Machine Learning with .NET: Model Serialization

This is part five in a series on Machine Learning with .NET.

So far in this series, we’ve looked at training a model at runtime. Even in the prior post, where we looked at predictions, we first had to train a model. With a Naive Bayes model against a small data set, that’s not an onerous task—it’s maybe a second or two the first time we instantiate our objects and we can leave the prediction engine around for the lifetime of our application, so we are able to amortize the cost fairly effectively.

Suppose, however, that we have a mighty oak of a neural network which we have trained over the past two months. It finally completed training and we now have a set of weights. Obviously, we don’t want to retrain this if we can avoid it, so we need to find a way to serialize our model and save it to disk somewhere so that we can re-use it later.

Saving a Model to Disk

Saving a model to disk is quite easy. Here’s the method I created to do just that:

public void SaveModel(MLContext mlContext, ITransformer model, string modelPath)
	using (var stream = File.Create(modelPath))
		mlContext.Model.Save(model, null, stream);

We create a file stream and use the built-in Save() method to write our data to that stream. Here is an example of me calling that method:

string modelPath = "C:\\Temp\\BillsModel.mdl";
bmt.SaveModel(mlContext, model, modelPath);

This generates a binary .mdl file which contains enough information to reconstitute our model and generate predictions off of it.

Loading a Model from Disk

Here’s where things get annoying again. If you have any kind of custom mapping, you need to do a bit of extra work, as I mentioned in my rant.

Next, when it comes time to deserialize the model, we need to register assemblies. I have two custom mappings but they’re both in the same assembly. Therefore, I only need to register the assembly using one of them. Should I desire to migrate these mappings into separate assemblies later, I would need to update the code to include each assembly at least once. I’m not sure yet which is a better practice: include all custom assemblies regardless of whether you need them, or go back and modify calling code later. What I do know is that it’s a bit annoying when all I really wanted was a simple string or integer translation. Anyhow, here’s the code:


Once I’ve registered those assemblies, I can reconstitute the model using a method call:

var newModel = bmt.LoadModel(mlContext, modelPath);

That method call wraps the following code:

public ITransformer LoadModel(MLContext mlContext, string modelPath)
	ITransformer loadedModel;
	using (var stream = File.OpenRead(modelPath))
		DataViewSchema dvs;
		loadedModel = mlContext.Model.Load(stream, out dvs);

	return loadedModel;

What we’re doing is building a model (of interface type ITransformer) as well as a DataViewSchema which I don’t need here. I get back my completed model and can use it like I just finished training. In fact, I have a test case in my GitHub repo which compares a freshly-trained model versus a saved and reloaded model to ensure function calls work and the outputs are the same:

private string GenerateOutcome(PredictionEngineBase<RawInput, Prediction&gt; pe)
	return pe.Predict(new RawInput
		Game = 0,
		Quarterback = "Josh Allen",
		Location = "Home",
		NumberOfPointsScored = 17,
		TopReceiver = "Robert Foster",
		TopRunner = "Josh Allen",
		NumberOfSacks = 0,
		NumberOfDefensiveTurnovers = 0,
		MinutesPossession = 0,
		Outcome = "WHO KNOWS?"

public void SaveAndLoadModel()
	string modelPath = "C:\\Temp\\BillsModel.mdl";
	bmt.SaveModel(mlContext, model, modelPath);

	// Register the assembly that contains 'QBCustomMappings' with the ComponentCatalog
	// so it can be found when loading the model.

	var newModel = bmt.LoadModel(mlContext, modelPath);

	var newPredictor = mlContext.Model.CreatePredictionEngine<RawInput, Prediction&gt;(newModel);
	var po = GenerateOutcome(predictor);
	var npo = GenerateOutcome(newPredictor);

	Assert.AreEqual(po, npo);

Results, naturally, align.


In today’s post, we looked at the ability to serialize and deserialize models. We looked at saving the model to disk, although we could save to some other location, as the Save() method on MLContext requires merely a Stream and not a FileStream. Saving and loading models is straightforward as long as you’ve done all of the pre-requisite work around custom mappings (or avoided them altogether).


Machine Learning with .NET: Predictions

This is part four in a series on Machine Learning with .NET.

Thanks to some travels, part 4 in the series is a bit later than I originally anticipated. In the prior two posts, we looked at different ways to create, train, and test models. In this post, we’ll use a trained model and generate predictions from it. We’ll go back to the model from part two.

Predictions in a Jar

Our first step will be to create a test project. This way, we can try out our code without the commitment of a real project. Inside that test project, I’m going to include NUnit bits and prep a setup method. I think I also decided to figure out how many functional programming patterns I could violate in a single file, but that was an ancillary goal. Here is the code file; we’ll work through it block by block.

