Python Execution in OpenVINO Model Server#

Introduction#

Starting with version 2023.3, OpenVINO Model Server supports execution of custom Python code. Such code can execute simple pre- or post-processing as well as complex tasks like image or text generation.

Python execution is enabled via MediaPipe by the built-in PythonExecutorCalculator that allows creating graph nodes to execute Python code. Python nodes can be used as standalone servables (single node graphs) or be part of larger MediaPipe graphs.

Check out the quickstart guide for a simple example that shows how to use this feature.

Check out Generative AI demos for real life use cases.

Building Docker Image#

The publicly available openvino/model_server image on Docker Hub supports Python, but does not come with external modules installed. If Python is all you need then you can use the public image without modification. Otherwise, you will need to extend the public image with additional layers that install any modules required for your Python code to run. For example, let’s say your code requires numpy. In that case, your Dockerfile may look like this:

FROM openvino/model_server:latest
USER root
ENV LD_LIBRARY_PATH=/ovms/lib
ENV PYTHONPATH=/ovms/lib/python
RUN apt update && apt install -y python3-pip git
RUN pip3 install numpy
ENTRYPOINT [ `/ovms/bin/ovms` ]

You can also modify requirements.txt from our python demos and from repository top level directory run make python_image

OvmsPythonModel class#

When deploying a Python node, the Model Server expects a Python file with an OvmsPythonModel class implemented:

class OvmsPythonModel:

    def initialize(self, kwargs):
        """
        `initialize` is called when model server loads graph definition.
        It allows to initialize and maintain state between subsequent execute() calls
        and even graph instances. For gRPC unary, graphs are recreated per request.
        For gRPC streaming, there can be multiple graph instances existing at the same time.
        OvmsPythonModel object is initialized with this method and then shared between all graph instances.
        Implementing this function is optional.

        Parameters:
        -----------
        kwargs : dict
            Available arguments:
            * node_name: string
                Name of the node in the graph
            * input_names: list of strings
                List of input stream names defined for the node in graph
                configuration
            * output_names: list of strings
                List of output stream names defined for the node in graph
                configuration
        -----------
        """
        print("Running initialize...")

    def execute(self, inputs):
        """
        `execute` is called in `Process` method of PythonExecutorCalculator
        which in turn is called by the MediaPipe framework. How MediaPipe
        calls the `Process` method for the node depends on the configuration
        and the two configurations supported by PythonExecutorCalculator are:

        * Regular: `execute` is called with a set of inputs and returns a set of outputs.
        For unary endpoints it's the only possible configuration.

        * Generative: `execute` is called with a set of inputs and returns a generator.
        The generator is then called multiple times with no additional input data and produces
        multiple sets of outputs over time. Works only with streaming endpoints.

        Implementing this function is required.

        Parameters:
        -----------
        * inputs: list of pyovms.Tensor
        -----------

        Returns: list of pyovms.Tensor or generator
        """
        ...
        return outputs

    def finalize(self):
        """
        `finalize` is called when model server unloads graph definition.
        It allows to perform any cleanup actions before the graph definition
        is removed. Implementing this function is optional.
        """
        print("Running finalize...")

initialize#

initialize is called when model server loads graph definition. It allows to initialize and maintain state between subsequent execute calls and even graph instances.

For unary endpoint, graphs are recreated per request.

For gRPC streaming, there can be multiple graph instances existing at the same time.

OvmsPythonModel object is initialized with this method and then shared between all graph instances.

Parameters and return value#

initialize is called with kwargs parameter which is a dictionary. kwargs contain information from node configuration. Considering a sample:

node {
  name: <NODE_NAME>
  ...
  input_stream: "<INPUT_TAG>:<INPUT_NAME>"
  input_stream: "<INPUT_TAG>:<INPUT_NAME>"
  ...
  output_stream: "<OUTPUT_TAG>:<OUTPUT_NAME>"
  output_stream: "<OUTPUT_TAG>:<OUTPUT_NAME>"
  ...
}

All keys are strings. Available keys and values:

Key

Value type

Description

node_name

string

Name of the node in the graph. <NODE_NAME> in the sample above

input_names

list of strings

List of <INPUT_NAME> from all input streams in the sample above

outputs_names

list of strings

List of <OUTPUT_NAME> from all output streams in the sample above

base_path

string

Path to the folder containing handler script

initialize is not expected to return any value.

Error handling#

Signaling that something went wrong should be done by throwing an exception. When model server catches exception from initialize it cleans up all Python resources in the graph (including those belonging to the correctly loaded nodes) and sets the whole graph in unavailable state.

Note: Run Model Server with --log_level DEBUG parameter to get information about errors in the server logs.

