Windows 10: Announcing ML.NET 0.9 – Machine Learning for .NET

Brink · Jan 10, 2019

ML.NET is an open-source and cross-platform machine learning framework (Windows, Linux, macOS) for .NET developers. Using ML.NET, developers can leverage their existing tools and skillsets to develop and infuse custom AI into their applications by creating custom machine learning models.

ML.NET allows you to create and use machine learning models targeting common tasks such as classification, regression, clustering, ranking, recommendations and anomaly detection. It also supports the broader open source ecosystem by proving integration with popular deep-learning frameworks like TensorFlow and interoperability through ONNX. Some common use cases of ML.NET are scenarios like Sentiment Analysis, Recommendations, Image Classification, Sales Forecast, etc. Please see our samples for more scenarios.

Today we’re happy to announce the release of ML.NET 0.9. (ML.NET 0.1 was released at //Build 2018). This release focuses on: API improvements, model explainability and feature contribution, support for GPU when scoring ONNX models and significant clean up of the framework internals.

This blog post provides details about the following topics in the ML.NET 0.9 release:

Feature Contribution Calculation (FCC) and other Model Explainability improvements

Added GPU support for ONNX Transform

New Visual Studio ML.NET project templates preview

Additional API improvements in ML.NET 0.9

Feature Contribution Calculation and other model explainability improvements

Feature Contribution Calculation (FCC)

The Feature Contribution Calculation (FCC for short) shows which features are most influential for a model’s prediction on a particular and individual data sample by determining the amount each feature contributed to the model’s score for that particular data sample.

FCC is particulary important when you initialy have a lot of features/attributes in your historic data and you want to select and use only the most important features because using too many features (especially if including features that don’t influence the model) can reduce the model’s performance and accuracy. Therefore, with FCC you can identify the most influential positive and negative contributions from the initial attribute set.

You can use FCC to produce feature contributions with code like the following:

Code: // Create a Feature Contribution Calculator // Calculate the feature contributions for all features given trained model parameters var featureContributionCalculator = mlContext.Model.Explainability.FeatureContributionCalculation(model.Model, model.FeatureColumn, numPositiveContributions: 11, normalize: false); // FeatureContributionCalculatingEstimator can be use as an intermediary step in a pipeline. // The features retained by FeatureContributionCalculatingEstimator will be in the FeatureContribution column. var pipeline = mlContext.Model.Explainability.FeatureContributionCalculation(model.Model, model.FeatureColumn, numPositiveContributions: 11) .Append(mlContext.Regression.Trainers.OrdinaryLeastSquares(featureColumn: "FeatureContributions"));
Click to expand...

The output of the above code is:

Label Score BiggestFeature Value Weight Contribution

24.00 27.74 RoomsPerDwelling 6.58 98.55 39.95
21.60 23.85 RoomsPerDwelling 6.42 98.55 39.01
34.70 29.29 RoomsPerDwelling 7.19 98.55 43.65
33.40 27.17 RoomsPerDwelling 7.00 98.55 42.52

FCC can be used as a step in the ML pipeline and complements the current explainability tools in ML.NET like Permutation Feature Importance (PFI). With ML.NET 0.8, we already provided initial APIs for model explainability to help machine learning developers better understand the feature importance of models (“Overall Feature Importance”) and create (“Generalized Additive Models”)

Sample for FCC.

Additional model explainability improvements for features selection

In addition to FCC, we also extended the capabilities of Permutation Feature Importance (PFI) and Generalized Additive Models (GAMs):

PFI now supports most learning tasks: Regression, Binary Classification, Multiclass Classification, and Ranking.

PFI now allows you to calculate confidence intervals on feature importance scores to allow you to get a better estimate of the mean.

GAMs now supports Feature Contribution Calculation (FCC) so you can quickly see which features drove an individual prediction.

Sample for PFI.

Added GPU support for ONNX Transform

In ML.NET 0.9 we added the capability to score/run ONNX models using CUDA 10.0 enabled GPUs (such as most NVIDIA GPUs), by integrating the high performance ONNX Runtime library. GPU support for ONNX models is currently available only on Windows 64-bit (not x86,yet), with Linux and Mac support coming soon. Learn here about supported ONNX/CUDA formats/version.

Sample code plus a Test here.

New Visual Studio ML.NET project templates preview

We are pleased to announce a preview of Visual Studio project templates for ML.NET. These templates make it very easy to get started with machine learning. You can download these templates from Visual Studio gallery here.