Setting Things Up

First, we have our setup block.

public class Tests
	MLContext mlContext;
	BillsModelTrainer bmt;
	IEstimator<ITransformer&gt; trainer;
	ITransformer model;
	PredictionEngineBase<RawInput, Prediction&gt; predictor;

	public void Setup()
		mlContext = new MLContext(seed: 9997);
		bmt = new BillsModelTrainer();

		var data = bmt.GetRawData(mlContext, "Resources\\2018Bills.csv");
		var split = mlContext.Data.TrainTestSplit(data, testFraction: 0.25);

		trainer = mlContext.MulticlassClassification.Trainers.NaiveBayes(labelColumnName: "Label", featureColumnName: "Features");
		model = bmt.TrainModel(mlContext, split.TrainSet, trainer);

		predictor = mlContext.Model.CreatePredictionEngine<RawInput, Prediction&gt;(model);

In here, I have mutable objects for context, the Bills model trainer class, the output trainer, the ouptut model, and a prediction engine which takes data of type RawInput and generates predictions of type Prediction. This means that any data we feed into the model must have the same shape as our raw data.

The Setup() method will run before each batch of tests. In it, we build a new ML context with a pre-generated seed of 9997. That way, every time I run this I’ll get the same outcomes. We also new up our Bills model trainer and grab raw data from the Resources folder. I’m going to use the same 2018 game data that we looked at in part 2 of the series, as that’s all of the data I have. We’ll pull out 25% of the games for testing and retain the other 75% for training. For the purposes of this demo, I’m using the Naive Bayes classifier for my trainer, and so TrainModel() takes but a short time, returning me a model.

To use this model for prediction generation, I call mlContext.Model.CreatePredictionEngine with the model as my parameter and explain that I’ll give the model an object of type RawInput and expect back an object of type Prediction. The Prediction class is trivial:

public class Prediction
	public string Outcome { get; set; }

I would have preferred the option to get back a simple string but it has to be a custom class.

Knocking Things Down

Generating predictions is easy once you have the data you need. Here’s the test function, along with five separate test cases:

[TestCase(new object[] { "Josh Allen", "Home", 17, "Robert Foster", "LeSean McCoy", "Win" })]
[TestCase(new object[] { "Josh Allen", "Away", 17, "Robert Foster", "LeSean McCoy", "Win" })]
[TestCase(new object[] { "Josh Allen", "Home", 17, "Kelvin Benjamin", "LeSean McCoy", "Win" })]
[TestCase(new object[] { "Nathan Peterman", "Home", 17, "Kelvin Benjamin", "LeSean McCoy", "Loss" })]
[TestCase(new object[] { "Josh Allen", "Away", 7, "Charles Clay", "LeSean McCoy", "Loss" })]
public void TestModel(string quarterback, string location, float numberOfPointsScored,
	string topReceiver, string topRunner, string expectedOutcome)
	var outcome = predictor.Predict(new RawInput
		Game = 0,
		Quarterback = quarterback,
		Location = location,
		NumberOfPointsScored = numberOfPointsScored,
		TopReceiver = topReceiver,
		TopRunner = topRunner,
		NumberOfSacks = 0,
		NumberOfDefensiveTurnovers = 0,
		MinutesPossession = 0,
		Outcome = "WHO KNOWS?"

	Assert.AreEqual(expectedOutcome, outcome.Outcome);

What I’m doing here is taking in a set of parameters that I’d expect my users could input: quarterback name, home/away game, number of points scored, leading receiver, and leading scorer. I also need to pass in my expected test result—that’s not something I’d expect a user to give me, but is important for test cases to make sure that my model is still scoring things appropriately.

The Predict() method on my predictor takes in a RawInput object, so I can generate that from my input data. Recall that we ignore some of those raw inputs, including game, number of sacks, number of defensive turnovers, and minutes of possession. Instead of bothering my users with that data, I just pass in defaults because I know we don’t use them. In a production-like scenario, I’m more and more convinced that I’d probably do all of this cleanup before model training, especially when dealing with wide data sets with dozens or maybe hundreds of inputs.

Note as well that I need to pass in the outcome. That outcome doesn’t have to be our best guess or anything reasonable—it just needs to have the same data type as our original input data set.

What we get back is a series of hopefully-positive test results:

Even in this model, Kelvin Benjamin drops the ball.

Predictions are that easy: a simple method call and you get back an outcome of type Prediction which has a single string property called Outcome.