Implementing this function is optional

execute#

execute is called in Process method of PythonExecutorCalculator which in turn is called by MediaPipe framework. How MediaPipe calls Process for the node depends on the configuration and the two configurations supported by PythonExecutorCalculator are:

Regular#

execute is called with a set of inputs and returns a set of outputs. For unary endpoints it’s the only possible configuration. On the implementation side, to use that mode, execute should return outputs.

def execute(self, inputs):
    ...
    return outputs

More information along with the configuration aspect described can be found in execution modes section.

Generative#

execute is called with a set of inputs and returns a generator. The generator is then called multiple times with no additional input data and produces multiple sets of outputs over time. Works only with streaming endpoints. On the implementation side, to use that mode, execute should yield outputs.

def execute(self, inputs):
    # For single set on inputs generate 10 sets of outputs
    for _ in range(10):
        ...
        yield outputs

More information along with the configuration aspect described can be found in execution modes section.

Parameters and return value#

execute is called with inputs parameter which is a list of pyovms.Tensor.

Depending on the mode it should return:

  • For regular mode: list of pyovms.Tensor

  • For generative mode: generator that generates list of pyovms.Tensor

So depending on the mode execute must always either return or yield a list of pyovms.Tensor

Returning multiple Python outputs from the graph

Note that this method returns outputs as a list, but since each output is a separate packet in MediaPipe flow, they do not arrive together to their destination. If the node outputs are also outputs from the graph the behavior differs depending on the kind of endpoint used:

  • For unary endpoints model server gathers all outputs from the graph and sends them all together in a single response

  • For streaming endpoints model server packs output and sends it in the response as soon as it arrives. It means that if execute returns a list of X outputs, the client will receive those outputs in X separate responses. The outputs can then be gathered using timestamp that can be found in received responses.

Error handling#

Signaling that something went wrong should be done by throwing an exception. The exception is caught by the PythonExecutorCalculator which logs it and returns non-OK status. Model Server then reads that status and sets graph in an error state. Then it closes all graph’s input streams and waits until in-progress actions are finished. Once it’s done the graph gets removed.

This behavior has different effect on the client depending on the kind of endpoint used - unary or streaming:

  • Unary

    With unary endpoint a graph is created, executed and destroyed for every request. If execute encounters an error, model server logs it and sends error message in response immediately.

  • Streaming

    With streaming endpoint a graph is created for the first request in the stream and then reused by all subsequent requests.

    If execute encounters an error on the first request (for example the Python code doesn’t work as expected), model server logs it and sends error message in response immediately. The graph gets destroyed.

    If execute encounters an error on one of the subsequent requests (for example wrong data has been received), model server logs it and MediaPipe sets error in the graph, but the client won’t receive error message until it sends another request. When the next request is read from the stream, model server checks if graph has an error, destroys it and sends response to the client. Rework of that behavior, so that error is being sent immediately is planned.

    As of now, graphs are not recoverable, so if execute encounters an error the stream gets closed and you need to create a new one.

Note: Run Model Server with --log_level DEBUG parameter to get information about errors in the server logs.

Implementing this function is required.

finalize#

finalize is called when model server unloads graph definition. It allows to perform any cleanup actions before the graph is removed.

Parameters and return value#

finalize does not have any parameters and is not expected to return any value.

Error handling#

Signaling that something went wrong should be done by throwing an exception. When model server catches exception from finalize it logs it and proceeds with the unload.

Note: Run Model Server with --log_level DEBUG parameter to get information about errors in the server logs.

Implementing this function is optional.

Python Tensor#

PythonExecutorCalculator operates on a dedicated Tensor class that wraps the data along with some additional information like name, shape or datatype. Objects of that class are passed to execute method as inputs and returned as output. They are also wrapped and exchanged between nodes in the graph and between graph and model server core.

This Tensor class is a C++ class with a Python binding that implements Python Buffer Protocol. It can be found in a built-in module pyovms.

Accessing Tensor Contents#

pyovms.Tensor attributes:

Name

Type

Description

name

string

Name of the string that also associates it with input or output stream of the node

shape

tuple

Tuple of numbers defining the shape of the tensor

datatype

string

Type of elements in the buffer.

data

memoryview

Memoryview of the underlying data buffer

size

number

Size of data buffer in bytes

Note: datatype attribute is not part of buffer protocol implementation. Buffer protocol uses format value that uses struct format characters. It can be read from data memoryview. There’s a mapping between those two - see datatype considerations.

As pyovms.Tensor implements buffer protocol it can be converted to another types that also implement buffer protocol:

def execute(self, inputs):
    input_tensor_bytes = bytes(inputs[0])
    ...
    import numpy as np
    input_tensor_ndarray = np.array(inputs[1])
    ...

