Keras Convert and Upload Within Wallaroo

How to convert Keras ML models and upload them to Wallaroo using the Wallaroo auto-conversion method.


Machine Learning (ML) models can be converted into a Wallaroo Model object and uploaded into Wallaroo workspace using the Wallaroo Client convert_model(path, source_type, conversion_arguments) method. This conversion process transforms the model into an open format that can be run across different frameworks at compiled C-language speeds.

The following tutorial is a brief example of how to convert a Keras or Tensor ML model to ONNX. This allows organizations that have trained Keras or Tensor models to convert them and use them with Wallaroo.

This tutorial assumes that you have a Wallaroo instance and are running this Notebook from the Wallaroo Jupyter Hub service.

This tutorial demonstrates how to:

  • Convert a keras ML model and upload it into the Wallaroo engine.
  • Run a sample inference on the converted model in a Wallaroo instance.

This tutorial provides the following:

  • A pre-trained keras sentiment model to be converted. This has 100 columns.

Conversion Steps

To use the Wallaroo autoconverter convert_model(path, source_type, conversion_arguments) method takes 3 parameters. The parameters for keras conversions are:

  • path (STRING): The path to the ML model file.
  • source_type (ModelConversionSource): The type of ML model to be converted. As of this time Wallaroo auto-conversion supports the following source types and their associated ModelConversionSource:
    • sklearn: ModelConversionSource.SKLEARN
    • xgboost: ModelConversionSource.XGBOOST
    • keras: ModelConversionSource.KERAS
  • conversion_arguments: The arguments for the conversion based on the type of model being converted. These are:
    • wallaroo.ModelConversion.ConvertKerasArguments: Used for converting keras type models and takes the following parameters:
      • name: The name of the model being converted.
      • comment: Any comments for the model.
      • input_type: A tensorflow Dtype called in the format ModelConversionInputType.{type}, where {type} is Float, Double, etc depending on the model.
      • dimensions: Corresponds to the keras xtrain in the format List[Union[None, int, float]].

Import Libraries

The first step is to import the libraries needed.

import wallaroo

from wallaroo.ModelConversion import ConvertKerasArguments, ModelConversionSource, ModelConversionInputType
from wallaroo.object import EntityNotFoundError

Configuration and Methods

The following will set the workspace, pipeline, model name, the model file name used when uploading and converting the keras model, and the sample data.

The functions get_workspace(name) will either set the current workspace to the requested name, or create it if it does not exist. The function get_pipeline(name) will either set the pipeline used to the name requested, or create it in the current workspace if it does not exist.

workspace_name = 'keras-autoconvert-workspace'
pipeline_name = 'keras-autoconvert-pipeline'
model_name = 'simple-sentiment-model'
model_file_name = ''
sample_data = 'simple_sentiment_testdata.json'

def get_workspace(name):
    wl = wallaroo.Client()
    workspace = None
    for ws in wl.list_workspaces():
        if == name:
            workspace= ws
    if(workspace == None):
        workspace = wl.create_workspace(name)
    return workspace

def get_pipeline(name):
    wl = wallaroo.Client()
        pipeline = wl.pipelines_by_name(pipeline_name)[0]
    except EntityNotFoundError:
        pipeline = wl.build_pipeline(pipeline_name)
    return pipeline

Connect to Wallaroo

Connect to your Wallaroo instance and store the connection into the variable wl.

wl = wallaroo.Client()

Set the Workspace and Pipeline

Set or create the workspace and pipeline based on the names configured earlier.

workspace = get_workspace(workspace_name)


pipeline = get_pipeline(pipeline_name)
name keras-autoconvert-pipeline
created 2022-07-07 16:27:57.437207+00:00
last_updated 2022-07-07 16:28:42.403022+00:00
deployed False
steps simple-sentiment-model

Set the Model Autoconvert Parameters

Set the paramters for converting the simple-sentiment-model. This includes the shape of the model.

model_columns = 100

model_conversion_args = ConvertKerasArguments(
    comment="simple keras model",
    dimensions=(None, model_columns)
model_conversion_type = ModelConversionSource.KERAS

Upload and Convert the Model

Now we can upload the convert the model. Once finished, it will be stored as {unique-file-id}-converted.onnx.

Converted Model
# converts and uploads model.
model_wl = wl.convert_model('', model_conversion_type, model_conversion_args)
{'name': 'simple-sentiment-model', 'version': 'c76870f8-e16b-4534-bb17-e18a3e3806d5', 'file_name': '14d9ab8d-47f4-4557-82a7-6b26cb67ab05-converted.onnx', 'last_update_time': datetime.datetime(2022, 7, 7, 16, 41, 22, 528430, tzinfo=tzutc())}

Test Inference

With the model uploaded and converted, we can run a sample inference.

Add Pipeline Step and Deploy

We will add the model as a step into our pipeline, then deploy it.

Waiting for deployment - this will take up to 45s .... ok
name keras-autoconvert-pipeline
created 2022-07-07 16:27:57.437207+00:00
last_updated 2022-07-07 16:41:23.615423+00:00
deployed True
steps simple-sentiment-model

Run a Test Inference

We can run a test inference from the simple_sentiment_testdata.json file, then display just the results.

sample_data = 'simple_sentiment_testdata.json'
result = pipeline.infer_from_file(sample_data)
Waiting for inference response - this will take up to 45s .... ok
        [0.9964503 ]])]

Undeploy the Pipeline

With the tests complete, we will undeploy the pipeline to return the resources back to the Wallaroo instance.

Waiting for undeployment - this will take up to 45s ................................... ok
name keras-autoconvert-pipeline
created 2022-07-07 16:27:57.437207+00:00
last_updated 2022-07-07 16:57:06.402657+00:00
deployed False
steps simple-sentiment-model