """This node type is an example of one where the intent of the attribute values is not easily apparent
    from their names. For that reason the documentation for what they are is embedded here in the function
    docstring. The function itself is calculating the normalized vector between two points. The attributes
    are documented explicitly here to provide more information to the user.
        a: Point considered to be the origin of the vector
        b: Point considered to be the ending point of the vector
        out_0: Normalized vector from point a to point b
    return (b - a) / np.linalg.norm(b - a)

Attribute Metadata

Like node types, attributes also have useful metadata such as ui_name that is not possible to provide directly. Some metadata values will take on useful defaults while others, such as minimum value, will just be absent. It is an implementation detail about attributes that while their structure is specified in the arguments of a decorated AutoNode function, or in the .ogn file, the actual Attribute object is copied from that structure and contains its own unique metadata.

For that reason, you can work around the missing metadata if you really need to by accessing the attributes directly and modifying the metadata after a node of the new type has been created. Here is a brute force method to get all nodes of the type omni.graph.normalize and set the output attribute’s UI name to “Normalized Vector”.

import omni.graph.core as og
import omni.graph.tools.ogn as ogn
for graph in og.get_all_graphs():
    for node in graph.get_nodes():
        if node.get_type_name() != "omni.graph.normalize":
        node.get_attribute("outputs:out_0").set_metadata(ogn.MetadataKeys.UI_NAME, "Normalized Vector")

Custom Naming Of Attribute Outputs

As you have probably noted from the examples, the output attributes all have numbered names such as out_0, out_1, etc. whereas node types defined through a .ogn file have explicit names. The name of an attribute is intentionally immutable so there is no API that will let you assign a new name to the output as there is for setting the UI name through the metadata. Short of doing something heroic, like setting up an indirection table to remap all of your attribute names from the ones you want to use to the numbered names, there is no way to use a more meaningful name for the output attributes.

Attribute Memory Types

In a .ogn file you can specify an attribute’s memory affinity using the memoryType keyword, either at the node type level or on a per-attribute basis. What this does is to cause the generated code to request the attribute’s value on the appropriate device without need for user intervention. With the AutoNode wrapping around the actual data retrieval there is no access to allow specification for attribute values to be retrieved from anywhere other than the CPU.

Outside of a compute method you are still free to request that the data be retrieved from any arbitrary device but inside the decorated function you are restricted to CPU access only.

Extended Attribute Types

While you can specify extended attribute types as part of an AutoNode decorated function argument list you will be unable to resolve the attribute types to a concrete type at runtime. Therefore they won’t be very useful types as they will never contain real values.

Instead, stick with concrete data types. AutoNode functions are meant to be simple anyway so overloading a function to handle many different types of data is against its purpose. Favor a quick function that adds two integers together over a more complex node type definition that reads the underlying types and dispatches the correct add code to the resolved data type values.

Runtime Pieces

In addition to the structural elements above there are a few runtime features that are similarly not available when using AutoNode node type definitions, since it is just a single function.

Defining Methods Other Than compute()

As you might have guessed from some of the above limitations, there are some features on the node type that rely on access to other functions that might be overridden in a node type implementation. The more common ones of these are initialize(), initialize_type(), and release().

As AutoNode functions are not full fledged class objects it is only possible for them to override a single function, which will always be the compute(). (In fact it actually wraps the function in some helper code that sets up and tears down the data for the function call.)

Some features of these extra functions can be handled in other ways, such as the use of the metadata keyword in the decorator arguments, or the brute force addition of attribute metadata seen above. For anything more complex you are better off creating a full node type implementation with a .ogn/.py file pair.

Accessing Node And Context In compute()

As the decorated function assumes every argument passed in to it is an input attribute there is no room for accessing other types of information. In particular, in a normal compute() override the function will have access to both the node and the context of the current evaluation. (In generated node types the information is further wrapped in a more feature-rich database class.)

Although the node and context are available in the wrapper function they cannot be accessed directly as they have not been passed in to the decorated function. However you can make use of the Python inspect module and some implementation information that the arguments will always be named node and context to look up that information.

import inspect

import omni.graph.core as og
import omni.graph.core.types as ot

def autonode_access_node() -> ot.string:
    """This node type is an example of how to use the inspect module to access the node that is evaluating
    this function as its compute. (The context could be accessed the same way.) The node type takes no
    inputs and returns as output the full path to the node being evaluated. It relies on some implementation
    details not changing but for the most part should be reliable.
    frame = inspect.currentframe().f_back
    return frame.f_locals.get("node").get_prim_path()

Emitting Errors And Warnings In compute()

The database generated when using a .ogn file has some useful utilities, one of which is the ability to log an error or a warning with the node so that the UI can report on execution failures. As the AutoNode function has no database this convenience cannot be used.

