Composer

The Composer is an GUI application designed to create AI Application pipeline through an easy-to-use graphic interface, reducing the complexity of application development significantly and thus reducing time to market.

The Composer is based on the Omniverse SDK which provides a highly responsive hardware-accelerated GUI.

User Interface

The user interface contains the following windows:

User Interface	Description
Graph Editor Window	Used to edit/create graphs, multiple windows can be created.
Node Template Window	Used to create nodes from saved templates.
Component List Window	Displays list of available components from Registry.
Details Window	Shows the details on any selected object.
Console Window	Display the console output
Tool Bar	Tool buttons for miscellaneous features

Menu Bar

The menu bar on the top of the main window contains the main menus of the Composer, providing entries to execute common commands.

File Menu

Menu Items	Action	Hot key
New Graph	Open an empty Graph Editor Window	Ctrl + N
Open Graph	Open an existing graph file in an new Graph Editor Window	Ctrl + O
Save Graph	Save the current graph	Ctrl + S
Save Graph As	Save the current graph with a different name and location	Shift + Ctrl + S
Close Graph	Close the current graph	Ctrl + W
Exit	Quit the application	Ctrl + Q

Edit Menu

Menu Items	Action	Hot key
Delete	Remove the selected instance	Del
Copy	Copy the selected instance to the clipboard	Ctrl + C
Cut	Copy the selected instance to the clipboard and remove it from the graph	Ctrl + X
Paste	Paste the instance from the clipboard to the graph	Ctrl + V
Undo	Undo a command	Ctrl + Z
Redo	Redo a command	Ctrl + Y

Registry Menu

Menu Items	Action
Add NGC Private Repo	Add NGC Private Repository
Sync Repo	Sync Repository
Remove Repo	Remove Repository
Clear Default Repo	Clear Default Repository
Set Cache Path	Set Registry Cache Path
Clear Cache	Clear Registry Cache

View Menu

Menu Items	Action
Debug View	Toggle the Debug View

Tools Menu

Menu Items	Action
Generate Extension	Open Generate Extension for Gstreamer Elements window
Container Builder	Launch the Container Builder

Here is the link to Generate Extension page:

ref

Generate Gstreamer Extension

Help Menu

Menu Items	Action
About	Display versions of the tools

Window Menu

Menu Items	Action
Property	Toggle the Details Window
Console	Toggle the Console Window

Context Menu

Context menu is invoked while a user right-clicks on the canvas with a mouse, which provides the entries to commands related to the current graph.

Menu Items	Action
Rename the Graph	Rename the current graph
Export Parameters	Export the flagged parameters in the current graph to a file
Load Parameters	Load the value of parameters from a parameter file
Open All	Expand all the nodes in the graph with all the details
Minimize All	Minimize all the nodes in the graph by showing connected ports only
Close All	Close all the nodes in the graph by hiding all the ports
Layout All	Automatically layout all the nodes
Focus on All	Automatically zoom and pan the canvas to fit the entire graph
Focus on the Selection	Automatically zoom and pan the canvas to focus on the current selection

Toolbar

Toolbar on the top of the main window provides shortcuts of available functions of the current graph

Toolbar Items	Action
Expand Nodes	Expand all the nodes in the graph with all the details
Minimize Nodes	Minimize all the nodes in the graph by showing connected ports only
Close Nodes	Close all the nodes in the graph by hiding all the ports
Container Builder	Build container for x86 or aarch64
Run Graph	Run current graph on local or remote machine
Stop Graph	Stops the execution of the currently running graph

Component List

The component list window shows the information of all the available components in various extensions, which are fetched from the remote registry and could be used to build different graphs by the user. By default, the list is showed in a two-level hierarchical view grouped by extensions.

You can choose a component from the Component List by a left click on the mouse to display the detailed information of that component in the Details window. The details include the component type, description, details of the extension which includes the component, and the details of the properties supported by the component. The Component List supports direct Drag and Drop, allowing users to create an instance of a specific component in a graph by simply dragging it from the list and dropping to the canvas or a destination node in the graph. If you drops it to the canvas, a default node will be created automatically to accommodate the instance and its input and output port will be initialized as much as possible.

Create New Application

1. Navigate to the File menu from the main window menu bar and choose New Graph.

2. An empty graph window is created and displayed in the graph editor workspace.

3. You may create multiple graph editor windows to work on different graphs simultaneously.

Open and Save Application Graphs

Note

Be sure the component list is not empty before opening any graph, trying opening a graph file with an empty component list will generate a corrupted graph.

1. To save the current graph, use the ‘File’ menu and choose Save the Graph (Ctrl + S) or Save the Graph as (Shift + Ctrl +S), if the graph has never been saved before or you want to save it to a different file.