The templates cover the following scenarios:

ML.NET Console Application – Sample app that demonstrates how you can use a machine learning model in your application.

ML.NET Model Library – Creates a new machine learning model library which you can consume from within your application.

Additional API improvements in ML.NET 0.9

In this release we have also added other enhancements to our APIs such as the following.

Text data loading is simplified

In ML.NET 0.9, when using the TextLoader class you can either directly provide the attributes/columns in the file as you were able to do it in previous versions or as a new improvement and optional choice you can instead specify those columns/attributes through a data-model class.

Before ML.NET v0.9 you always needed to have explicit code like the following:

Code: // //... Your code... var mlContext = new MLContext(); // Create the reader: define the data columns and where to find them in the text file. var reader = mlContext.Data.CreateTextReader(new[] { new TextLoader.Column("IsOver50K", DataKind.BL, 0), new TextLoader.Column("Workclass", DataKind.TX, 1) },hasHeader: true ); var dataView = reader.Read(dataPath);[/quote] With 0.9, you can simply load the type as follows.

Code: // //... Your code in your class... var mlContext = new MLContext(); // Read the data into a data view. var dataView = mlContext.Data.ReadFromTextFile<InspectedRow>(dataPath, hasHeader: true); // The data model. This type will be used from multiple code. private class InspectedRow { [LoadColumn(0)] public bool IsOver50K { get; set; } [LoadColumn(1)] public string Workclass { get; set; } }[/quote] Sample code.

Get prediction confidence factor

With Calibrator Estimators, in addition to the score column you can get when evaluating the quality of your model you can now get a probability column as well (probability of this example being on the predicted class; prediction confidence indicator).

For instance, you could get a list of the probabilities per each predicted value, like in the following list:

Score - 0.458968 Probability 0.4670409
Score - 0.7022135 Probability 0.3912723
Score 1.138822 Probability 0.8703266

Sample code

New Key-Value mapping estimator and transform

This feature replaces the TermLookupTransform and provides a way to specify the mapping betweeen two values (note this is specified and not trained). You can specify the mapping by providing a keys list and values list that must be equal in size.

Sample code

Other improvements and changes

Allow ML.NET to run on Windows Nano containers or Windows machines without Visual C++ runtime installed.

Metadata Support In DataView Construction with information about the model, like the evaluation metrics which is encoded metadata into the model and can be programatically extracted and therefore visualized in any tool. This feature can be useful for ISVs.

For a with list of breaking changes in v0.9 that impacted the ML.NET samples, check this Gist here

Moving forward

While on the past 9 months we have been adding new features and improving ML.NET, in the forthcoming 0.10, 0.11 and upcoming releases before we reach v1.0, we will focus on the overall stability of the package, continue to refine the API, increase test coverage and improve documentation and samples.

Provide your feedback through the new ML.NET survey!

ML.NET is new, and as we are developing it, we would love to get your feedback! Please fill out the brief survey below and help shape the future of ML.NET by telling us about your usage and interest in Machine Learning and ML.NET.

Take the survey now!

Get started!

If you haven’t already get started with ML.NET here.

Next, going further explore some other resources:

Tutorials and resources at the Microsoft Docs ML.NET Guide

Code samples at the machinelearning-samples GitHub repo

Important ML.NET concepts for understanding the new API are introduced here

“How to” guides that show how to use these APIs for a variety of scenarios can be found here

We will appreciate your feedback by filing issues with any suggestions or enhancements in the ML.NET GitHub repo to help us shape ML.NET and make .NET a great platform of choice for Machine Learning.

Thanks and happy coding with ML.NET!

The ML.NET Team.

This blog was authored by Cesar de la Torre and Pranav Rastogi plus additional contributions from the ML.NET team

[/quote]
Source: Announcing ML.NET 0.9 – Machine Learning for .NET | .NET Blog

:)

Brink · Jan 10, 2019

Announcing ML.NET 0.8 - Machine Learning for .NET

ML.NET is an open-source and cross-platform framework (Windows, Linux, macOS) which makes machine learning accessible for .NET developers.

ML.NET allows you to create and use machine learning models targeting scenarios to achieve common tasks such as sentiment analysis, issue classification, forecasting, recommendations, fraud detection, image classification and more. You can check out these common tasks at our GitHub repo with ML.NET samples.

Today we’re happy to announce the release of ML.NET 0.8. (ML.NET 0.1 was released at //Build 2018). This release focuses on adding improved support for recommendation scenarios, model explainability in the form of feature importance, debuggability by previewing your in-memory datasets, API improvements such as caching, filtering, and more.