Fortunately the logging capability is in the node’s API so all you need to do to manually log messages is to access the node using the inspect module as above and call the log function with the appropriate logging level set. Here is an example that uses both warnings and errors to provide status information to the UI.

import inspect

import omni.graph.core as og
import omni.graph.core.types as ot

def autonode_logging(numerator: ot.double, denominator: ot.double) -> ot.double:
    """This node type is an example of how to use the inspect module to access the node that is evaluating
    this function in order to log error and warning messages. The function divides the first number by the
    second number, providing an error when the denominator is zero and a warning when both the numerator and
    denominators are zero, otherwise returning the first number divided by the second.
    if denominator != 0.0:
        return numerator / denominator

    frame = inspect.currentframe().f_back
    node = frame.f_locals.get("node")
    if numerator == 0.0:
        node.log_compute_message(og.Severity.WARNING, "Indeterminate result from dividing 0 by 0")
        node.log_compute_message(og.Severity.WARNING, f"Infinite result from dividing {numerator} by 0")
    # While you can also log an error by setting severity to og.Severity.ERROR, it is better to use an
    # exception so that the compute function will fail and no further computation will take place.
    # if numerator == 0.0:
    #     raise og.OmniGraphError("Indeterminate result from dividing 0 by 0")
    return 0.0

Creating New Output Bundles In compute()

Normally when you want to output a bundle you would construct one using the generated database’s convenience function for doing that. You can also implement this yourself, however you need to emulate that same convenience function, which requires access to both the node and the context via the inspect module.

import inspect

import omni.graph.core as og
import omni.graph.core.types as ot

def autonode_output_bundles() -> ot.bundle:
    """This node type is an example of how to use the inspect module to access the node and context of evaluation
    in order to construct a bundle for use as an output value.
    frame = inspect.currentframe().f_back
    node = frame.f_locals.get("node")
    context = frame.f_locals.get("context")
    new_bundle = og.BundleContents(context, node, "outputs_out_0", read_only=False, gpu_by_default=False)
    new_bundle.add_attributes([og.Type(og.BaseDataType.INT)], ["fizzbin"])

    return new_bundle

Accessing Per-Node Data

As with the message logging, the generated database provides access to internal state data that is associated with each individual node. This data is controlled by the user. Again since there is no database with an AutoNode node type this feature is not available, though with the help of the inspect module it can be implemented locally with not too much effort:

import inspect

import omni.graph.core as og
import omni.graph.core.types as ot

unique_id = 57
per_node_data = {}

def autonode_per_node_data(increment: ot.int) -> ot.int:
    """This node type is an example of how to use the inspect module to access the node that is evaluating
    this function in order to create per-node personal data. This particular example provides a monotonic ID
    that increases for each node created by maintaining a local per-node dictionary that uses the node's
    internal ID as a key.
    nonlocal unique_id
    nonlocal per_node_data
    frame = inspect.currentframe().f_back
    node_id = frame.f_locals.get("node").node_id()
    if node_id not in per_node_data:
        per_node_data[node_id] = unique_id
        unique_id += increment
    return per_node_data[node_id]

Saving The Node Type Definition

As the node type definition is created at runtime there is no artifact to be referenced to find that node type definition in another session. Although the node type definition may have a particular extension associated with it there is no guarantee that merely loading that extension will recreate that node type definition. Indeed it may be impossible if some of the parameters used in its creation are session-dependent, such as date of creation.

To warn the user that they might be saving node types that cannot be retrieved a warning will be issued when a file is saved that contains an instantiated node of any AutoNode type. And in future when required extensions are saved with the USD files any extension called out in an AutoNode definition will be included.

Concrete Extension Dependencies

Further to the above, it is quite possible that an AutoNode decorated function could have dependencies on any arbitrary extension that is currently loaded. For this reason it’s also possible that another user running the exact same decorated function may not get a functioning node type if they do not have the requisite extensions enabled.

Again, another reason why nodes with these types will not be silently saved as they could cause the graph to become non-functional when read into another session.