2. A file browser dialog pops up prompting you to pick a path in the file system.

Compose an Application Graph

Finding the right component

Components are grouped by extensions. Information of a specific component and the extension can be found in the details window if you click on the corresponding item showed in the Component List, including its type, the purpose, and supported properties. With the search bar on the top of the Component List window, you can also filter out the candidate components.

By choosing the versions from the Version list, the extension with a specific version will be loaded.

From Composer perspective, a Component falls into one of the 3 categories below based on the role it plays in the graph:

Component Type	Description	Example
Primary Component	Compute Components which are executed by scheduler.	Decoder
Secondary Component	Used by primary components for data exchange, config, scheduling rules.	SinkPad
Connection Component	Responsible to connect input and output components	Connection
Subgraph Component	Used to load a subgraph	Subgraph

Before creating an AI Application graph, a user should always figure out what kind of primary components are needed. Only after all the primary components have been properly picked, a user could consider initializing them and then setting up connections.

The component within an entity has two parts:

• The header - Displays the name of the component, which is its type by default. Double-click on the header to change its name.

• One or more handles - They are displayed under the header part if the component can link to the other components. Number of supported handles, the type and names of those handles are all predefined by the component type.

Creating a Component Instance

Steps to create component instances:

Using Drag and Drop

1. To add a new component instance to a graph, find the type of the component from Component window and drag and drop it to the graph editor window.

2. If the component is dropped to the vacant area an entity will be added automatically to contain the component

3. If the component is dropped to an existing entity it is going to be added to that entity.

Using New empty node

You can also choose to first create an empty node by right-clicking on the canvas and choosing New Empty Node, then drag and drop a component to the node. You must manually initialize all the handles of the instance if it is a primary component.

Using Context Menu

Select Create from the context menu of a handle of a primary component, which only works for creating or using secondary components. Since the compatible components have already been filtered for the specific handle, the list becomes much shorter and it is always recommended to create secondary components including Input and output components in this way.

Understanding the Component Handles

A component can have handles which allows it to interact with other components. Essentially a handle is a special property supported by the component, which is associated with a specific component type. To make a handle functional, you’ll need to initialize it first, based on its type.

There are some special handles called Input and Output handles who manage the data exchange of a component, and to enable data exchange of a component a user needs to initialize them with appropriate Input and Output components.

Linking and Unlinking components

There are two options for a user to initialize a Handle: one is through its context menu, which is mentioned above; the other one is using Link. The Link operation can be done by dragging and dropping a curve from the handle to its required component, which means the required component must be created at first. As a reverse operation, Unlink operation turns a handle back to uninitialized state, which can be invoked through the context menu of the handle.

Note

One handle can only be initialized with (or linked to) one component instance at the same time

Steps to link or unlink the components:

1. To link two components, left-click with your cursor on an output port (the full circle) in one component and drag a curve to the input port (the half circle) of another component.

2. Links can only be set up among compatible components, which means the type of the link source must be the same type or the base type of a link target.

3. To see the port type, hover the mouse cursor on the port.

4. To remove the link, hover the cursor on the curve and right click the mouse, then choose Unlink from the context menu.

Node

A node is the container of a set of components who work together to get certain tasks done. Usually there is one primary and several secondary components required by the primary one, including the input, output, and others. An empty node can be created through context menu triggered by a right click on the canvas, and an existing node can be deleted from File Menu, the Context Menu of the node or the corresponding hot key.

Nodes can be categories based on the roles and types of the components contained inside, different category of nodes will be showed in different style:

Node Category	Description	Example
compute	General Computing Components	NvDsAnalytics
cloud	Cloud Components	NvDsMsgBroker
input	Input Components	NvDsSingleSrcInput
neural networks	Inference Components with Neural Networks	NvDsInferVideo
output	Output Components	NvDsVideoRenderer
muxer	Muxing/Demuxing Components	NvDsStreamDemux
utility	Utilities	RealTimeClock
image processing	Image Processing Components	NvDsDewarper

Setting up a Connection from an Input to an Output

A connection represents the data flow from an input component to an output one, and it can be created when a user drags a curve from an output and drop it to a compatible input. It is important to understand that the input and the output of a connection must be compatible, that is to say, the type of data from the output to the input must be exchangeable. Connections can only be set up between two different nodes. Multiple connections are allowed on a single input or output. To remove connections on an Input or Output, users can use the “Disconnect” from the context menu by right clicking on the connection.

Changing the Component Properties

To select a component, left or right-click on a component or its handles in the graph editor. This displays the component information including its properties in the detail window accordingly, where you can make changes on any writable property. It displays the property name [1], the text field [2] to edit the property, the check box [3] to export a file property to a resource file and another check box [4] to export the property to a different yaml file.

