Skip to content

Latest commit

 

History

History
executable file
·
633 lines (467 loc) · 44.9 KB

usage.md

File metadata and controls

executable file
·
633 lines (467 loc) · 44.9 KB
Raw

User Guide

To Wrapyfi your code:

from wrapyfi.connect.wrapper import MiddlewareCommunicator

class TheClass(MiddlewareCommunicator)
        ...
           
        @MiddlewareCommunicator.register(...)
        @MiddlewareCommunicator.register(...)
        def encapsulated_method(...):
            ...
            return encapsulated_a, encapsulated_b
        
        def encapsulating_method(...)
            ...
            encapsulated_a, encapsulated_b = self.encapsulated_method(...)
            ...
            return result,


the_class = TheClass()
the_class.activate_communication(the_class.encapsulated_method, mode="publish")
while True:
    the_class.encapsulating_method(...)

The primary component for facilitating communication is the MiddlewareCommunicator. To register the methods for a given class, the class should inherit the MiddlewareCommunicator. Any method decorated with @MiddlewareCommunicator.register(<Data structure type>, <Communicator>, <Class name>, <Topic name>) is automatically registered by Wrapyfi.

The <Data structure type> is the publisher/listener type for a given method's return. The supported data types are listed here.

The <Communicator> defines the communication medium e.g.: yarp, ros2, ros, or zeromq. The default communicator is zeromq but can be replaced by setting the environment variables WRAPYFI_DEFAULT_COMMUNICATOR or WRAPYFI_DEFAULT_MWARE (WRAPYFI_DEFAULT_MWARE overrides WRAPYFI_DEFAULT_COMMUNICATOR when both are provided) to the middleware of choice e.g.:

        export WRAPYFI_DEFAULT_COMMUNICATOR=yarp

The <Class name> serves no purpose in the current Wrapyfi version, but has been left for future support of module-level decoration, where the methods don't belong to a class, and must therefore have a unique identifier for declaration in the configuration files.

The <Topic name> is the name used for the connected topic and is dependent on the middleware platform. The listener and publisher receive the same topic name.

The @MiddlewareCommunicator.register decorator is defined for each of the method's returns in the same order. As shown in the example above, the first decorator defines the properties of encapsulated_a's publisher and listener, whereas the second decorator belongs to encapsulated_b. A decorated method must always return a tuple which can easily be enforced by adding a comma after the return in case a single variable is returned. Lists are also supported for single returns e.g.:


        @MiddlewareCommunicator.register([..., {...}], [..., {...}], [...])
        @MiddlewareCommunicator.register(...)
        def encapsulated_method(...):
            ...
            encapsulated_a = [[...], [...], [...]]
            ...
            return encapsulated_a, encapsulated_b

Methods with a single return should be followed by a comma e.g., return encapsulated a, . This explicitly casts the return as a tuple to avoid confusion with list returns as single return element/s

Each of the list's returns is encapsulated with its own publisher and listener, with the named arguments transmitted as a single dictionary within the list. Notice that encapsulated_a returns a list of length 3, therefore, the first decorator contains 3 list configurations as well. This is useful especially when transmitting multiple images or audio chunks over YARP, ROS, and ROS 2. Note that by using a single NativeObject as a <Data structure type>, the same can be achieved. However, the implementation of the NativeObject for most middleware serializes the objects as strings before transmission. The NativeObject may result in a greater overhead and should only be used when multiple nesting depths are required or the objects within a list are not within the supported data structure types.

Argument Passing

The $ symbol is used in Wrapyfi to specify that a decorator should update its arguments according to the arguments of the decorated method. This can be useful when the decorator needs to modify its behavior during runtime. For instance:

...
        @MiddlewareCommunicator.register('NativeObject', 
           '$0', 'ExampleCls', '/example/example_arg_pass', 
           carrier='tcp', should_wait='$blocking')
          def example_arg_pass(self, mware, msg='', blocking=True):

Setting the decorator's keyword argument should_wait='$blocking' expects the decorated method to receive a boolean blocking argument, altering the encapsulating decorator's behavior when the encapsulated method is called. Setting the decorator's second argument to $0 acquires the value of mware (the first argument passed to example_arg_pass) and sets it as the middleware for that method. These arguments take effect on the first invocation of a method. Changing arguments after the first invocation results in no change in behavior, unless a MiddlewareCommunicator inheriting class for a given method is closed.

Closing and Deleting Classes

Currently, closing a connection requires closing all connections established by every method within that class.