Creating Output Tensors#

Inputs will be provided to the execute function, but outputs must be prepared by the user. Output objects can be created using pyovms.Tensor class constructor:

Tensor(name, data, shape=None, datatype=None)

  • name: a string that associates Tensor data with specific name. This name is also used by PythonExecutorCalculator to push data to the correct output stream in the node. More about it in node configuration section.

  • data: an object that implements Python Buffer Protocol. This could be an instance of some built-in type like bytes or types from external modules like numpy.ndarray.

  • shape (optional): a tuple or list defining the shape of the data. This value is directly assigned to shape attribute of the Tensor. By default, shape attribute is inherited from the data object. Providing shape to the constructor will override inherited value, so use it only if you know what you are doing.

  • datatype (optional): a string defining the type of the data. This value is directly assigned to datatype attribute of the Tensor. By default, datatype attribute is inherited from the data object. Providing datatype to the constructor will override inherited value, so use it only if you know what you are doing.

Note: shape and datatype arguments do not modify internal structure of the data - there are no reshapes and type conversions. They only override Tensor.shape and Tensor.datatype attributes, so the user can provide custom context to the next node or server response. It means they can be completely detached from the data buffer properties and it’s user’s responsibility to correctly interpret these attributes while reading the Tensor in the next node or the server response on the client side.

import numpy as np
from pyovms import Tensor

class OvmsPythonModel:
    def execute(self, inputs):
        # Create Tensor called my_output1 with encoded text
        output1 = Tensor("my_output1", "some text".encode())

        # Create Tensor called my_output with batch of string in numpy array
        # with overriding tensor datatype to match numpy array dtype
        npy_arr = np.array(["my", "batched", "string", "output"])
        output2 = Tensor("my_output2", npy_arr, datatype=np_arr.dtype.str)

        # A list of Tensors is expected, even if there's only one output
        return [output1, output2]

As Tensor gets created from another type it adapts all fields required by the buffer protocol as its own. Depending on how Tensor is created shape or datatype may be overridden. If they are not provided Tensor will adapt another buffer shape as it’s own and will map it’s format to a datatype. Learn more in datatype considerations section.

If the node is connected to another Python node, then Tensors pushed to the output of this node, are inputs of another node.

Datatype considerations#

There are two places where pyovms.Tensor objects are created and accessed:

  • in execute method of OvmsPythonModel class

  • in model server core during serialization and deserialization if Python node inputs or outputs as also graph inputs or outputs

Model Server receives requests and sends responses on interface via KServe API which defines expected data types for tensors. On the other hand Python Buffer Protocol requires format to be specified as struct format characters.

In order to let users work with KServe types without enforcing the usage of struct format characters on the client side, model server attempts to do the mapping as follows when creating Tensor objects from the request:

KServe Type

Format Character

BOOL

?

UINT8

B

UINT16

H

UINT32

I

UINT64

Q

INT8

b

INT16

h

INT32

i

INT64

q

FP16

e

FP32

f

FP64

d

The same mapping is applied the other way around when creating Tensor from another Python object in execute method (unless datatype argument is provided to the constructor ).

Tensor object always holds both values in Tensor.datatype and Tensor.data.format attributes so they can be used in deserialization and serialization, but also in another node in the graph.

In some cases, users may work with more complex types that are not listed above and model server also allows that.

BYTES datatype#

If datatype “BYTES” is specified and data is located in bytes_contents field of input(for gRPC) or in JSON body(for REST) OVMS converts it to pyovms.Tensor buffer according to the format where every input is preceded by four bytes of its size.

For example this gRPC request: bytes_content: [<240 byte element>, <1024 byte element>, <567 byte element>]

would be converted to this pyovms.Tensor.data contents: | 240 | < first element> | 1024 | | 567 | |

Custom types#

The datatype field in the tensor is a string and model server will not reject datatype that is not among above KServe types. If some custom type is defined in the request and server cannot map it to a format character it will translate it to B treating it as a 1-D raw binary buffer. For consistency the shape of the underlying buffer in that case will also differ from the shape defined in the request. Let’s see it on an example:

  1. Model Server receives request with the following input:

    • datatype: “my_string”

    • shape: (3,)

    • data: “null0terminated0string0” string encoded in UTF-8

  2. Model Server creates pyovms.Tensor with:

    • Tensor.datatype: “my_string”

    • Tensor.shape: (3,)

    • Tensor.data.format: “B”

    • Tensor.data.shape: (23,)

In execute user has access to both information from the request as well as how the internal buffer looks like.

pyovms.Tensor objects produced inside execute inherit most of the fields from the objects they are created from and by default tensor will try to map buffer protocol format to datatype according to the mapping mentioned before.

If it fails, the datatype is set to format, so that if such tensor is the output tensor of the graph, client receives the most valuable information about the type of output data.

In case this approach is insufficient, user can manually set datatype attribute to more suitable one, by providing optional datatype argument to the Tensor constructor.

