.Net Foundation ML.NET, 24 Jun 2019

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: Auto Train Machine Learning Models with ML.NET and AutoML

Description

Using the AutoML package to find a best ML.NET model for the given data.

Code - https://github.com/jwood803/MLNetExamples/tree/master/MLNetExamples/AutoML

Data - https://www.kaggle.com/camnugent/california-housing-prices

AutoML documentation - https://docs.microsoft.com/en-us/dotnet/machine-learning/how-to-guides/how-to-use-the-automl-api#load-data

Contact:
Twitter: https://twitter.com/JWood/
Blog: https://jonwood.co/

A

Hey everyone so a while back I did a video about the model builder, which was kind of a Visual Studio extension that gives a GUI interface to build to automatically build machine learning algorithms. But behind the scenes the model builder uses something called Auto ml from Microsoft and basically Auto ml, just iterates over your data set, and it goes through different machine learning. Algorithms, with different hyper parameters and it'll help select the best model based on that data.

A

And what we're going to do here in this video is that we're gonna use, auto ml programmatically in a c-sharp program and kind of do the same thing that a model builder does so I'm here in Visual, Studio I have a done deck or console program created and I already have and I already have my data file included here, and this is the California housing prices data set here, and it has several features: latitude, longitude, total rooms, total bedrooms, median income, ocean proximity and what we're gonna use is our label.

A

And so what we're going to predict here is the median house value.

A

All right so to get started, I'm going to add some dependencies from nigut.

A

Yeah first I need email net and this is going to be used in version 1.1.

A

Right and then to use the auto ml features, I need to bring in the Microsoft ml to oto mo packet.

A

All right so, first things: first I need to create a new email context. I.

A

Start bringing in some using statements.

A

Next, we'll read in our our training data.

A

Take it from the context of data and we'll load from text file, not usual. When we load from a text file, we need to give it the kind of the input schema and so I put a placeholder here for the housing data class that I'll have to create in just a bit now I read in the housing CSV.

A

As a hitter and I'll give it a separator character, there's a column all right. So let's create this epic imma.

A

Now just copy this over, so you don't have to see me type in everything I mean to bring in using statements from there to there go so easy, pretty much all the properties, all the features that who showed in the data file and the median house value, I'm gonna tell basically rename it to a call name of label right. So what we can do.

A

Create an experiment and on the context we can run a function called auto, and that is an extension method on that package that we brought in and just like the model builder.

A

Currently, you can create a binary classification with the class clefts classification or regression experiment now do a regression experiment and right now, I just give it a parameter of the max experiment, time in seconds I'll just put in 10, so it'll take 10 seconds to run through all the different algorithms to help renders for the data and with that we'll get the result from the experiments by running the execute method on it and we'll give it our training data and then for a simple start.

A

I just tell it, the label will be just label label, column, name.

A

And with that result we can read in and give our get our best run from that. So I'm gonna copy some print statements here.

A

Basically, it just gives us our best run on a trainer name from this result object here. It gives us that best run. We can look at the trainer, the algorithm that I picked. We can look at the validation metrics on it to suggest r-squared and mean absolute error. So let's run this real quick.

A

The best run was too fast tree regression and we have r-squared of point zero five and it mean absolute error of almost 30,000. But there's a bit more that we can do here that the auto AML package will. Let us do so. The first thing we can do is we can give it a couple of settings and we have regression experiment, settings class that we can use here and have some properties there. The main one is we had that max experiment, time in seconds proper that we can use now continue to use tin.

A

There's a couple other things that we can do can give it a cache directory where it'll load all the training data in the cache into memory. Pretty much so that'll give you some faster.

A

Training times that way give it a cancellation token. So if the user wants to cancel that the the all the way in the whole experiment, they can do that we can. We can also tell what metrics to optimize for when it does their run. Only I think it does the r-squared and the regression, and we get a list of the trainers that are going to run through and with this that we can, if they add new trainers to it, we can do that manually or three. No, we don't want to use a trainer.

A

We're gonna, just remove it from here, but right here, I'll just leave, leave it with the next experiment, time in seconds and on the crate regression experiment I can just pass in the settings.

A

Another interesting thing we can do is since this runs for 10 seconds, as we saw when we ran it earlier, there's no kind of progress that we did while I was running so it'd be nice to kind of get some some progress.

A

Information, they tell us what's what's going on what is currently working on, and this is the same information that we got when we were running with the model builder. To do dance, imma create a new progress, and that is gonna. Take a run detail which is going to take regression metrics.

A

And that takes in an action, so I can use a blend of function and before I do anything I'm going to share the validation, metrics mixture as make sure it's not. No, because we're going to be using that and now I can just console console out kind of the.

A

Sorry, it's just regression matrix.

A

And I bring that in yeah we good now gonna tell current results and then I can check current trainer name. I can check on the validation metrics the current r-squared and.

A

Then I can also check the current mean absolute errand.

A

There we go so basically, what's it's going to do is as it updates during the experiment, run, it's going to run this method here, so to use that we will just pass it in until you execute and pass them into as the progress handler parameter.

A

And now, let's see what this looks like.

A

You know so each we're getting information for each of the current run. So doing so, the trainer named r-squared and mean I'm salute air and we still, we actually good a difference trainer name now there we go. That's just the kind of a quick overview of the auto ml package and how to use it programmatically.

A

So this can actually be used, your own pipelines, where every time you know you get some new data, you can run through this code and automatically generate kind of the best model that you know you can use within your internet projects.

A

Alright. So thanks for watching and we'll see you next time, Thanks.