Selectively deactivating method connections is not supported [![planned](https://custom-icon-badges.demolab.com/badge/planned%20for%20Wrapyfi%20v0.5-%23C2E0C6.svg?logo=hourglass&logoColor=white)](https://github.com/modular-ml/wrapyfi/issues/99 "planned link")

To close and delete a MiddlewareCommunicator inheriting class means that the middleware connection will be disconnected gracefully. The class references will be removed from all registries, the communication ports will be freed, and the instance will be destroyed. To close a class instance:

# assuming an existing instance-> example_instance = ExampleCls()
example_instance.close()
del example_instance

Configuration

In order to establish communication using any communication pattern supported by Wrapyfi, the methods need to be activated by setting them to behave according to a compatible communication mode

Communication Modes

Modes are configurations that define the behavior of a method, i.e., should the method be executed?, should the method await a message from another method?, should the method publish its message?, or should the method await an acknowledgement from a message requester?

Selecting the mode for each method is tha starting point for establishing communication through Wrapyfi. Wrapyfi supports the common PUB/SUB and REQ/REP communication patterns, with different modes supported for each. The MiddlewareCommunicator's child class method modes can be independently set to accomodate the communication pattern. We separate the modes into their corresponding patterns:

Mode-Agnostic

  • none(default): Run the method as usual without triggering publish, listen, receive, reemit, transceive, request or reply. hint: Setting the mode to None (or null within a yaml configuration file) has the same effect
  • disable: Disables the method and returns None for all its returns. Caution should be taken when disabling a method since it could break subsequent calls

Modes for Publishers and Listeners|Subscribers (PUB/SUB)

  • publish: Run the method and publish the results using the middleware's transmission protocol. The results of executing the method are returned directly from the method: The method runs as expected, with the additional publishing of its returns

  • listen: Skip the method and wait for the publisher with the same port name to transmit a message, eventually returning the received message

  • receive: Similar in behavior to the listen mode, but receives the input as an argument and executes the method assuming the argument/s are received from a publisher. The method is then run based on the received argument/s value/s and the returns are resulting from running the method itself, rather than passively listening for input only

  • reemit: Similar in behavior to the receive mode, but also publishes the returns eventually, similar to the publish mode

  • transceive: Similar in behavior to the publish mode, however, unlike publishing methods, it listens for returns from another publisher, and returns the resulting message instead, similar to the listen mode

The possible mode combinations for each method to communicate successfully with its corresponding methods operating on the same topic are:

  • publish <-> listen: Both method returns are identical
  • publish <-> disable: Method returns are not identical: This only works when publish decorater argument should_wait=False, since there is no listener, given disable mode always returns None and does not establish a connection with the publisher
  • publish <-> none: Method returns are not identical: This only works when publish decorater argument should_wait=False, since there is no listener, given "none"/None mode always runs the method and acquires its returns but does not establish a connection with the publisher
  • publish <-> receive: Method returns are not identical
  • transceive <-> reemit: Both method returns could be identical

Modes for Servers and Clients (REQ/REP)

  • reply: Run the method and publish the results using the middleware's transmission protocol. Arguments are received from the requester
  • request: Send a request to the replier in the form of arguments passed to the method. Skip the method and wait for the replier with the same port name to transmit a message, eventually returning the received message

The possible mode combinations for each method to communicate successfully with its corresponding methods operating on the same topic are:

  • request <-> reply: Both method returns are identical. For every request (client) there should be a corresponding reply (server) and would not operate (hangs indefinitely) without acknowledgement from the server

Activating Communication Modes

The mode of each method can be set by calling:

activate_communication(<Method name>, mode=<Mode>)

where <Method name> is the method's name (string name of method by definition) and <Mode> is the transmission mode ("publish", "listen","receive", "transceive", reemit, "reply", request, "none" | None, "disable") depending on the communication pattern . The activate_communication method can be called multiple times. <Method name> could also be a class instance method, by calling:

activate_communication(<MiddlwareCommunicator instance>.method_of_class_instance, mode=<Mode>)

Configuration Files

for each decorated method within the class. This however requires modifying your scripts for each machine or process running on Wrapyfi. To overcome this limitation, use the ConfigManager e.g.:

from wrapyfi.config.manager import ConfigManager
ConfigManager(<Configuration file path *.yml>)

The ConfigManager is a singleton that must be called once before the initialization of any MiddlewareCommunicator. Initializing it multiple times has no effect. This limitation was created by design to avoid loading the configuration file multiple times.