Trying Out Different Algorithms

We can also use this test project to try out different algorithms and see what performs better given our data. To do that, I’ll build an evaluation tester. The entire method looks like this:

[TestCase(new object[] { "Naive Bayes" })]
[TestCase(new object[] { "L-BFGS" })]
[TestCase(new object[] { "SDCA Non-Calibrated" })]
public void BasicEvaluationTest(string trainerToUse)
	mlContext = new MLContext(seed: 9997);
	bmt = new BillsModelTrainer();

	var data = bmt.GetRawData(mlContext, "Resources\\2018Bills.csv");
	var split = mlContext.Data.TrainTestSplit(data, testFraction: 0.4);

	// If we wish to review the split data, we can run these.
	var trainSet = mlContext.Data.CreateEnumerable<RawInput&gt;(split.TrainSet, reuseRowObject: false);
	var testSet = mlContext.Data.CreateEnumerable<RawInput&gt;(split.TestSet, reuseRowObject: false);

	IEstimator<ITransformer&gt; newTrainer;
	switch (trainerToUse)
		case "Naive Bayes":
			newTrainer = mlContext.MulticlassClassification.Trainers.NaiveBayes(labelColumnName: "Label", featureColumnName: "Features");
		case "L-BFGS":
			newTrainer = mlContext.MulticlassClassification.Trainers.LbfgsMaximumEntropy(labelColumnName: "Label", featureColumnName: "Features");
		case "SDCA Non-Calibrated":
			newTrainer = mlContext.MulticlassClassification.Trainers.SdcaNonCalibrated(labelColumnName: "Label", featureColumnName: "Features");
			newTrainer = mlContext.MulticlassClassification.Trainers.NaiveBayes(labelColumnName: "Label", featureColumnName: "Features");

	var newModel = bmt.TrainModel(mlContext, split.TrainSet, newTrainer);
	var metrics = mlContext.MulticlassClassification.Evaluate(newModel.Transform(split.TestSet));

	Console.WriteLine($"Macro Accuracy = {metrics.MacroAccuracy}; Micro Accuracy = {metrics.MicroAccuracy}");
	Console.WriteLine($"Confusion Matrix with {metrics.ConfusionMatrix.NumberOfClasses} classes.");

	Assert.AreNotEqual(0, metrics.MacroAccuracy);

Basically, we’re building up the same set of operations that we did in the Setup() method, but because that setup method has mutable objects, I don’t want to run the risk of overwriting the “real” model with one of these tests, so we do everything locally.

The easiest way I could think of to handle multiple trainers in my test case was to put in a switch statement over my set of valid algorithms. That makes the method brittle with respect to future changes (where every new algorithm I’d like to test involves modifying the test method itself) but it works for our purposes here. After choosing the trainer, we walk through the same modeling process and then spit out a few measures: macro and micro accuracy, as well as a confusion matrix.

To understand macro versus micro accuracy (with help from ML.NET and Sebastian Raschka):

  • Micro accuracy sums up all of your successes and failures, whether the outcome is win, lose, draw, suspension due to blizzard, or cancellation on account of alien abduction. We add up all of those cases and if we’re right 95 out of 100 times, our micro accuracy is 95%.
  • Macro accuracy calculates the harmonic mean of all of the individual classes. If our accuracies for different outcomes are: [ 85% (win), 80% (loss), 50% (draw), 99% (suspension due to blizzard), 33% (cancellation on account of alien abduction) ], we sum them up and divide by the number of outcomes: (0.85 + 0.8 + 0.5 + 0.99 + 0.33) / 5 = 69.4%.

Both of these measures are useful but macro accuracy has a tendency to get pulled down when a small category is wrong. Let’s say that there are 3 alien abduction scenarios out of 100,000 games and we got 1 of those 3 correct. Unless we were trying to predict for alien abductions, the outcome is uncommon enough that it’s relatively unimportant that we get it right. Our macro accuracy is 69.4% but if we drop that class, we’re up to 78.5%. For this reason, micro accuracy is typically better for heavily imbalanced classes.

Finally, we print out a confusion matrix, which shows us what each model did in terms of predicted category versus actual outcome.


Generating predictions with ML.NET is pretty easy, definitely one of the smoothest experiences when working with this product. Although we looked at a test in the blog post, applying this to real code is simple as well as you can see in my MVC project demo code (particularly the controller). In this case, I made the conscious decision to train anew at application start because training is fast. In a real project, we’d want to save a model to disk and load it, which is what we’ll cover next time.

DataCamp Course: Time Series Analysis in SQL Server

I have a DataCamp course which just launched entitled Time Series Analysis in SQL Server.

Course Description

SQL Server has a robust set of tools to prepare, aggregate, and query time series data. This course will show you how to build and work with dates, parse dates from strings (and deal with invalid strings), and format dates for reporting. From there, you will see how SQL Server’s built-in aggregation operators and window functions can solve important business problems like calculating running totals, finding moving averages, and displaying month-over-month differences using realistic sample data sets. You will also see how taking a different perspective on your data can solve difficult problems.

Chapter 1 is available for free, so you can check it out at no risk of accidentally giving me money. This chapter also includes one of the biggest tricks behind working with time series data in SQL Server: calendar tables. And add in a bonus bit on the APPLY operator. The other three chapters are great too, but those are for closers subscribers.

Machine Learning with .NET: Model Builder

This is part three in a series on Machine Learning with .NET.