Configuration and deployment#

Python is enabled via MediaPipe by built-in PythonExecutorCalculator, therefore, in order to execute Python code in OVMS you need to create a graph with a node that uses this calculator.

The way the graph is configured has a huge impact on the whole deployment. It defines things like:

  • inputs and outputs of the graph

  • inputs and outputs of each node in the graph

  • connections between the nodes

  • graph and nodes options

  • input stream handlers (defines conditions that must be met to launch Process in the node)

PythonExecutorCalculator#

Main part of the configuration is the node setting. Python nodes should use PythonExecutorCalculator and must be named. See a basic example:

node {
  name: "python_node"
  calculator: "PythonExecutorCalculator"
  input_side_packet: "PYTHON_NODE_RESOURCES:py"
  input_stream: "INPUT:input"
  output_stream: "OUTPUT:output"
  node_options: {
    [type.googleapis.com / mediapipe.PythonExecutorCalculatorOptions]: {
      handler_path: "/ovms/workspace/model.py"
    }
  }
}

Let’s break it down:

  • name: the name by which the node will be identified in the model server. Every Python node in a graph must have a unique name.

  • calculator: indicates the calculator to be used in the node. Must be PythonExecutorCalculator.

  • input_side_packet: a shared data passed from the model server to the Python nodes. It allows to share OvmsPythonModel state between multiple graph instances. Must be PYTHON_NODE_RESOURCES:py.

  • input_stream: defines input in form [TAG]:[NAME]. MediaPipe allows configurations with indexes i.e. [TAG]:[INDEX]:[NAME], but PythonExecutorCalculator ignores it.

  • output_stream: defines output in form [TAG]:[NAME]. MediaPipe allows configurations with indexes i.e. [TAG]:[INDEX]:[NAME], but PythonExecutorCalculator ignores it.

  • handler_path: the only one options so far in PythonExecutorCalculator. It’s a path to the Python file with OvmsPythonModel implementation.

Input and output streams in Python code#

How node input and output streams are configured has direct impact on the names of pyovms.Tensor objects in execute method of OvmsPythonModel. In previous simple configuration there are:

input_stream: "INPUT:input"
output_stream: "OUTPUT:output"

Both input and output streams are constructed as [TAG]:[NAME]. So in this example there’s:

  • input with tag INPUT and name input

  • output with tag OUTPUT and name output

In the Python code you should always refer to the [NAME] part. So inside execute there would be:

from pyovms import Tensor

class OvmsPythonModel:
    def execute(self, inputs):
        my_input = inputs[0]
        my_input.name == "input" # true
        my_output = Tensor("output", "some text".encode())
        return [my_output]

Access inputs via index#

In basic configurations, when execute runs with all expected inputs the order of Tensors in inputs list is not random. When PythonExecutorCalculator iterates through input streams to create Tensors, the streams are sorted by their tags. That knowledge can be useful while writing execute method to directly access data from particular streams. See an example:

node {
  name: "python_node"
  calculator: "PythonExecutorCalculator"
  input_side_packet: "PYTHON_NODE_RESOURCES:py"
  input_stream: "B:b"
  input_stream: "A:a"
  input_stream: "C:c"
  output_stream: "OUTPUT:output"
  node_options: {
    [type.googleapis.com / mediapipe.PythonExecutorCalculatorOptions]: {
      handler_path: "/ovms/workspace/model.py"
    }
  }
}

In that case, inputs can be access like this:

from pyovms import Tensor

class OvmsPythonModel:
    def execute(self, inputs):
        a = inputs[0] # Tensor with name "a" from input stream with tag "A"
        b = inputs[1] # Tensor with name "b" from input stream with tag "B"
        c = inputs[2] # Tensor with name "c" from input stream with tag "C"
        ...

Note: Node configuration and execute implementation should always match. For example if the node is configured to work with incomplete inputs, then accessing Tensors via index will not be useful.

Graph input and output streams#

So far only node input and output streams have been mentioned, but the configuration also requires defining graph’s input and output streams. The rules are very similar to how it works on the node level, so the streams are described in form: [TAG]:[NAME], but there’s more to it.

On graph level the [TAG] helps model server in deserialization and serialization by providing information about the object type expected in the stream. Model server reads the tag and expects it to start with one of predefined prefixes. If graph input stream is connected to Python node the tag should begin with OVMS_PY_TENSOR, which tells the server that it should deserialize input from the request to pyovms.Tensor object.

There can’t be two or more the same tags among the input streams as well as there can’t be two or more the same tags among the output streams. In such cases, prefix must be followed by some unique string.

input_stream: "OVMS_PY_TENSOR_IMAGE:image"
input_stream: "OVMS_PY_TENSOR_TEXT:text"
output_stream: "OVMS_PY_TENSOR:output"

Note: The same rule applies to node input and output streams.