The <Configuration file path *.yml>'s configuration file has a very simple format e.g.:

TheClass:
  encapsulated_method: "publish"

where TheClass is the class name, encapsulated_method is the method's name, and publish is the transmission mode. This is useful when running the same script on multiple machines, where one is set to publish and the other listens. Multiple instances of the same class' method can have different modes, which can be set independently using the configuration file. This can be achieved by providing the mode as a list:

TheClass:
  encapsulated_method: 
        "publish"
        null
        "listen"
        "listen"
        "disable"
        null

where the list element index corresponds to the instance index. When providing a list, the number of list elements should correspond to the number of instances. If the number of instances exceeds the list length, the script exits and raises an error.

Communication Patterns

Wrapyfi supports the publisher-subscriber (PUB/SUB) pattern as well as the request-reply (REQ/REP) pattern. The PUB/SUB pattern assumes message arguments are passed from the publisher-calling script to the publishing method. The publisher executes the method and the subscriber (listener) merely triggers the method call, awaits the publisher to execute the method, and returns the publisher's method returns. The REQ/REP pattern on the other hand assumes arguments from the client (requester) are sent to the server (responder or replier). Once the server receives the request, it passes the arguments to its own method, executes it, and replies to the client back with its method returns.

Communication patterns in Wrapyfi are set by passing the configuration mode argument to activate_communication method as described in the configuration documentation.

in REQ/REP, the requester transmits all arguments passed to the method as a dictionary encoded as a string. This is not ideal for predefined services, where the service expects a certain object/message type. A better approach would include the option to pass a single item of a certain value and type [![planned](https://custom-icon-badges.demolab.com/badge/planned%20for%20Wrapyfi%20v0.5-%23C2E0C6.svg?logo=hourglass&logoColor=white)](https://github.com/modular-ml/wrapyfi/issues/99 "planned link")

Publishers and Listeners|Subscribers (PUB/SUB)

The publishers and listeners of the same message type should have identical constructor signatures. The current Wrapyfi version supports 4 universal message types for all middleware. The extended types such as ROSMessage and ROS2Message are exclusive to the provided middleware.

YARP:

YARP publishers remain [persistent](https://www.yarp.it/latest/persistent_connections.html#:~:text=When%20a%20connection%20is%20made%20between%20two%20YARP,made%20whenever%20possible.%20These%20are%20called%20%22persistent%20connections%22.). To disable persistence, pass the argument `persistent=False` to the `@MiddlewareCommunicator.register` decorator.

All messages are transmitted using the yarp Python bindings

  • Image: Transmits and receives a cv2 or numpy image using either yarp.BufferedPortImageRgb or yarp.BufferedPortImageFloat. When JPG conversion is specified, it uses a yarp.BufferedPortBottle message carrying a JPEG encoded string instead
  • AudioChunk: Transmits and receives a numpy audio chunk with the sound properties using yarp.Port transporting yarp.Sound
  • NativeObject: Transmits and receives a json string supporting all native Python objects, numpy arrays and other formats using yarp.BufferedPortBottle
  • Properties: Transmits properties planned

ROS:

ROS requires a custom message to handle audio. This message must be compiled first before using Wrapyfi with ROS Audio. 
Refer to [these instructions for compiling Wrapyfi ROS services and messages](https://github.com/modular-ml/wrapyfi_ros_interfaces/blob/master/README.md).

All messages are transmitted using the rospy Python bindings as topic messages

  • Image: Transmits and receives a cv2 or numpy image using sensor_messages.msg.Image. When JPG conversion is specified, uses the sensor_messages.msg.CompressedImage message instead
  • AudioChunk: Transmits and receives a numpy audio chunk using wrapyfi_ros_interfaces.msg.ROSAudioMessage
  • NativeObject: Transmits and receives a json string supporting all native Python objects, numpy arrays, and other formats using std_msgs.msg.String
  • Properties: Transmits and receives parameters to/from the parameter server using the methods rospy.set_param and rospy.get_param respectively
  • ROSMessage: Transmits and receives a single ROS message per return decorator. Note that currently, only common ROS interface messages are supported and detected automatically. This means that messages defined in common interfaces such as std_msgs, geometry_msgs, and sensor_msgs can be directly returned by the method do not need to be converted to native types

ROS2:

ROS 2 requires a custom message to handle audio. This message must be compiled first before using Wrapyfi with ROS 2 Audio. 
Refer to [these instructions for compiling Wrapyfi ROS 2 services and messages](https://github.com/modular-ml/wrapyfi_ros2_interfaces/blob/master/README.md).