Last time on 36 Chambers, we looked at building our own model to explain why Kelvin Benjamin is awful. Today, we’re going to use the ML.NET Model Builder to understand how awful airlines are. The unofficial motto of this blog, after all, is “Parades need rain, too.”

The ML.NET Model Builder is still in preview, so expect things to change; some of those changes might be breaking changes, too. If you want to play along, download the Model Builder from the Visual Studio Marketplace. This works for Visual Studio 2017 and Visual Studio 2019, so you’ll need one of those. Community Edition is supported as well, so you don’t need a fancy version of Visual Studio.

After you install the extension, you’ll need to be in a solution with at least one project in it. The Model Builder creates new projects but is a context menu option off of a project itself. If you want to play along at home, clone my GitHub repo. I also have some instructions on the repo, which is rare for me.

First up, right-click on any of the projects and select Add --> Machine Learning.

We’ll add machine learning, all right, just not here. Not Europa.

This brings up a wizard tab to build a machine learning model.

Choose the form of the destructor.

You have four options from which to choose: two-class classification, multi-class classification, regression, or Choose Your Own Adventure. Today, we’re going to create a two-class classification model. Incidentally, they’re not kidding about things changing in preview—last time I looked at this, they didn’t have multi-class classifiers available.

Once you select Sentiment Analysis (that is, two-class classification of text), you can figure out how to feed data to this trainer.

Twitter is a garbage hole and airlines are great garbage magnets, so this should be grand.

As of right now, we can use delimited files or read from a table in SQL Server. Note the 1 GB limit for maximum file size; this is not a solution for very large data sets.

Once we select a file, we get the option to predict a label. As of the time of writing, this label must have a value of either 1 or 0. Our tweets are in the text column, so we’ll use that to predict sentiment.

From there, we move on to training.

This model is definitely a 10.

We don’t specify the algorithm or algorithms to use, but we do specify the number of seconds allowed for training. This is an interesting specification parameter; basically, the more seconds you allow for training, the more likely you are to find a good model. I don’t have any insight into the specifics of what happens during this training period; my expectation is that they iterate through a set of algorithms and try progressively more complicated sets of inputs and transformations. The ML.NET team does have a set of guidelines around how long you might want to train given data sizes. I’m going to train for 4 minutes because I didn’t plan well enough ahead to give myself ten minutes to let it wait. Here it is in progress:

Hard at work on training, boss. Training all kinds of stuff here.

Eventually, we run out of time and we have a winner:

Like a proper horse race, this one changed just as we ran out of time. The most exciting 4 minutes in sports.

Now it is time to evaluate. We get to see a table with the five best models on a few criteria: accuracy, Area Under the Curve (AUC), Area Under the Precision-Recall Curve (AUPRC), F1 score, and the amount of time it took to train.

One of these things is not like the others.

One thing I wish I could see here was a breakdown of our key correlation matrix measures—AUC, AUPRC, and F1 score are fine combination metrics and are very useful, but when you have major discrepancies in class makeup, recall by itself can mean a lot. For example, suppose we have a test for a medical condition which affects 1 in 100,000 people. We can get a 99.999% accuracy if we simply return “Negative result” for everybody, but that’s an awful test because the recall is zero: we never catch any of the positive results. It could be that I am not clever enough to interpret these results and infer the base values, though.

Regardless, our next step is to generate some code. Click on the “Code” link and you get the opportunity to add new projects. Note that you do not get to choose which of the models you want to use; the Model Builder picks the one with the highest accuracy.

Time to make the donuts.

We now have two new projects.

I made the donuts.

We have a model project and a console application. The model project contains our trained model (as a zip file) and input and output files. Here is the input file:

using Microsoft.ML.Data;

namespace DotNetMachineLearningML.Model.DataModels
    public class ModelInput
        [ColumnName("airline_sentiment"), LoadColumn(0)]
        public bool Airline_sentiment { get; set; }

        [ColumnName("text"), LoadColumn(1)]
        public string Text { get; set; }

And here is the output file:

using System;
using Microsoft.ML.Data;

namespace DotNetMachineLearningML.Model.DataModels
    public class ModelOutput
        // ColumnName attribute is used to change the column name from
        // its default value, which is the name of the field.
        public bool Prediction { get; set; }

        public float Score { get; set; }

The console application contains a model builder and a Program file. The ModelBuilder C# file allows us to modify or re-train our model should we desire, but is not needed for regular operation. The console application grabs a row at random from our tweet set and shows the prediction versus actual values. Instead of doing that, I’m going to take user input to classify a tweet. Replace the main method with this code:

static void Main(string[] args)
	MLContext mlContext = new MLContext();

	ITransformer mlModel = mlContext.Model.Load(GetAbsolutePath(MODEL_FILEPATH), out DataViewSchema inputSchema);
	var predEngine = mlContext.Model.CreatePredictionEngine<ModelInput, ModelOutput&gt;(mlModel);

	Console.WriteLine("Enter a sample tweet:");
	var text = Console.ReadLine();

	// Create sample data to do a single prediction with it 
	ModelInput sampleData = new ModelInput { Text = text };

	// Try a single prediction
	ModelOutput predictionResult = predEngine.Predict(sampleData);
	string outcome = predictionResult.Prediction ? "Positive tweet" : "Negative tweet";
	Console.WriteLine($"Single Prediction --&gt; Predicted value: {outcome}");

	Console.WriteLine("=============== End of process, hit any key to finish ===============");

Change the console application to be the startup project and away we go:

Too soon for TWA jokes?