When it comes to the [NAME] part, it is used to connect graph inputs and output with the nodes. They are also the input and output names in server requests and responses.

Multiple nodes#

Here is what you should know if you want to have multiple Python nodes in the same graph:

  • Every Python node must have a unique name in graph scope

  • Every Python node has it’s own instance of OvmsPythonModel that is not shared even if two nodes have identical handler_path

  • Nodes based on PythonExecutorCalculator can be connected directly without need for converters

  • Nodes may reuse the same Python file, but every Python file used by the server must have a unique name, otherwise some nodes might not work as expected. For example: /ovms/workspace1/model.py and /ovms/workspace2/model.py will result in only one model.py effectively loaded (this is supposed to be changed in the future versions).

Basic example#

Let’s see a complete example of the configuration with three Python nodes set in sequence:

input_stream: "OVMS_PY_TENSOR:first_number"
output_stream: "OVMS_PY_TENSOR:last_number"

node {
  name: "first_python_node"
  calculator: "PythonExecutorCalculator"
  input_side_packet: "PYTHON_NODE_RESOURCES:py"
  input_stream: "INPUT:first_number"
  output_stream: "OUTPUT:second_number"
  node_options: {
    [type.googleapis.com / mediapipe.PythonExecutorCalculatorOptions]: {
      handler_path: "/ovms/workspace/incrementer.py"
    }
  }
}

node {
  name: "second_python_node"
  calculator: "PythonExecutorCalculator"
  input_side_packet: "PYTHON_NODE_RESOURCES:py"
  input_stream: "INPUT:second_number"
  output_stream: "OUTPUT:third_number"
  node_options: {
    [type.googleapis.com / mediapipe.PythonExecutorCalculatorOptions]: {
      handler_path: "/ovms/workspace/incrementer.py"
    }
  }
}

node {
  name: "third_python_node"
  calculator: "PythonExecutorCalculator"
  input_side_packet: "PYTHON_NODE_RESOURCES:py"
  input_stream: "INPUT:third_number"
  output_stream: "OUTPUT:last_number"
  node_options: {
    [type.googleapis.com / mediapipe.PythonExecutorCalculatorOptions]: {
      handler_path: "/ovms/workspace/incrementer.py"
    }
  }
}

In that example client will send an input called first_number and receive output called last_number. Since user has access to input and output names in the Python code, the code for incrementation can be generic and reused in all nodes.

incrementer.py

from pyovms import Tensor

def increment(input):
    # Some code for input incrementation
    ...
    return output

class OvmsPythonModel:
    # Assuming this code is used with nodes
    # that have single input and single output

    def initialize(self, kwargs):
        self.output_name = kwargs["output_names"][0]

    def execute(self, inputs):
        my_input = inputs[0]
        my_output = Tensor(self.output_name, increment(my_input))
        return [my_output]

Model Server configuration file#

Once Python code and the pbtxt file with graph configuration is ready the model server configuration is very simple and could look like this:

{
    "model_config_list": [],
    "mediapipe_config_list": [
        {
            "name":"python_graph",
            "graph_path":"/ovms/workspace/graph.pbtxt"
        }
    ]
}

Where name defines the name of the graph and graph_path contains the path to graph configuration file.

Client side considerations#

Inference API and available endpoints#

Since Python execution is supported via MediaPipe serving flow, it inherits it’s enhancements and limitations. First thing to note is that MediaPipe graphs are available only via KServe API

From the client perspective model server serves a graph and user interacts with a graph. Single node in the graph cannot be accessed from the outside.

For a graph client can:

  • request status

  • request metadata

  • request inference

Learn more about how MediaPipe flow works in OpenVINO Model Server

For inference, data can be send both via gRPC API and KServe API(only for unary calls). If the graph has a OvmsPyTensor output stream, then the data in the KServe response can be found in raw_output_contents field (even if data in the request has been placed in InferTensorContents).

The data passed in the request is accessible in execute method of the node connected to graph input via data attribute of pyovms.Tensor object. For data of type BYTES send in bytes_contents field of input(for gRPC) or in JSON body(for REST) OVMS converts it to pyovms.Tensor buffer according to the format where every input is preceded by four bytes of its size.

Inputs and outputs also define shape and datatype parameters. Those values are also accessible in pyovms.Tensor. For outputs, datatype and shape are by default read from the underlying buffer, but it is possible to overwrite them (see pyovms.Tensor constructor. If you specify datatype as BYTES in your requests, make sure to review datatype considerations, since this type is treated differently than the others.