All messages are transmitted using the rclpy Python bindings as topic messages

  • Image: Transmits and receives a cv2 or numpy image using sensor_messages.msg.Image. When JPG conversion is specified, uses the sensor_messages.msg.CompressedImage message instead
  • AudioChunk: Transmits and receives a numpy audio chunk using wrapyfi_ros2_interfaces.msg.ROS2AudioMessage
  • NativeObject: Transmits and receives a json string supporting all native Python objects, numpy arrays, and other formats using std_msgs.msg.String
  • Properties: Transmits properties planned
  • ROS2Message: Transmits and receives a single ROS 2 message per return decorator

ZeroMQ:

ZeroMQ exchanges in REQ/REP rely on a broker with a dedicated socket. By default, Wrapyfi will not spawn a new connection to the socket when multiple threads are created. For multi-threaded applications, this leads to race conditions. We avoid that by detecting whether a new instance of the socket is available in the thread's local storage. This multi-threading-friendly mode is enabled by passing `multi_threaded=True` to the `@MiddlewareCommunicator.register` decorator. This is only recommended when registering methods that are going to be multi-threaded.

All messages are transmitted using the zmq Python bindings. Transmission follows the proxied XPUB/XSUB pattern

  • Image: Transmits and receives a cv2 or numpy image wrapped in the NativeObject construct. Note that all Image types are transmitted as multipart messages, where the first element is the topic name and the second element is the header (e.g., timestamp), and the third element is the image itself
  • AudioChunk: Transmits and receives a numpy audio chunk wrapped in the NativeObject construct
  • NativeObject: Transmits and receives a json string supporting all native Python objects, numpy arrays and other formats using zmq context.socket(zmq.PUB).send_multipart for publishing and zmq context.socket(zmq.SUB).receive_multipart for receiving messages. The zmq.PUB socket is wrapped in a zmq.proxy to allow multiple subscribers to the same publisher. Note that all NativeObject types are transmitted as multipart messages, where the first element is the topic name and the second element is the message itself (Except for Image)
  • Properties: Transmits properties planned

Websocket:

Websocket assumes a server is running on the specified address and port. The forwarding of messages can only be done manually by the user. An example server can be found [here](https://github.com/fabawi/wrapyfi/tree/main/wrapyfi/examples/websockets/websocket_server.py) 

Unlike the majority of middleware supported by Wrapyfi, websockets are bidirectional, meaning that the publisher can also be a listener. This allows Wrapyfi to support multiple publishers on the same topic ([namespaces](https://socket.io/docs/v4/namespaces/) and [rooms](https://socket.io/docs/v4/rooms/))

All messages are transmitted using the python-socketio Python bindings. Transmission follows the socket.io protocol

  • Image: Transmits and receives a cv2 or numpy image wrapped in the NativeObject construct
  • AudioChunk: Transmits and receives a numpy audio chunk wrapped in the NativeObject construct
  • NativeObject: Transmits and receives a json string supporting all native Python objects, numpy arrays and other formats using socketio.emit for publishing and socketio.on for receiving messages
  • Properties: Transmits properties planned

Zenoh:

All messages are transmitted using the eclipse-zenoh Python bindings. Transmission follows the zenoh protocol

  • Image: Transmits and receives a cv2 or numpy image wrapped in the NativeObject construct as zenoh.Bytes
  • AudioChunk: Transmits and receives a numpy audio chunk wrapped in the NativeObject construct as zenoh.Bytes
  • NativeObject: Transmits and receives a json string supporting all native Python objects, numpy arrays and other formats as zenoh.Bytes using zenoh.session.key.put for publishing and an asynchronus callback for receiving messages

MQTT:

MQTT runs on a public online broker by default broker.emqx.io for convenience (no setup required), however, it is recommended to use a local broker like [Mosquitto](https://mosquitto.org/download/) for production.

All messages are transmitted using the paho-mqtt Python bindings. Transmission follows the MQTT protocol

  • Image: Transmits and receives a cv2 or numpy image wrapped in the NativeObject construct
  • AudioChunk: Transmits and receives a numpy audio chunk wrapped in the NativeObject construct
  • NativeObject: Transmits and receives a json string supporting all native Python objects, numpy arrays and other formats using client.publish for publishing and client.on_message for receiving messages
  • Properties: Transmits properties planned

Servers and Clients (REQ/REP)

The servers and clients of the same message type should have identical constructor signatures. The current Wrapyfi version supports 3 universal message types for all middleware. The extended types such as ROSMessage and ROS2Message are exclusive to the provided middleware.

