MLEM, developed by iterative.ai - is an open-source tool that allows users to simplify machine learning model packaging and deployment. With MLEM, users can run their machine learning models anywhere, by wrapping models as a Python package or Docker Image, or deploying them to Heroku (SageMaker, Kubernetes, and more platforms coming soon). MLEM is written in Python, and runs as a microservice with applications calling for a simple HTTP API. This may concern some users who did not develop their applications in Python, which is why we decided to take a closer look at using models on different platforms.
This .NET HTTP client is designed to help software developers use a machine learning approach to create intelligent applications, based on .NET, without needing a deep knowledge of MLEM.This client is a library that allow users to connect with MLEM API simply and easily, as well as integrate the MLEM model functionality to their .NET project. It also supports request body validation by schema received from the MLEM server, and also offers a standard logging interface to use.
The client provides several methods for using MLEM technologies with given rules. The main of them are PredictAsync and CallAsync.
PredictAsyncmakes a request for the "predict" methodCallAsyncmakes a request for any methods (including "predict")
This is the core functionality of MLEM client for .NET apps. Having a stable application with minimum functionality is a good way to support it now and make for easier improvements in the future.
Before using the client make sure that you have a deployed mlem model (local or remote). Read MLEM docs to know how to deploy a model. A list of sample models you can find below. Also, clone the repository and build the MlemApi project.
After you have a link to a deployed model, prepare your application that will use MLEM capabilities. Refer it to the built MlemApi.dll.
There is Example project in this perository as an usage example, where you can see several cases (see Utilities/TestCases.cs).
Next, a sample case with the Iris model is described.
Define a request object class
Create a C# class according to the deployed model.
using Newtonsoft.Json;
public class Iris
{
[JsonProperty("sepal length (cm)")]
public double SepalLength { get; set; }
[JsonProperty("sepal width (cm)")]
public double SepalWidth { get; set; }
[JsonProperty("petal length (cm)")]
public double PetalLength { get; set; }
[JsonProperty("petal width (cm)")]
public double PetalWidth { get; set; }
}Create the client
The simplest overload of the client requests a link to deployed model.
MlemApiClient mlemClient = new("https://example-mlem-get-started-app.herokuapp.com");Define the input data
You can use an object instance for a request or a list of the objects
List<Iris> inputData = new
{
new Iris
{
SepalLength = 5.1,
SepalWidth = 3.5,
PetalLength = 1.4,
PetalWidth = 0.2
},
new Iris
{
SepalLength = 7.6,
SepalWidth = 3.0,
PetalLength = 6.6,
}
};Send a request
PredictAsync
List<long> result = await mlemClient.PredictAsync<List<long>, Iris>(inputData);or CallAsync
List<List<double>> result = await mlemClient.CallAsync<List<List<double>>, Iris>(
"predict_proba",
inputData);Mlem client provides validatiion functionality for request objects and response based on api schema of the deployed mlem model. Since the mlem server does not provide a verbosed error in case of data format error - this feature can be useful to catch data format issues easier.
While request object validation allows you to double-check that input data was provided properly to the client, the response validation helps to check that a mlem model returns a response according to the schema.
For the request validation it is required to pass a map from model column names to request data class property names (modelColumnToPropNamesMap), so the mlem client will be able to associate relevant data fields from a request object and schema.
You can turn on/off this feature using ArgumentTypesValidationIsOn and ResponseValidationIsOn properties of the mlem client for request and response objects validation respectively (turned off by default).
See more examples of validation cases/input data prepared in the Example project.
There is an upcoming feature for classes generation - for request and response data.
You can use GenerateClasses of ModelClassesGenerator classes to generate request and response data classes - to use it further in your add as data contracts.
See the Example project for the example of usage (TestCases.ClassGeneration test case).
There are the following sample models, that can be used for deployment and requests testing.
- Digits
- Iris
- Wine
They are in ModelRepository project in Additional models folder.
They are built using LearnModelScript.py scripts for each model.
Mlem client support the following types:
- Pandas types: dataframe
- Numpy types: ndarray
- List type
- Primitive types: float64, int (int8, int16,int32), uint (uint8, uint16, uint32, uint64), bool, str
Please note that mlem already sorts dataframe fields in right order - so there is no need to worry about columns order for this datatype (treat it just like a dictionary type).
See Example_CSharp project for more examples.
Some resources should be the same in .NET and Java clients. So, this reprository uses git submodules to get files like Error_messages.json and Log_messages.json from another private reporsitory - see README.md file in [root]\ResourcesGenerator\CommonResources for more details.
This file is located in [root]\ResourcesGenerator\CommonResources folder, that is a clone of that private repository in fact. So, this folder can contain any shared files defined there.
Regarding to this .NET repository, the client doesn't use these .json files directly. For example, Error_messages.json file converts into [root]\MlemApi\Resources\EM.cs file via [root]\Update_resources.cmd script; the same for Log_messages.json file and [root]\MlemApi\Resources\LM.cs. After cloning of the repository you don't need to do anything additional to get the resources. They are already defined in EM.cs and LM.cs file.
To update resources do the following:
- Open
[root]\ResourcesGenerator\CommonResources\Error_messages.jsonfile (orLog_messages.json) - Make some changes, it will affect the both .NET and Java clients
- Open
[root]\ResourcesGenerator\CommonResourcesfolder via git, here is the submodule local repository - Make a commit and push it to the
mainbranch (make sure you have necessary permissions) - Next you need to update submodule dependencies. Just run
[root]\Update_resources.cmdscript. It will update[root]\MlemApi\Resources\EM.csand[root]\MlemApi\Resources\LM.csfile and add submodule changes to the git index of the current .NET repository - Commit and push the new changes
MLEM makes the process of packaging and deployment of machine learning models much easier. .NET client developed by Akvelon make it possible to integrate MLEM models to non-Python projects. Use the client with your existing or new applications: Web (ASP.Net), Mobile (Xamarin), Desktop (WPF, WinForms). Forget about the need to create spaghetti code and get access to the advantages of MLEM’s features in your apps with the http clients developed by Akvelon's engineers!