This blog post provides details about the following topics in the ML.NET 0.8 release:

New Recommendation scenarios

Improved debuggability

Model explainability

Additional API improvements

Click to expand...

Read more: Announcing ML.NET 0.8 Machine Learning for .NET | .NET Blog

Tweet

— Twitter API (@User) View on Twiiter

Brink · Jan 10, 2019

Announcing ML.NET 0.7 (Machine Learning .NET)

We’re excited to announce today the release of ML.NET 0.7 – the latest release of the cross-platform and open source machine learning framework for .NET developers (ML.NET 0.1 was released at //Build 2018). This release focuses on enabling better support for recommendation based ML tasks, enabling anomaly detection, enhancing the customizability of the machine learning pipelines, enabling using ML.NET in x86 apps, and more.

This blog post provides details about the following topics in the ML.NET 0.7 release:

Enhanced support for recommendation tasks with Matrix Factorization

Enabled anomaly detection scenarios – detecting unusual events rel

Improved customizability of ML pipelines

x86 support

NimbusML – experimental Python bindings for ML.NET

Get started with ML.NET v07

Enhanced support for recommendation tasks with Matrix Factorization

Recommender systems enable producing a list of recommendations for products in a catalog, songs, movies, and more. We have improved support for creating recommender systems in ML.NET by adding Matrix factorization (MF), a common approach to recommendations when you have data on how users rated items in your catalog. For example, you might know how users rated some movies and want to recommend which other movies they are likely to watch next.

We added MF to ML.NET because it is often significantly faster than Field-Aware Factorization Machines (which we added in ML.NET 0.3) and it can support ratings which are continuous number ratings (e.g. 1-5 stars) instead of boolean values (“liked” or “didn’t like”). Even though we just added MF, you might still want to use FFM if you want to take advantage of other information beyond the rating a user assigns to an item (e.g. movie genre, movie release date, user profile). A more in-depth discussion of the differences can be found here.

Sample usage of MF can be found here. The example is general but you can imagine that the matrix rows correspond to users, matrix columns correspond to movies, and matrix values correspond to ratings. This matrix would be quite sparse as users have only rated a small subset of the catalog.

ML.NET’s MF uses LIBMF.

Enabled anomaly detection scenarios – detecting unusual events

Anomaly detection enables identifying unusual values or events. It is used in scenarios such as fraud detection (identifying suspicious credit card transactions) and server monitoring (identifying unusual activity).

ML.NET 0.7 enables detecting two types of anomalous behavior:

Spike detection: spikes are attributed to sudden yet temporary bursts in values of the input data. These could be outliers due to outages, cyber-attacks, viral web content, etc.

Change point detection: change points mark the beginning of more persistent deviations in the behavior of the data. For example, if product sales are relatively consistent and become more popular (monthly sales double), there is a change point when the trend changes.

These anomalies can be detected on two types of data using different ML.NET components:

IidSpikeDetector and IidChangePointDetector are used on data assumed to be from one stationary distribution (each data point is independent of previous data, such as the number of retweets of each tweet).

SsaSpikeDetector and SsaChangePointDetector are used on data that has a season/trend components (perhaps ordered by time, such as product sales)

Sample code using anomaly detection with ML.NET can be found here.

Improved customizability of ML.NET pipelines

ML.NET offers a variety of data transformations (e.g. processing text, images, categorical features, etc.). However, some use cases require application-specific transformations, such as calculating cosine similarity between two text columns. We have now added support for custom transforms so you can easily include custom business logic.

The CustomMappingEstimator allows you to write your own methods to process data and bring them into the ML.NET pipeline. Here is what it would look like in the pipeline:

var estimator = mlContext.Transforms.CustomMapping<MyInput, MyOutput>(MyLambda.MyAction, "MyLambda")
.Append(...)
.Append(...)

Below is the definition of what this custom mapping will do. In this example, we convert the text label (“spam” or “ham”) to a boolean label (true or false).

public class MyInput
{
public string Label { get; set; }
}
public class MyOutput
{
public bool Label { get; set; }
}
public class MyLambda
{
[Export("MyLambda")]
public ITransformer MyTransformer => ML.Transforms.CustomMappingTransformer<MyInput, MyOutput>(MyAction, "MyLambda");
[Import]
public MLContext ML { get; set; }
public static void MyAction(MyInput input, MyOutput output)
{
output.Label= input.Label == "spam" ? true : false;
}
}

A more complete example of the CustomMappingEstimator can be found here.

x86 support in addition to x64

With this release of ML.NET you can now train and use machine learning models on x86 / 32-bit architecture devices. Previously, ML.NET was limited to x64 devices.