YARP:

All messages are transmitted using the yarp Python bindings for RPC communication. The requester encodes its arguments as a json string supporting all native Python objects, numpy arrays, and other formats using yarp.Bottle. The requester formats its arguments as ([args], {kwargs})

  • Image: Transmits and receives a cv2 or numpy image encoded as a json string using yarp.Bottle. JPG conversion is currently not supported
  • AudioChunk: Transmits and receives a numpy audio chunk encoded as a json string using yarp.Bottle
  • NativeObject: Transmits and receives a json string supporting all native Python objects, numpy arrays, and other formats using yarp.Bottle

(ROS):

ROS requires a custom service to handle audio. This service must be compiled first before using Wrapyfi with ROS Audio. 
Refer to [these instructions for compiling Wrapyfi ROS services and messages](https://github.com/modular-ml/wrapyfi_ros_interfaces/blob/master/README.md).

All messages are transmitted using the rospy Python bindings as services. The requester encodes its arguments as a json string supporting all native Python objects, numpy arrays, and other formats using std_msgs.msg.String. The requester formats its arguments as ([args], {kwargs})

  • Image: Transmits and receives a cv2 or numpy image using sensor_messages.msg.Image JPG conversion is currently not supported
  • AudioChunk: Transmits and receives a numpy audio chunk using wrapyfi_ros_interfaces.msg.ROSAudioMessage
  • NativeObject: Transmits and receives a json string supporting all native Python objects, numpy arrays, and other formats using std_msgs.msg.String

ROS2:

ROS 2 requires custom services to handle arbitrary messages. These services must be compiled first before using Wrapyfi in this mode. 
Refer to [these instructions for compiling Wrapyfi ROS 2 services](https://github.com/modular-ml/wrapyfi_ros2_interfaces/blob/master/README.md).

All messages are transmitted using the rclpy Python bindings as services. The requester encodes its arguments as a json string supporting all native Python objects, numpy arrays, and other formats using std_msgs.msg.String. The requester formats its arguments as ([args], {kwargs})

  • Image: Transmits and receives a cv2 or numpy image using sensor_messages.msg.Image
  • AudioChunk: Transmits and receives a numpy audio chunk using wrapyfi_ros2_interfaces.msg.ROS2AudioMessage
  • NativeObject: Transmits and receives a json string supporting all native Python objects, numpy arrays, and other formats using std_msgs.msg.String

ZeroMQ:

All messages are transmitted using the zmq Python bindings. Transmission follows the proxied XREP/XREQ pattern The requester encodes its arguments as a json string supporting all native Python objects, numpy arrays, and other formats using zmq context.socket(zmq.REQ).send_multipart. The requester formats its arguments as ([args], {kwargs})

  • Image: Transmits and receives a cv2 or numpy image wrapped in the NativeObject construct
  • AudioChunk: Transmits and receives a numpy audio chunk wrapped in the NativeObject construct
  • NativeObject: Transmits and receives a json string supporting all native Python objects, numpy arrays, and other formats using zmq context.socket(zmq.REP) for replying and zmq context.socket(zmq.REQ) for receiving messages

Publisher- and Listener-specific Arguments

Differences are expected between the returns of publishers and listeners, sometimes due to compression methods (e.g., setting `jpg=True` when transmitting an **Image** compresses the image but the encoding remains the same), intentional setting of different devices for different tensors (refer to [device mapping for tensors](#device-mapping-for-tensors)), and differences in library versions between receiving and transmitting plugins (refer to [plugins](#plugins)).

To direct arguments specifically toward the publisher or subscriber without exposing one or the other to the same argument values, the corresponding arguments can be added to the dictionary listener_kwargs to control the listener only, or publisher_kwargs to control the publisher only. Both dictionaries can be passed directly to the Wrapyfi decorator. Since the transmitting and receiving arguments should generally be the same regardless of the communication pattern, publisher_kwargs and listener_kwargs also apply to the servers and clients respectively.

Communication Schemes

We introduce three communication schemes: Mirroring, Channeling, and Forwarding. These schemes are communication forms that can be useful in different scenarios.

Mirroring

For the REQ/REP pattern, mirroring is a communication scheme that allows a client to send arguments to a server, and receive the method returns back from the server. As for the PUB/SUB pattern, mirroring allows a publisher to send the returns of a method to a subscriber based on the publisher's method arguments. Following both patterns, the returns of a method are mirrored on the receiver and the sender side. This is useful when the pipeline for each receiver is identical, but we would like to delegate the processing to different publishers when processing requires more resources than a single publisher can provide.

Mirroring Example

In the mirroring_example.py, the module transmits a user input message from the publisher to a listener (PUB/SUB pattern), and displays the message along with other native objects on the listener and publisher side. Similarly, we transmit a user input message from the server to a client (REQ/REP pattern), when the client requests the message from the server. The example can be run from the examples/communication_schemes/ directory.