Let’s see it on an example:

# client.py

import tritonclient.grpc as grpcclient
...
client = grpcclient.InferenceServerClient("localhost:9000")
inputs = []
with open("image_path", 'rb') as f:
    image_data = f.read()
image_input = grpcclient.InferInput("image", [len(image_data)], "BYTES")
image_input._raw_content = image_data

text_encoded = "some text".encode()
text_input = grpcclient.InferInput("text", [len(text_encoded)], "BYTES")
text_input._raw_content = text_encoded

numpy_array = np.array([[1.0, 2.0, 3.0], [4.0, 5.0, 6.0]])
numpy_input = grpcclient.InferInput("numpy", numpy_array.shape, "FP32")
numpy_input.set_data_from_numpy(numpy_array)

results = client.infer("model_name", [image_input, text_input, numpy_input])
# model.py

from pyovms import Tensor
from PIL import Image
import io
import numpy as np
...
class OvmsPythonModel:

    def execute(self, inputs):
        # Read inputs
        image_input = inputs[0]
        print(image_input.shape) # (<image_binary_size>, )
        print(image_input.datatype) # "BYTES"

        text_input = inputs[1]
        print(text_input.shape) # (<string_binary_size>, )
        print(text_input.datatype) # "BYTES"

        numpy_input = inputs[2]
        print(text_input.shape) # (2, 3)
        print(text_input.datatype) # "FP32"

        # Convert pyovms.Tensor objects to more useful formats

        # Pillow Image created from image bytes
        image = Image.open(io.BytesIO(bytes(image_data)))
        # Python string "some text"
        text = bytes(text_input).decode()
        # Numpy array with shape (2, 3) and dtype float32
        ndarray = np.array(numpy_input)
        ...

Timestamping#

Mediapipe graph works with packets and every packet has its timestamp. The timestamps of packets on all streams (both input and output) must be ascending.

When requesting inference, user can decide to use automatic timestamping, or send timestamps themself along with the request as OVMS_MP_TIMESTAMP parameter. Learn more about timestamping

When it comes to Python node PythonExecutorCalculator:

  • for regular execution mode simply propagates timestamp i.e. uses input timestamp as output timestamp.

  • for generative execution mode it saves timestamp of the input and sends first set of outputs downstream with this timestamp. Then timestamp gets incremented with each generation, so next sets of output packages have ascending timestamp.

Multiple generation cycles on a single graph instance

Keep in mind that node keeps the timestamp and overwrites it every time new input arrives. It means that if you want to run multiple generation cycles on a single graph instance you must use manual timestamping - next request timestamp must be larger than the one received in the last response.

Outputs synchronization in gRPC streaming#

Timestamping has a crucial role when synchronizing packets from different streams both on the inputs and outputs as well as inside the graph. MediaPipe provides outputs of the graph to the model server and what happens next depends on what endpoint is used:

  • on gRPC unary endpoints server waits for the packets from all required outputs and sends them in a single response.

  • on gRPC streaming endpoints server serializes output packets as soon as they arrive and sends them back in separate responses.

It means that if you have a graph that has two or more outputs and use gRPC streaming endpoint you will have to take care of gathering the outputs. You can do that using OVMS_MP_TIMESTAMP.

timestamp = result.get_response().parameters["OVMS_MP_TIMESTAMP"].int64_param

Advanced Configuration#

Execution modes#

Python nodes can be configured to run in two execution modes - regular and generative.

In regular execution mode the node produces one set of outputs per one set of inputs. It works via both gRPC/REST unary and gRPC streaming endpoints and is a common mode for use cases like computer vision.

In generative execution mode the node produces multiple sets of outputs over time per single set of inputs. It works only via gRPC streaming endpoints and is useful for use cases where total processing time is big and you want to return some intermediate results before the execution is completed. That mode is well suited to Large Language Models to serve them in a more interactive manner.

Depending on which mode is used, both the Python code and graph configuration must be in line.

Regular mode#

When using regular mode, the execute method in OvmsPythonModel class must return value.

from pyovms import Tensor
...
  def execute(self, inputs):
        ...
        my_output = Tensor("output", data)
        return [my_output]

When execute returns, the PythonExecutorCalculator grabs the outputs and pushes them down the graph. Node Process method is called once per inputs set. Such implementation can be paired with basic graph setting, like:

node {
  name: "python_node"
  calculator: "PythonExecutorCalculator"
  input_side_packet: "PYTHON_NODE_RESOURCES:py"
  input_stream: "INPUT:input"
  output_stream: "OUTPUT:output"
  node_options: {
    [type.googleapis.com / mediapipe.PythonExecutorCalculatorOptions]: {
      handler_path: "/ovms/workspace/model.py"
    }
  }
}

Generative mode#

When using generative mode, the execute method in OvmsPythonModel class must yield value.

from pyovms import Tensor
...
  def execute(self, inputs):
        ...
        for data in data_stream:
          my_output = Tensor("output", data)
          yield [my_output]