Editor Features

1. Zoom and Pan the Canvas

   The scroll wheel of the mouse enables zoom in and zoom out, and pressing the scroll wheel allows users to pan the canvas.

2. Copy, Cut and Paste

Before doing a Copy or Cut, first select a instance from a graph editor, then trigger the operation through Edit menu or corresponding hot key. Copy and Cut is supported for both component instances and node instances in the graph editors. Either operation creates a copy of the instance. When the copy of a node is created, all the components inside the node will be created accordingly preserving their relationship. There is a global clipboard in the Composer where a copied instance (a component or a node) is kept temporarily. When a Paste command is triggered the instance in the clipboard will be added back to the graph. To paste a component instance, the user needs to choose the destination node first, otherwise the Paste fails.

3. Undo and Redo

   Undo and Redo are supported for following commands:

   • New Node

   • New Component

   • Link a Handle

   • Unlink a Handle

   • Set up a Connection

   • Disconnect an Input/Output

   • Delete a Node

   • Delete a Component

   • Rename a Node

   • Rename a Component

4. Debug View

   The composer provides an option for users to turn on the debug view on the graph, which represents the graph in the exact same topology its implementation architecture. In the debug view all the abstractions used for visualization simplification have been restored to their raw structures:

   • Connections be expanded to a standalone node with a connection component inside, whose handles be linked to an input component and an output component.

   • Connections cannot be set up from an Output to an Input directly, even after both of them have been linked to some Input and Output components

   Debug view is useful in inspecting your graph if you notice anything abnormal.

Subgraph

Introduction of the Subgraph concept increases the modularity and reusability of Component based programming significantly. Subgraph is nothing different from a regular graph except that it has exported interfaces which allow it to be re-used in other graphs, and a subgraph itself can also use other subgraphs.

When a subgraph is used in other graphs, it will be abstracted as a “Subgraph” type component inside a subgraph node, only the exported interfaces are showed in the same node with all the other details being concealed. Components in the parent graph can interact with the subgraph through those interfaces.

1. Creating a Subgraph

   There is not too much difference between creating a regular graph and a subgraph. After creating a regular graph the user can export interfaces so that the graph can be used as a subgraph in other graphs. To export interfaces, the user need to add an interface node from the context menu by right clicking on the canvas. Then if the user right click an the newly created interface he or she can add interfaces from the context menu of that interface node. After adding interfaces the user can point them to the actual component which is to be exported by dragging a connection from the interface to the target component. To make the exported interfaces effective the graph file needs to be saved.

2. Using a Subgraph

   In order to use a subgraph in the current graph, the user needs to Drag and Drop a “Subgraph” node from “Miscellaneous” group in the catalog to the canvas, then update the “location” property to load the subgraph from the filesystem.

3. Viewing a Subgraph

   A double-click on the subgraph component will trigger a switch of the current graph view to the subgraph view, where the internal details of the subgraph is displayed. A navigation bar will show up on the top of the graph window once a subgraph view is activated so that the user can switch the view between subgraphs and their parent graph.

Backdrop

Backdrop is a tool used to group the nodes in the current graph, which can be created through Drag&Drop from “Miscellaneous” group in the catalog. After a Backdrop rectangle is created, it can be moved and resized to include a number of nodes, so to group them together. Once a Backdrop is created, the nodes inside will be moved altogether while the Backdrop itself is moved.

User can rename a Backdrop by double-clicking on its name or change its color through the color picker.

Run Graph

A Graph can be run through Composer using the Play button on the toolbar. Output of the running graph can be seen on the console window. A running graph can be stopped using the Stop button on the toolbar.

Options

System Options

Options	Description	Range
Local Machine	Run Graph on Local Machine
Remote Machine	Run Graph on Remote Machine
Platform Config File	Select a platform config file. For x86_64, Default value is /opt/nvidia/graph-composer/target_x86_64.yaml. For ARM 64, Default value is /opt/nvidia/graph-composer /target_aarch64.yaml	Required

Runtime Options

Options	Description	Range
Environment Variables	Separated List of environment variables to be set before running the graph	Optional
Manifest File	Use an existing manifest file	Optional
Use Fresh Manifest	Re-install graph and generate a manifest file	Optional
Resources File	Use a existing resources file If not provided, tool will generate a temp resources file in /tmp/	Optional
Remove Install Directory	Remove graph install directory during exit	Optional
Subgraphs	Comma separated list of paths to subgraphs used by the graph. Paths can be absolute or relative to the main graph.	Optional

Remote Machine Options

Options	Description	Range
Target	Address of the remote machine uname@host	Required when Remote Machine is selected
Password	Password of the remote machine	Required when Remote Machine is selected

Note

To run graph on the remote machine, install the following packages:

1. openssh-client

2. sshfs

Also generate a ssh key pair using ssh-keygen.