Forwarding

Forwarding is a communication scheme that allows a server or publisher to forward the method arguments to another server or publisher (acting as a client or listener), and in return, forwards the received messages to another client or listener. This is useful when the server or publisher is not able to communicate with the client or listener directly due to limited middleware support on the client or listener side. The middle server or publisher can then act as a bridge between the two, and forward the messages between them, effectively chaining the communication. The chain can be extended and is not limited to two servers or publishers.

Forwarding Example

In the forwarding_example.py, the module constantly publishes a string from chain_A to a listener on chain_A. The chain_A listener then forwards the message by publishing to chain_B. The string is then forwarded to a third instances which listens exclusively to chain_B, without needing to support the middleware used by chain_A. The example can be run from the examples/communication_schemes/ directory.

Channeling

Channeling differs from mirroring, in that there are multiple returns from a method. Disabling one or more of these returns is possible, allowing the server or publisher to transmit the message to multiple channels, each with a different topic, and potentially, a different middleware. This is useful for transmitting messages using the same method, but to different receivers based on what they choose to receive. Not all clients or subscribers require all the messages from a method, and can therefore selectively filter out what is needed and operate on that partial return.

Channeling Example

In the channeling_example.py, the module constantly publishes three data types (NativeObject, Image, and AudioChunk) over one or more middlware. The listeners can then choose to receive one or more of these data types, depending on the middleware they support. When --mware_... for one of the channels is not provided, it automatically disables the topic for that channel/s and returns a None type value. The example can be run from the examples/communication_schemes/ directory.

Middleware

The `<Data structure type>` is the object type for a given method's return. The supported data types are listed [here](<#data-structure-types>) section.

Wrapyfi natively supports a number of middleware. However, more middleware could be added by:

  • Creating a derived class that inherits from the base classes wrapyfi.connect.Listener, wrapyfi.connect.Publisher, wrapyfi.connect.Client, or wrapyfi.connect.Server depending on the communication pattern to be supported
  • Decorating the classes inside scripts residing within:
    • the listeners directory with @Listeners.register(<Data structure type>, <Communicator>)
    • the publishers directory with @Publishers.register(<Data structure type>, <Communicator>)
    • the clients directory with @Clients.register(<Data structure type>, <Communicator>)
    • the servers directory with @Servers.register(<Data structure type>, <Communicator>)
  • Appending the script path where the class is defined to the WRAPYFI_MWARE_PATHS environment variable
  • Ensure that the middleware communication pattern scripts reside within directories named listeners, publishers, clients, or servers nested inside the WRAPYFI_MWARE_PATH and that the directory contains an __init__.py file

Natively Supported Middleware

  • YARP
  • ROS
  • ROS 2
  • ZeroMQ [beta feature]:
    • should_wait trigger introduced with event monitoring
    • Event monitoring currently cannot be disabled planned
  • Websocket Only PUB/SUB [alpha support]
  • Zenoh Only PUB/SUB [alpha support]
  • MQTT Only PUB/SUB [alpha support]

Plugins

The NativeObject message type supports structures beyond native Python objects. Wrapyfi already supports a number of non-native objects including numpy arrays and tensors. Wrapyfi can be extended to support objects by using the plugin API. All currently supported plugins by Wrapyfi can be found in the plugins directory. Plugins can be added by:

  • Creating a derived class that inherits from the base class wrapyfi.utils.Plugin
  • Overriding the encode method for converting the object to a json serializable string. Deserializing the string is performed within the overridden decode method
  • Specifying custom object properties by defining keyword arguments for the class constructor. These properties can be passed directly to the Wrapyfi decorator
  • Decorating the class with @PluginRegistrar.register and appending the plugin to the list of supported objects
  • Appending the script path where the class is defined to the WRAPYFI_PLUGIN_PATHS environment variable
  • Ensure that the plugin resides within a directory named plugins nested inside the WRAPYFI_PLUGIN_PATHS and that the directory contains an __init__.py file

Plugin Example

An example for adding a plugin for a custom Astropy object is provided in the astropy_example.py example. In the example, we append the example's directory to the WRAPYFI_PLUGIN_PATHS environment variable and import the plugin. The plugin (astropy_tables.py) in the plugins directory is then used to encode and decode the custom object (from within the examples/encoders/ directory):

# create the publisher with default middleware (changed with --mware). The plugin is automatically loaded
Python3 astropy_example.py --mode publish
# create the listener with default middleware (changed with --mware). The plugin is automatically loaded
Python3 astropy_example.py --mode listen

from the two terminal outputs, the same object should be printed after typing a random message and pressing enter:

Method result: [{'message': 'hello world', 'astropy_table': <Table length=3>
  name     flux 
 bytes8  float64
-------- -------
source 1     1.2
source 2     2.2
source 3     3.1, 'list': [1, 2, 3]}, 'string', 0.4344, {'other': (1, 2, 3, 4.32)}]

Due to differences in versions, the decoding may result in inconsitent outcomes, which must be handled for all versions e.g., MXNet plugin differences are handled in the existing plugin.