The Model Builder won’t replace an actual machine learning pipeline anytime soon, but it’s a good way to get started with processing clean data when you have no clue which algorithms you should try.

Machine Learning with .NET: Modeling

This is part two in a series on Machine Learning with .NET.

In the first post in this series, I took a look at the “why” behind ML.NET as well as some of its shortcomings in data processing. In this post, I want to look at an area where it does much better: training models.

A Simple Model: Predicting Victory

In this first example, I’m going to put together a small but complete demonstration of a business problem.

In 2018, the Buffalo Bills went 6-10. Previously on 36 Chambers, we learned how much Kelvin Benjamin dragged the team down. Now we’re going to re-learn it but this time in .NET. We will solve a classification problem with two cases: win or loss.

Data Preparation

Our input features include the starting quarterback, location (home or away), number of points scored, top receiver (by yards), top rusher (by yards), team number of sacks, team number of defensive turnovers, team minutes of possession, and the outcome (our label).

We can represent all of this in a class, RawInput:

public class RawInput
	//1,Peterman,Away,3,Zay Jones,Marcus Murphy,1,1,25,Loss

	public float Game { get; set; }
	public string Quarterback { get; set; }
	public string Location { get; set; }
	public float NumberOfPointsScored { get; set; }
	public string TopReceiver { get; set; }
	public string TopRunner { get; set; }
	public float NumberOfSacks { get; set; }
	public float NumberOfDefensiveTurnovers { get; set; }
	public float MinutesPossession { get; set; }
	public string Outcome { get; set; }

There are a couple of points I want to make here:

  • Each attribute receives a tag which represents the order in which we load columns.
  • Every numeric feature must be a float. Even integers.
  • We need to specify the label with its own column name.

With a class, I can create a quick function to load my raw data into a list of RawData branded data:

public IDataView GetRawData(MLContext mlContext, string inputPath)
	return mlContext.Data.LoadFromTextFile<RawInput&gt;(path: inputPath, hasHeader: true, separatorChar: ',');

The IDataView interface is our .NET version of the DataFrame in R or Pandas. The good news here is that just by creating a POCO with a few attributes, I can interact with ML.NET. Right now, loading from text files is the primary data load scenario, but I could see hitting SQL Server or other ODBC sources as well as Excel files, etc. in the future.

Build a Trainer

My next function trains a model. We’re going to use Naive Bayes here as well, just to keep consistent with the prior blog post.

Here are the transformations I’d like to do before feeding in my data:

  1. Translate quarterback name based on a simple rule: Josh Allen maps to Josh Allen and every other QB maps to Nate Barkerson, the man of a million faces interceptions.
  2. Translate number of points scored based on a simple rule: if they scored double digits, return true; otherwise, return false.
  3. Drop the columns for the number of sacks, number of defensive turnovers, and number of minutes of possession. These columns are probably useful but we aren’t going to use them in this Naive Bayes model.
  4. Drop the Game feature, which represents the game number. We don’t need it.

Now if you’ll allow me a rant.

Code Plus a Rant

In order to perform operation #1, I need to perform a custom mapping using mlContext.Transforms.CustomMapping. My rule is exceedingly simple; here it is in C# lambda expression form: name => name == "Josh Allen" ? "Josh Allen" : "Nate Barkerson". Real easy…except it’s not.

See, first I need to build input and output classes for my custom mapping, so it’s really mlContext.Transforms.CustomMapping<QBInputRow, QBOutputRow>. I can’t use a simple type here, either: it has to be a class.

So let’s create some classes:

public class QBInputRow
	public string Quarterback { get; set; }

public class QBOutputRow
	public string QuarterbackName { get; set; }

Okay, now that I have classes, I need to put in that lambda. I guess the lambda could change to qb => qb.Quarterback == "Josh Allen" ? "Josh Allen" : "Nate Barkerson" and that’d work except for one itsy-bitsy thing: if I do it the easy way, I can’t actually save and reload my model. Which makes it worthless for pretty much any real-world scenario.