Comparison of Different Work Flows

Work Flow	Description	Host Requirement	Target Requirement	Remote Requirement	DS Installation
Local Execution	both the Composer and graphs run on the host machine	Ubuntu 20.04 with Nvidia GPU	N/A	N/A	DS 6.1 required on the host
Remote Execution	the Composer runs on the host machine and graphs on the target machine	Ubuntu 20.04 with Nvidia GPU	Ubuntu 20.04 on Jetson Xavier/Orin	N/A	DS 6.1 require on the target
Remote Access	the Composer runs on the host machine and user access the UI from a remote machine	Ubuntu 20.04 with Nvidia GPU	Ubuntu 20.04 on Jetson Xavier/Orin	Linux/Windows with Browser Support	DS 6.1 required on host or target

Build Container Image

A container image can be build through the Composer using the Docker button on the toolbar.

Options

Runtime Options

Options	Description	Range
Configuration File	Select a Container Builder configuration file	Required
Platform Config File	Select a platform config file	Required

Buttons

Options	Description
Create	Create a new Container Builder Config File using a template
Edit	Edit an existing Container Builder Config File
Build	Build the Container

Generate Gstreamer Extension

Note

If the Deepstream SDK package is not installed, Generate Extension button will be disabled.

Options

Options	Description	Range
GStreamer Elements List	Comma Separated Elements List	Required
Target Output Directory	Directory where the extension will be generated	Required

Buttons	Description
Install Dependencies	Install extension development dependencies
Generate	Generates the extension
Build	Builds the extension using Bazel

Remote Access

Composer supports remote access through two streaming options. 1) WebSocket 2) WebRTC

Comparison of available options

Technology	Devices	Benefits
WebSocket	Web browsers (all devices)	Simple to distribute, and to customize for custom experiences. Supports software encoding if hardware encoders are not present on the host.
WebRTC	Web browsers (all devices)	Simple to distribute. Strong support across devices and platforms. Lower latency then the WebSocket alternative.

WebSocket streaming

By leveraging the new WebSocket live-streaming, it is now possible to stream Graph Composer to web browsers.

Using the WebSocket live-stream is a straightforward process:

1. In a command prompt or terminal, run the command: composer --enable-streaming websocket

2. Find the IP address of the machine hosting the Graph Composer application on your local network

3. Navigate to the stream page in your favorite web browser: http://<server IP address>:8211/streaming/client

4. Click the “play” button to start streaming your application

To enable software encoding on the hardware where hardware encoding is not available, use the following command:

composer --enable-streaming websocket --enable-software-encoding

Note

Sometimes the above URL is not accessible because the port 8211 might have been occupied, in the case we need to replace 8211 with the right HTTP port number displayed in the console.

Note

Browser should be updated to the latest version.

Note

For a best user experience, it is recommended that only a single User at a time be provided with mouse or keyboard control while others remain in spectator mode.

WebRTC streaming

By leveraging the new WebRTC live-streaming, it is now possible to stream Graph Composer to web browsers.

Using the WebRTC live-stream is a straightforward process:

1. In a command prompt or terminal, run the command: composer --enable-streaming webrtc

2. Find the IP address of the machine hosting the Graph Composer application on your local network

3. Navigate to the stream page in your favorite web browser: http://<server IP address>:8011/streaming/webrtc-client?server=<server IP address

4. Click the “play” button to start streaming your application

Note

Sometimes the above URL is not accessible because the port 8011 might have been occupied, in the case we need to replace 8011 with the right HTTP port number displayed in the console.

Note

For a best user experience, it is recommended that only a single User at a time be provided with mouse or keyboard control while others remain in spectator mode.

Note

WebRTC Streaming is supported only on RTX GPUs.

Restrictions

1. Currently the Composer is well supported only on NVIDIA GPU setup.

Trouble Shooting

Instructions for resolving issues when running the Graph Composer.

Component Type not found

In some cases users will see such messages as “Component type invalid” while opening a graph, which means some component types are not available in the current component list thus the loaded graph become incomplete. There are 2 major reasons for the issue: 1. The component list is empty or standalone 2. The graph is not compatible with the latest component list It is recommended for users to close the graph without saving it and perform “repo sync” through the registry menu and try again with the updated list. If the issue still occurs, users should review the graph to figure out which component is missing and replace it with the proper one from the latest component list.

Failure During the Repo Sync

Sometimes the “repo sync” process takes too long or gets stuck in the middle due to network issues and users can’t close the “repo sync” progress bar, in this case it is recommended to close the Graph Composer app and restart it again.