Data Structure Types

Wrapyfi primarily supports Image, AudioChunk, and NativeObject types, with additional types supported for different middleware and communication patterns. Other than native Python objects, the following objects are supported by NativeObject:

  • numpy.ndarray and numpy.generic
  • pandas.DataFrame and pandas.Series (pandas v1)
  • torch.Tensor
  • tensorflow.Tensor and tensorflow.EagerTensor
  • mxnet.nd.NDArray
  • jax.numpy.DeviceArray
  • trax.ArrayImpl -> jaxlib.xla_extension.ArrayImpl
  • paddle.Tensor
  • PIL.Image
  • pyarrow.StructArray
  • xarray.DataArray and xarray.Dataset
  • cupy.ndarray
  • dask.array.Array and dask.dataframe.DataFrame
  • zarr.core.Array and zarr.core.Group
  • pint.Quantity

Device Mapping for Tensors

To map tensor listener decoders to specific devices (CPUs/GPUs), add an argument to tensor data structures with direct GPU/TPU mapping to support re-mapping on mirrored nodes e.g.,

@PluginRegistrar.register
class MXNetTensor(Plugin):
    def __init__(self, load_mxnet_device=None, map_mxnet_devices=None, **kwargs):

where map_mxnet_devices should be {'default': mxnet.gpu(0)} when load_mxnet_device=mxnet.gpu(0) and map_mxnet_devices=None. For instance, when load_mxnet_device=mxnet.gpu(0) or load_mxnet_device="cuda:0", map_mxnet_devices can be set manually as a dictionary representing the source device as key and the target device as value for non-default device maps.

Suppose we have the following Wrapyfied method:


        @MiddlewareCommunicator.register("NativeObject", args.mware, "Notify", "/notify/test_native_exchange",
                                         carrier="tcp", should_wait=True, load_mxnet_device=mxnet.cpu(0), 
                                         map_mxnet_devices={"cuda:0": "cuda:1", 
                                                             mxnet.gpu(1): "cuda:0", 
                                                             "cuda:3": "cpu:0", 
                                                             mxnet.gpu(2):  mxnet.gpu(0)})
        def exchange_object(self):
            msg = input("Type your message: ")
            ret = {"message": msg,
                   "mx_ones": mxnet.nd.ones((2, 4)),
                   "mxnet_zeros_cuda1": mxnet.nd.zeros((2, 3), ctx=mxnet.gpu(1)),
                   "mxnet_zeros_cuda0": mxnet.nd.zeros((2, 3), ctx=mxnet.gpu(0)),
                   "mxnet_zeros_cuda2": mxnet.nd.zeros((2, 3), ctx=mxnet.gpu(2)),
                   "mxnet_zeros_cuda3": mxnet.nd.zeros((2, 3), ctx=mxnet.gpu(3))}
            return ret,

then the source and target gpus 1 & 0 would be flipped, gpu 3 would be placed on cpu 0, and gpu 2 would be placed on gpu 0. Defining mxnet.gpu(1): mxnet.gpu(0) and cuda:1: cuda:2 in the same mapping should raise an error since the same device is mapped to two different targets.

The plugins supporting remapping are:

  • mxnet.nd.NDArray
  • torch.Tensor
  • paddle.Tensor
  • cupy.ndarray ONLY SUPPORTS CUDA DEVICES

Serialization

When encoding dictionaries, `json` supports string keys only and converts any instances of `int` keys to string, causing a difference between the publisher and subscriber returns. It is best to avoid using `int` keys, otherwise handle the difference on the receiving end.

Wrapyfi currently supports JSON as the only serializer. This introduces a number of limitations (beyond serializing native Python objects only by default), including:

  • dictionary keys cannot be integers. Integers are automatically converted to strings
  • Tuples are converted to lists. Sets are not serializable. Tuples and sets are encoded as strings and restored on listening, which resolves this limitation but adds to the encoding overhead. This conversion is supported in Wrapyfi

Environment Variables

Wrapyfi reserves specific environment variable names for the functionality of its internal components:

  • WRAPYFI_PLUGIN_PATHS: Path/s to plugin extension directories
  • WRAPYFI_MWARE_PATHS: Path/s to middleware extension directories. These are simply middleware classes that are not part of the core library
  • WRAPYFI_DEFAULT_COMMUNICATOR or WRAPYFI_DEFAULT_MWARE (WRAPYFI_DEFAULT_MWARE overrides WRAPYFI_DEFAULT_COMMUNICATOR when both are provided): Name of default when none is provided as the second argument to the Wrapyfi decorator