So no easy lambda-based solution for us. Instead, we need a delegate. That’s going to be another class with a static method and a GetMapping() action:

public class QBCustomMappings : CustomMappingFactory<QBInputRow, QBOutputRow&gt;
	// This is the custom mapping. We now separate it into a method, so that we can use it both in training and in loading.
	public static void QBMapping(QBInputRow input, QBOutputRow output) =&gt; output.QuarterbackName =
		(input.Quarterback == "Josh Allen") ? "Josh Allen" : "Nate Barkerson";

	// This factory method will be called when loading the model to get the mapping operation.
	public override Action<QBInputRow, QBOutputRow&gt; GetMapping()
		return QBMapping;

After creating the QBMapping() function, I can finally reference it: mlContext.Transforms.CustomMapping(QBCustomMappings.QBMapping, nameof(QBCustomMappings.QBMapping)). I need to create three separate classes to do a simple mapping. Oh, and three more classes to map my points scored. That’s six classes I would never have had to create in R or Python.

That’s a lot of boilerplate code considering in my mind, it’s a simple transformation. This leads me to advise against using custom transformations if you can. Instead, do all of your transformations before getting the data, but I think that means you can’t use the easy data load method I showed above (though I could be wrong on that score).

Rant over. Now that I have my mapping classes all built out, my training method looks like this:

public TransformerChain<Microsoft.ML.Transforms.KeyToValueMappingTransformer&gt; TrainModel(
	MLContext mlContext, IDataView data)
	var pipeline =
		mlContext.Transforms.CustomMapping<QBInputRow, QBOutputRow&gt;(
			QBCustomMappings.QBMapping, nameof(QBCustomMappings.QBMapping))
		.Append(mlContext.Transforms.CustomMapping<PointsInputRow, PointsOutputRow&gt;(
			PointsCustomMappings.PointsMapping, nameof(PointsCustomMappings.PointsMapping)))
		// We could potentially use these features for a different model like a fast forest.
		.Append(mlContext.Transforms.DropColumns(new[] { "NumberOfSacks", "NumberOfDefensiveTurnovers",
			"MinutesPossession" }))
		.Append(mlContext.Transforms.DropColumns(new[] { "Game", "Quarterback" }))
		.Append(mlContext.Transforms.Concatenate("FeaturesText", new[]
		.Append(mlContext.Transforms.Text.FeaturizeText("Features", "FeaturesText"))
		// Label is text so it needs to be mapped to a key
		.Append(mlContext.Transforms.Conversion.MapValueToKey("Label"), TransformerScope.TrainTest)
		.Append(mlContext.MulticlassClassification.Trainers.NaiveBayes(labelColumnName: "Label", featureColumnName: "Features"))
		.Append(mlContext.Transforms.Conversion.MapKeyToValue("PredictedOutcome", "PredictedLabel"));

	var model = pipeline.Fit(data);

	return model;

I’m building out a data pipeline here, which performs transformations in a series, using the Append() method to link parts together similar to |> in F# or %>% in R. It’s not nearly as pretty as either of those solutions, but it’s the best we’re getting with C#.

Our first two operations are the data transformations to get our QB name and “did they score double-digit points?” features. After that, we drop unused features using the DropColumns() method.

The next part deserves a bit of discussion. With ML.NET, we’re only allowed to send in one text column, so we need to combine together all of our string features and “featurize” them. The combination of Concatenate() and FeaturizeText() does this for us.

After we finish that part of the job, we need to turn our “Win” and “Loss” values into key-value mappings. ML.NET needs keys for binary and multi-class classification, as it will not train on labels. We want to keep the labels so we understand which class we’re in, so we compromise by using the MapValueToKey() method.

Then, we want to train using the Naive Bayes algorithm. ML.NET classifies Naive Bayes as a multi-class classifier and not a binary classifier, so we need to use the multi-class set even though our data set has only wins and losses. Finally, after we get back a class key, we need to map that key back to a value and return it. This way, we know our class name.

Finally, we fit the model to our data and return the fitted model.

Training and Evaluating the Model

The actual process of training the model has us retrieve data, split it into training and test data sets, and perform model training. Here is an example:

MLContext mlContext = new MLContext(seed: 9997);
BillsModelTrainer bmt = new BillsModelTrainer();

var data = bmt.GetRawData(mlContext, "Resources\\2018Bills.csv");
var split = mlContext.Data.TrainTestSplit(data, testFraction: 0.4);

// If we wish to review the split data, we can run these.
var trainSet = mlContext.Data.CreateEnumerable<RawInput&gt;(split.TrainSet, reuseRowObject: false);
var testSet = mlContext.Data.CreateEnumerable<RawInput&gt;(split.TestSet, reuseRowObject: false);

ITransformer model = bmt.TrainModel(mlContext, split.TrainSet);
var metrics = mlContext.MulticlassClassification.Evaluate(model.Transform(split.TestSet));

Console.WriteLine($"Macro Accuracy = {metrics.MacroAccuracy}; Micro Accuracy = {metrics.MicroAccuracy}");
Console.WriteLine($"Confusion Matrix with {metrics.ConfusionMatrix.NumberOfClasses} classes.");

I also threw in model evaluation here because it’s pretty easy to do. We generate an ML context, load our data, and then split it into test and training data. Interestingly, I set the test fraction to 0.4 (or 40%) but it only pulled 25% of my data. I imagine that with a larger data set, I’d see closer to 40% reserved for testing but it’s luck of the draw with just 16 rows. By the way, never trust a model with 12 data points.