When execute yields, the PythonExecutorCalculator saves the generator. Then it repeatedly calls it until it reaches the end of generated sequence. Node Process method is called multiple times per single inputs set. To trigger such behavior a specific graph configuration is needed. See below:

node {
  name: "python_node"
  calculator: "PythonExecutorCalculator"
  input_side_packet: "PYTHON_NODE_RESOURCES:py"
  input_stream: "INPUT:input"
  input_stream: "LOOPBACK:loopback"
  input_stream_info: {
    tag_index: 'LOOPBACK:0',
    back_edge: true
  }
  input_stream_handler {
    input_stream_handler: "SyncSetInputStreamHandler",
    options {
      [mediapipe.SyncSetInputStreamHandlerOptions.ext] {
        sync_set {
          tag_index: "LOOPBACK:0"
        }
      }
    }
  }
  output_stream: "OUTPUT:output"
  output_stream: "LOOPBACK:loopback"
  node_options: {
    [type.googleapis.com / mediapipe.PythonExecutorCalculatorOptions]: {
      handler_path: "/ovms/workspace/model.py"
    }
  }
}

Apart from basic configuration present also in regular mode, this graph contains some additional content. Let’s review it.

  1. LOOPBACK input and output stream

    input_stream: "LOOPBACK:loopback"
    ...
    output_stream: "LOOPBACK:loopback"
    

    This set of additional input and output stream enables internal cycle inside the node. It is used to trigger Process calls without any incoming packets and therefore call the generator without new data. The value in both input and output stream must be exactly the same and the PythonExecutorCalculator always expects the tag to be LOOPBACK.

    LOOPBACK input is not passed to execute method and user does not interact with it in any way.

  2. Back Edge Annotation

    input_stream_info: {
      tag_index: 'LOOPBACK:0',
      back_edge: true
    }
    

    This part says that the input stream with tag LOOPBACK and index 0 is used to create a cycle. If there are more than one index for LOOPBACK tag, the PythonExecutorCalculator will ignore it.

  3. SyncSetInputStreamHandler

    input_stream_handler {
      input_stream_handler: "SyncSetInputStreamHandler",
      options {
        [mediapipe.SyncSetInputStreamHandlerOptions.ext] {
          sync_set {
            tag_index: "LOOPBACK:0"
          }
        }
      }
    }
    

    In regular configuration DefaultInputStreamHandler is used by default, but for generative mode it’s not sufficient. When default handler is defined, node waits for all input streams before calling Process. In generative mode Process should be called once for data coming from the graph and then multiple times only by receiving signal on LOOPBACK, but inputs from a graph and LOOPBACK will never be present at the same time.

    For generative mode to work, inputs from the graph and LOOPBACK must be decoupled, meaning Process can be called with a set of inputs from the graph, but also with just LOOPBACK. It can be achieved via SyncSetInputStreamHandler. Above configuration sample creates a set with LOOPBACK, which also, implicitly creates another set, with all remaining inputs. Effectively there are two sets that do not depend on each other:

    • LOOPBACK

    • … every other input specified by the user.

It’s recommended not to reuse the same graph instance when the cycle is finished. Instead, if you want to generate for new data, create new gRPC stream.

For working configurations and code samples see the demos.

Incomplete inputs#

There are usecases when firing Process with only a subset of inputs defined in node configuration is desired. By default, node waits for all inputs with the same timestamp and launches Process once they’re all available. Such behavior is implemented by the DefaultInputStreamHandler which is used by default. To configure the node to launch Process with only a subset of inputs you should use a different input stream handler for different input policy.

Such configuration is used in generative execution mode, but let’s see another example:

node {
  name: "python_node"
  calculator: "PythonExecutorCalculator"
  input_side_packet: "PYTHON_NODE_RESOURCES:py"
  input_stream: "INPUT1:labels"
  input_stream: "INPUT2:image"
  input_stream_handler {
    input_stream_handler: "ImmediateInputStreamHandler",
  }
  output_stream: "OUTPUT:result"
  node_options: {
    [type.googleapis.com / mediapipe.PythonExecutorCalculatorOptions]: {
      handler_path: "/ovms/workspace/model.py"
    }
  }
}

Node configured with ImmediateInputStreamHandler will launch Process when any input arrives (no synchronization at all). Such configuration must be in line with the OvmsPythonModel class implementation. For example:

from pyovms import Tensor

class OvmsPythonModel:

    def initialize(self, kwargs: dict):
        self.model = load_model(...)
        self.labels = []

    def execute(self, inputs: list):
        outputs = []
        for input in inputs:
            if input.name == "labels":
                self.labels = prepare_new_labels(input)
            else: # the only other name is "image"
                output = self.model.process(input, self.labels)
                return [Tensor("result", output)]

In a scenario above the node runs some processing on the image with provided set of labels. When configuration allows for sending incomplete inputs, then the client can send labels only one time and then send only images.