ZeroMQ requires socket configurations that can be passed as arguments to the respective middleware constructor (through the Wrapyfi decorator) or using environment variables. Note that these configurations are needed both by the proxy and the message publisher and listener. The downside to such an approach is that all messages share the same configs. Since the proxy broker spawns once on first trigger (if enabled) as well as a singleton subscriber monitoring instance, using environment variables is the recommended approach to avoid unintended behavior. This can be achieved by setting:

  • WRAPYFI_ZEROMQ_SOCKET_IP: IP address of the socket. Defaults to "127.0.0.1"
  • WRAPYFI_ZEROMQ_SOCKET_PUB_PORT: The publishing socket port. Defaults to 5555
  • WRAPYFI_ZEROMQ_SOCKET_SUB_PORT: The sub-socket port (listening port for the broker). Defaults to 5556
  • WRAPYFI_ZEROMQ_PUBSUB_MONITOR_TOPIC: The topic name for the pub-sub monitor. Defaults to "ZEROMQ/CONNECTIONS"
  • WRAPYFI_ZEROMQ_START_PUBSUB_MONITOR_BROKER: Spawn a new broker for enabling topic discovery since topic discovery is not natively supported on ZeroMQ. Defaults to "True"
  • WRAPYFI_ZEROMQ_PUBSUB_MONITOR_LISTENER_SPAWN: Either spawn the pub-sub monitor listener as a "process" or "thread". Defaults to "process"
  • WRAPYFI_ZEROMQ_START_PROXY_BROKER: Spawn a new broker proxy without running the standalone proxy broker. Defaults to "True"
  • WRAPYFI_ZEROMQ_PROXY_BROKER_SPAWN: Either spawn broker as a "process" or "thread". Defaults to "process")
  • WRAPYFI_ZEROMQ_PARAM_POLL_INTERVAL: Polling interval in milliseconds for the parameter server. Defaults to 1 (currently not supported)
  • WRAPYFI_ZEROMQ_PARAM_REQREP_PORT: The parameter server request-reply port. Defaults to 5659 (currently not supported)
  • WRAPYFI_ZEROMQ_PARAM_PUB_PORT: The parameter server pub-socket port. Defaults to 5655 (currently not supported)
  • WRAPYFI_ZEROMQ_PARAM_SUB_PORT: The parameter server sub-socket port. Defaults to 5656 (currently not supported)
  • WRAPYFI_WEBSOCKET_SOCKET_IP: IP address of the socket. Defaults to "127.0.0.1"
  • WRAPYFI_WEBSOCKET_SOCKET_PORT: The socket port. Defaults to 5000
  • WRAPYFI_WEBSOCKET_NAMESPACE: The socket namespace. Defaults to "/"
  • WRAPYFI_ZENOH_IP: IP address of the Zenoh socket. Defaults to "127.0.0.1"
  • WRAPYFI_ZENOH_PORT: The Zenoh socket port. Defaults to 7447
  • WRAPYFI_ZENOH_MODE: The Zenoh mode indicating whether to use the router as a broker or adopt peer-to-peer communication. Defaults to "peer"
  • WRAPYFI_ZENOH_CONNECT: The Zenoh connect endpoints seperated by a comma e.g., "tcp/127.0.0.1:7447,udp/127.0.0.1:7448". This overrides WRAPYFI_ZENOH_IP and WRAPYFI_ZENOH_PORT. Defaults to an empty list
  • WRAPYFI_ZENOH_LISTEN: The Zenoh listen endpoints seperated by a comma e.g., "tcp/127.0.0.1:7446". Defaults to an empty list
  • WRAPYFI_ZENOH_CONFIG_FILEPATH: The Zenoh configuration file path. Defaults to None. Conflicting keys are overriden by WRAPYFI_ZENOH_IP, WRAPYFI_ZENOH_PORT, WRAPYFI_ZENOH_CONNECT, and WRAPYFI_ZENOH_LISTEN
  • WRAPYFI_MQTT_BROKER_ADDRESS: The MQTT broker address. Defaults to "broker.emqx.io"
  • WRAPYFI_MQTT_BROKER_PORT: The MQTT broker port. Defaults to 1883

ROS and ROS 2 queue sizes can be set by:

  • WRAPYFI_ROS_QUEUE_SIZE: Size of the queue buffer. Defaults to 5
  • WRAPYFI_ROS2_QUEUE_SIZE: Size of the queue buffer. Defaults to 5