Speaking of models, we run the TrainModel() method and get back a model. From there, I can evaluate the model using the Evaluate() method and get back some metrics. For multi-class classification problems, I get back micro accuracy, macro accuracy, and a confusion matrix.

Macro Accuracy = 1; Micro Accuracy = 1
 Confusion Matrix with 2 classes.
 Confusion table
 PREDICTED ||     Loss |      Win | Recall
 TRUTH     ||======================
      Loss ||        3 |        0 | 1.0000
       Win ||        0 |        1 | 1.0000
 Precision ||   1.0000 |   1.0000 |

Oh, I had a 100% correct rate for my test data. Like I said, don’t trust models based off of 12 data points and don’t trust evaluations with 4 data points.

Model Changes

If I want to change the model I use for training, I can change my TrainModel() method. For multi-class classification, we have about a half-dozen models from which to choose:

Six models, ah ah ah

These models have a few trade-offs, including computational complexity, accuracy, and assumptions regarding the shape of data. Investigate and choose based on your problem and data, but don’t assume every one solves everything equally well.


One last thing I want to point out is cross-validation. Doing this with ML.NET is really easy, but I need to get the pipeline out of TrainModel(). If I want to use cross-validation in production, I’d probably have one method which returns the pipeline and a second method which takes a pipeline and training data and generates a model for me. For now, here’s the cross-validation part specifically:

var cvResults = mlContext.MulticlassClassification.CrossValidate(data, pipeline, numberOfFolds: 4);

var microAccuracies = cvResults.Select(r => r.Metrics.MicroAccuracy);

I built out four folds, so we train on 12 games and test on 4 games. The average micro-accuracy is .64 or 64%, which is about what I expected. It’s not a great accuracy, but then again, it’s 12 data points.


In today’s post, we looked at training, testing, and evaluating models. I think that overall, this is a reasonably good experience if you have clean data. As soon as you want to perform non-standard transformations in the data pipeline, though, things get busy fast, in a way that we don’t typically see in R or Python.

In the next post in the series, I’ll show a completely different method for building models: the Model Builder.

Machine Learning with .NET: Premise

This is part one in a series on Machine Learning with .NET.

Microsoft has officially released ML.NET 1.0. The idea behind ML.NET is to bring some of the data science techniques and algorithms we use in R and Python over to C# and F#. Over the course of this series, we will look at a few examples, but in this first post, I’d like to cover some of the reasoning for why you might want to use it.

An Early Digression: Documentation

For a new library, documentation is critical. Ideally, you want to have a series of examples covering the basics, as well as at least a couple of examples covering more advanced topics. That way, your first generation of users will discover and use your product more readily. As they gain skills, they go out and train others, leading to waves of acceptance if all goes well. But if you have poor documentation, your first wave of users may not fully grasp the power of what you’re giving them.

To its credit, ML.NET has a good documentation. Their tutorial gets you started with the Model Builder (which we’ll look at in the third post of this series) and they have other step-by-step tutorials which show how to hand-design models. They also have quite a few samples on their ML.NET Samples repo.

The First Question: Why?

I think this is a fair question, so here’s my attempt at answering it from the standpoint of a practitioner with a team familiar with R and Python.

A Long Time Go in a Server Far Away

If we go back ten or so years, statisticians and data analysts used tools like R, SAS, and Matlab to perform analyses and create models. If you wanted to turn this into production-worthy code, however, you didn’t simply spin up an R or SAS server and let the world bang away at it—those tools were relatively slow, inefficient, and (especially with R at that time) bug-prone.

Instead, data analysts tended more often to generate a model using one of these tools, describe the algorithm and weights to programmers, and the programmers would rewrite the model in a language like C. The upside to doing this was that you could count on the code working and being reasonably fast; the downside is that model changes necessitated code changes (or at least configuration file changes).

Enter the Present Era

Over the past 5-7 years, Python and R have come a long way, or two separate long ways. For Python, I see the turning point as the evolution of scikit-learn in about 2013-2014 leading this language into a space where earlier attempts (such as SciPy) were much less successful. Meanwhile, R always had the algorithms but it wasn’t until around 2013-2014 that we started to see the major improvements in stability and performance needed to take it seriously as a production-worthy language.

Today, both languages are fine for production purposes—for example, my team has production code running in R and in Python, and yet I sleep well at night.