Note that some components that are based on external dependencies (e.g. TensorFlow) will not be available in x86.

NimbusML – experimental Python bindings for ML.NET

NimbusML provides experimental Python bindings for ML.NET. We have seen feedback from the external community and internal teams regarding the use of multiple programming languages. We wanted to enable as many people as possible to benefit from ML.NET and help teams to work together more easily. ML.NET not only enables data scientists to train and use machine learning models in Python (with components that can also be used in scikit-learn pipelines), but it also enables saving models which can be easily used in .NET applications through ML.NET (see here for more details).

In case you missed it: provide your feedback on the new API

ML.NET 0.6 introduced a new set of APIs for ML.NET that provide enhanced flexibility. These APIs in 0.7 and upcoming versions are still evolving and we would love to get your feedback so you can help shape the long-term API for ML.NET.

Want to get involved? Start by providing feedback through issues at the ML.NET GitHub repo!

Additional resources

The most important ML.NET concepts for understanding the new API are introduced here.

A cookbook (How to guides) that shows how to use these APIs for a variety of existing and new scenarios can be found here.

A ML.NET API Reference with all the documented APIs can be found here.

Get started!

If you haven’t already, get started with ML.NET here Next, explore some other great resources:

Tutorials and resources at the Microsoft Docs ML.NET Guide

Code samples at the machinelearning-samples GitHub repo

We look forward to your feedback and welcome you to file issues with any suggestions or enhancements in the ML.NET GitHub repo.

This blog was authored by Gal Oshri and Cesar de la Torre

Thanks,

The ML.NET Team
Click to expand...

Source: Announcing ML.NET 0.7 (Machine Learning .NET) | .NET Blog

Brink · Jan 10, 2019

Announcing .NET Standard 2.0

The .NET Standard 2.0 specification is now complete. It is supported in .NET Core 2.0, in the .NET Framework 4.6.1 and later versions, and in Visual Studio 15.3. You can start using .NET Standard 2.0 today.

For the impatient: TL;DR

.NET Standard is for sharing code. .NET Standard is a set of APIs that all .NET implementations must provide to conform to the standard. This unifies the .NET implementations and prevents future fragmentation. It replaces Portable Class Libraries (PCLs) as the tool for building .NET libraries that work everywhere.

Much bigger API Surface: We have more than doubled the set of available APIs from 13k in .NET Standard 1.6 to 32k in .NET Standard 2.0. Most of them are existing .NET Framework APIs. These additions make it much easier to port existing code to .NET Standard, and, by extension, to any .NET implementation of .NET Standard, such as .NET Core 2.0 and the upcoming version of UWP.

.NET Framework compatibility mode: The vast majority of NuGet packages are currently still targeting .NET Framework. Many projects are currently blocked from moving to .NET Standard because not all their dependencies are targeting .NET Standard yet. That’s why we added a compatibility mode that allows .NET Standard projects to reference .NET Framework libraries. While this may not work in all cases (for instance, if the .NET Framework binaries use WPF) we found that with .NET Standard 2.0. So in practice it unblocks many projects.

Broad platform support. .NET Standard 2.0 is supported on the following platforms:

.NET Framework 4.6.1

.NET Core 2.0

Mono 5.4

Xamarin.iOS 10.14

Xamarin.Mac 3.8

Xamarin.Android 7.5

Upcoming version of UWP (expected to ship later this year)

Click to expand...

Read more: Announcing .NET Standard 2.0 | .NET Blog

Windows 10: Announcing ML.NET 0.9 – Machine Learning for .NET

Announcing ML.NET 0.9 – Machine Learning for .NET

Announcing ML.NET 0.9 – Machine Learning for .NET

Announcing ML.NET 0.9 – Machine Learning for .NET

Announcing ML.NET 0.9 – Machine Learning for .NET - Similar Threads - Announcing NET –

Announcing .NET 5.0

Announcing .NET Core 3.0 Preview 9

Announcing ML.NET 1.0 RC - Machine Learning for .NET

Announcing ML.NET 0.11 - Machine Learning for .NET

Announcing ML.NET 0.10 - Machine Learning for .NET

Announcing ML.NET 0.8 - Machine Learning for .NET

Announcing ML.NET 0.7 (Machine Learning .NET)

Announcing .NET Standard 2.0

Announcing .NET Framework 4.6