Note: Keep in mind that members of OvmsPythonModel objects are shared between all graph instances. It means that if in above scenario one client in one graph changes labels, then that change is also effective in every other graph instance (for every other client that sends requests to that graph). Saving data between executions that will be exclusive to a single graph instance is planned to be supported in future versions.

Incomplete outputs#

PythonExecutorCalculator allows you to return incomplete set of outputs in execute method. It can be useful especially when working with streaming endpoints that serialize each graph output in a separate response. See an example:

node {
 name: "python_node"
 calculator: "PythonExecutorCalculator"
 input_side_packet: "PYTHON_NODE_RESOURCES:py"
 input_stream: "INPUT:input"
 output_stream: "OUTPUT:result"
 output_stream: "ERROR:error_message"
 node_options: {
   [type.googleapis.com / mediapipe.PythonExecutorCalculatorOptions]: {
     handler_path: "/ovms/workspace/model.py"
   }
 }
}

Python code that would run such node could look like this:

from pyovms import Tensor

class OvmsPythonModel:

    def initialize(self, kwargs: dict):
        self.model = load_model(...)

    def execute(self, inputs: list):
        input = inputs[0]
        try:
            output = self.model(input)
        except Exception:
          return [Tensor("error_message", "Error occurred during execution".encode())]
        return [Tensor("result", output)]

In such case, the client could implement different actions depending on which output it receives on the stream.

Another example of such configuration is signaling that generation is finished when running in generative mode. This solution is used in image generation demo.

Calculator type conversions#

Python nodes work with a dedicated Python Tensor objects that can be used both on C++ and Python side. The downside of that approach is that usually other calculators cannot read and create such objects. It means that Python nodes cannot be directly connected to any other, non-Python nodes.

That’s why converter calculators exists. They work as adapters between nodes and implement necessary conversions needed to create a connection between calculators that work on two different types of packets.

PyTensorOvTensorConverterCalculator#

OpenVINO Model Server comes with a built-in PyTensorOvTensorConverterCalculator that provides conversion between Python Tensor and OV Tensor.

Currently PyTensorOvTensorConverterCalculator works with only one input and one output.

  • The stream that expects Python Tensor must have tag OVMS_PY_TENSOR

  • The stream that expects OV Tensor must have tag OVTENSOR

In future versions converter calculator will accept multiple inputs and produce multiple outputs, but for now the only correct configuration is with one input stream and one output stream. One of those stream must have tag OVMS_PY_TENSOR and the other OVTENSOR, depending on the conversion direction.

PyTensorOvTensorConverterCalculator can also be configured to use node options with tag_to_output_tensor_names map and it’s used in OV Tensor to Python Tensor conversion. It defines the name Python Tensor should be created with, based on output stream tag.

See a simplified example with both conversions taking place in the graph:

input_stream: "OVMS_PY_TENSOR:input"
output_stream: "OVMS_PY_TENSOR:output"

node {
  name: "PythonPreprocess"
  calculator: "PythonExecutorCalculator"
  input_side_packet: "PYTHON_NODE_RESOURCES:py"
  input_stream: "INPUT:input"
  output_stream: "OUTPUT:preprocessed_py"
  node_options: {
    [type.googleapis.com/mediapipe.PythonExecutorCalculatorOptions]: {
      handler_path: "/ovms/workspace/preprocess.py"
    }
  }
}

node {
  calculator: "PyTensorOvTensorConverterCalculator"
  input_stream: "OVMS_PY_TENSOR:preprocessed_py"
  output_stream: "OVTENSOR:preprocessed_ov"
}

node {
  calculator: "OpenVINOInferenceCalculator"
  input_side_packet: "SESSION:session" # inference session
  input_stream: "OVTENSOR:preprocessed_ov"
  output_stream: "OVTENSOR:result_ov"
}

node {
  calculator: "PyTensorOvTensorConverterCalculator"
  input_stream: "OVTENSOR:result_ov"
  output_stream: "OVMS_PY_TENSOR:result_py"
  node_options: {
    [type.googleapis.com/mediapipe.PyTensorOvTensorConverterCalculatorOptions]: {
      tag_to_output_tensor_names {
        key: "OVMS_PY_TENSOR"
        value: "result_py"
      }
    }
  }
}

node {
  name: "PythonPostprocess"
  calculator: "PythonExecutorCalculator"
  input_side_packet: "PYTHON_NODE_RESOURCES:py"
  input_stream: "INPUT:result_py"
  output_stream: "OUTPUT:output"
  node_options: {
    [type.googleapis.com/mediapipe.PythonExecutorCalculatorOptions]: {
      handler_path: "/ovms/workspace/postprocess.py"
    }
  }
}

See a CLIP demo for a complete example of a graph that uses Python nodes, OV Inference nodes and converter nodes.