The Burdens of Polyglot Infrastructure

I have to admit, though, that when your company has a bunch of .NET developers and your operations people are all used to working with .NET, the struggle is real when I tell them that we’re going to use R and Python. For more risk-averse companies, it might be okay to use R or Python for personal development, but when it comes time to move out, they want nothing but C# code.

If that’s the scenario you’re in, ML.NET can be useful. This way, you can build the entire pipeline in C#, integrate it easily with your existing C# code, and maintain it in C#.


Installation of ML.NET is pretty simple: there is a NuGet package. I have Microsoft.ML and Microsoft.ML.DataView installed in my solution.

Two important NuGet packages.

You will also want to download and install the Microsoft ML.NET Model Builder. As of the time of this post (May 28th, 2019), the Model Builder is in a public preview state.

Data Analysis

Let’s get a quick reminder of the Microsoft Team Data Science Process. If you’d like a fuller picture, try out this post, which is part 1 of my series on launching a data science project.

The Team Data Science Project Lifecycle (Source)

The first two phases of the lifecycle are business understanding and data acquisition & understanding. Frankly, .NET (ML.NET included) is pretty awful at both of those phases of the lifecycle.

I don’t dock it too many points for not being good at business understanding—R and Python aren’t any good at that step of the process, either. Instead, Excel, OneNote, and pen & paper are going to be some of your most valuable tools here.

Where I think it really falls short, however, is in the data analysis phase. R and Python are excellent for data analysis for a few reasons:

  1. They both make it easy to load data of various shapes and origins (Excel file, flat file, SQL database, API, etc.).
  2. They both make it easy to perform statistical analysis on data sets and get results back in a reasonable time frame. Statistical analysis can be something as trivial as a five-number summary and can scale up to more complicated statistical analyses.
  3. They both make it easy to transform and reshape data. If I have to define classes, I’m working too hard.
  4. They both make it easy to visualize data, whether that’s ggplot2, plotly, bokeh, matplotlib, or whatever.
  5. They both make it easy to explore data. In an R console, I can use functions like head() to grab the first couple of rows, see what they look like, and make sure that I get what I’m expecting. If I’m using R Studio, I get a built-in data frame viewer.
  6. They both have enormous libraries of statistical and analytical functionality developed over the course of decades.

On the C# side, here’s my argument:

  1. .NET has plenty of functionality around loading data from numerous sources. I don’t think it’s necessarily easy, particularly when you’re dealing with huge files with hundreds or thousands of columns. Do you seriously want me to create a class with all of those members just to perform exploratory analysis?
  2. .NET has Math.NET. I don’t think it’s as easy as what’s available in R and Python, but it’s solid.
  3. I have to define classes, so I’m working too hard.
  4. I guess I can use the Chart class with C#, but I don’t think it’s easy, particularly for throwaway stuff.
  5. Data exploration is a weak point, even with ML.NET. If I just want to see a few rows, I suppose I could build a unit test or console app, but that’s a lot of overkill. There is a C# Interactive which tries to mitigate some of the pain.
  6. Without castigating the work the Math.NET, Accord.NET, and ML.NET teams have done, C# is going to take the L here.

When it comes to F#, the data analysis story is a little better:

  1. Type providers make it a good bit easier to work with data without the expectation that I’m creating classes on my own. Record types are good here. I’d rate this as pretty solid once you get used to type providers.
  2. Same as the C# answer, so pretty solid.
  3. F# has its advantages, particularly around a very strict type system. I think that strict type system slows down exploratory work and
  4. FSharp.Charting is not bad, but it’s several rungs below the libraries I listed for R and Python. I haven’t tried XPlot yet, so maybe that will end up contradicting my gripe-fest here.
  5. F# does have a good REPL and you can create fsx scripts easily, so I give it credit there. I still think it feels slower exploring F# data sets than R or Python data sets. For example, I don’t know of an easy way to display a quick view of a data set like what we have in R Studio or even base R when running head().
  6. F# won’t add much to the table on this point.

In short, you can struggle through but there are much better experiences. I’m open to correction on the above points from people who spend considerably more time working with data science in the .NET space than I do.


In today’s post, I walked through some of the reasoning for ML.NET and looked at the area where it is weakest: data analysis. In the next post, we will look at an area where ML.NET is considerably stronger: data modeling.

Upcoming Events: TriPASS Data Science Group

Key Details

What: Triangle SQL Server User Group: Data Science SIG
Where: 3100 Tower Boulevard, Durham, NC
When: Tuesday, May 28th at 6 PM
Admission is free. Register on Meetup.

What I’m Presenting

06:00 PM — 08:00 PM — A Brief Look at ML.NET

This will include my example from my Machine Learning with .NET talk as well as a second quick example. This is a hands-on lab, so I have a bit more work than usual to do, as I’m going to have to build the lab guide.