What is Azure Machine Learning service and how data scientists use it


– Coming up, we’ll look at the new Azure Machine Learning service that gives data scientists
the capabilities they need across the
machine learning lifecycle, within a familiar notebook experience. Now we’re gonna show you
how Azure ML makes it easier for data scientists to build
reproduceable experiments with machine learning pipelines and communicate operational dependencies to their engineering counterparts as part of a new MLOps
approach as you deploy to the Cloud and the edge at scale. And how you can use the newly introduced Automated Machine Learning
capabilities in Azure ML to build machine learning models
in a fraction of the time. (upbeat music) So, I’m joined today by Chris Lauren from the Azure Machine Learning team. Welcome. – Thanks, I’m excited to be on the show. – So, Azure Machine
Learning is a new service, but at Microsoft we’ve been
at this space for a while. In the past on Mechanics for example, we’ve looked at things like
environments for data prep and experimentation and
analysis using things like Azure Data Bricks as
well as various frameworks. But how do we think about
Azure Machine Learning, this service and really what’s
new and what’s different? – In the past if you wanted
to use machine learning in production settings, you
needed to bring together a bunch of different services to support the full machine learning lifecycle. For example bringing together storage like Azure Blob Storage or
Azure Data Lakes Storage, because you can’t train machine learning models without data. Bring in some compute like
individual Virtual Machines or Spark cluster using HD Insight or Azure Databricks to run your code. And then to protect your
data for enterprise readiness you’d bring in your virtual networks or configure your compute
and data inside the same virtual network and bring
in some Azure Key vault to manage and secure your
credentials for example. And then if you wanted to
repeat your experiments using a consistent set of
machine learning libraries and the different versions
thereof, then you’d create docker containers and use
Azure Container registry to store those docker
containers and then put that inside your VNET
and then to run all that at scale you might use
Azure Kubernetes Service and include that inside your VNET. – Right and this all sounds
like a whole lot of stuff to piece together to
get the machine learning and all the models and everything to work. – Yeah, it is a lot to manage
and you spend more time managing all that stuff
than actually doing the data science work, so with Azure
Machine Learning Service, we remove that complexity for you. As a managed service it
comes with it’s own compute, hosted notebooks and capabilities
for model management, version control and model
reproducibility built right in. You can layer that on top of
your existing Azure services. For example you can plug
in the compute and storage that you already have as well as your other infrastructure services. Azure Machine Learning
connects and orchestrates them within a single environment so
that you have one end-to-end modular platform for your entire
machine learning lifecycle as you prepare data, build, train, package and deploy your machine learning models. – And as you mentioned
then, for the data scientist there’s nothing to worry
about there in that case. You’re basically getting
the services and everything configured for them and
not having to compile all these individual services together. – Right, but the best
part is you don’t need to learn how to use new tools. You can use Jupyter Notebooks, your favorite Python editor
like VS Code or PyCharm, and Azure Machine Learning
works with any Python machine learning framework library like TensorFlow, PyTorch,
SCIKIT-Learn, and more. – This is awesome in terms
of making that lifecycle a lot shorter, in terms
of getting everything up and running but for
all the Cloud operators and engineers running that
or maybe tenant admins, how do we ensure that our data scientists aren’t consuming too much resource? – Yeah, because it is an Azure service we have integrated
role-based access control built right in that you
can fulling customize in any way you like. So for example, you can
add a new data science role so you can grant access
to a data scientist and you can simply facilitate
them to run experiments but not create or edit
compute for example, so they can do their job
without running up your bills in a way that’s not approved. – Right, and the cool part of this service is something brand new, MLOps. Now this is a new concept
for a lot of people. Can you explain what
MLOps is and how it works? – So what really differentiates
Azure Machine Learning is that we have an
integrated DevOps approach for what we think of as MLOps. It makes it easier for data
scientists and engineers to work together,
because engineers already have a good understanding of
how this continuous integration and continuous deployment process works and the data scientists know
how to train great models. And so by enabling them to work together we can ensure that we have high-quality models at
scale in production. With MLOps incorporated as part of the Azure Machine Learning service, we add experimentation. This means for data scientists
that reproducibility is incorporated throughout
the Machine Learning lifecycle in your training, test and
production environments and they can create discreet
pipelines for each model making them easier to track and reproduce. For example, let’s say you have
a wind farm that you manage and you want to ensure that you can both optimize the energy output
and have a predictive maintenance scenario
so that you can ensure that you can avoid any downtime. And so each one of these pipelines can help you consistently
built, train, package and deploy machine learning
models to different windmills on the edge and
iterate as new data comes in, new telemetry signals, so
you can train and deploy new machine learning model
versions to make sure that you have the best
quality models in production. – Right, so this ML Ops-centric approach then makes a lot of
sense but what do you do in terms of getting the
service up and running for the first time? – So this an Azure service
so you would simply go to your list of services
in the Azure portal, log in just like in any other service. And then you can search
for Azure Machine Learning, you can select the workspaces here and if you’re just getting started you can easily click add and create a new Machine Learning Workspace. You can see I have a list of available workspaces here already. So I can simply select
one and then I can see that I have a list of
machine learning experiments that my team has been
working on, I have a list of compute targets, whether
it’s like Kubernete Services to deploy my models of Azure
Machine Learning Compute to be able to train my models. And I’ve got a list of
the models themselves as well so I can keep track of what’s deployed where and
how well it’s performing. Now, to get started on an
experiment or to continue working on the one I already
have, I even have lists of Notebook Virtual Machines
integrated directly here so those are like Cloud
workstations if you will. So I can log directly
into the Jupyter notebook from the portal here or
access the URL directly and open a notebook that
I’ve already been working on. You can see I’ve got a
bunch of open source Python libraries like pandas and numpy etc., and the Azure Machine Learning Python SDK which I can install and
use on my local laptop or use this Cloud workstation as well. – Right, and the nice thing
is the notebook environment’s really perfect for
collaboration and it’s good to see that Azure Machine
Learning effectively plugs right into that and it’s a tool that I think most data
scientists are used to. But can you walk me through
then a real scenario and get some experiments and
some models up and running? – Sure, so we’ve got
Contoso Auto as a fictional car manufacturer, right? And they are collecting IoT
data from all their vehicles that they have deployed
all around the world. Customers have bought and they’re driving but they’re getting a lot of calls, a higher volume than normal of complaints that they’re having issues. Now we can drill into
this Power BI dashboard that I have here and we can
see that for the different vehicles that I have out in the field, I’ve got an issue with the starter motor, there’s already a recall in progress and I’ve got an issue with batteries so I’m gonna drill in
here to see some issues and I can see that I’ve
got a higher than expected rate of failure for these batteries. So I’ll show you how to
use Azure Machine Learning to predict what kind of
failures are likely to occur so we can notify customers
proactively to bring their vehicles in for
maintenance before issues happen. – Okay, so now then you can
use these these data points effectively and like
machine learning in general it’s really doing some
predictions here in terms of customers in this case
that might have a likely failure so that we can call
them before they end up on the side of the road
needing to call a tow truck which’ll be even more expensive for us. – That’s absolutely right. So to get started with
Azure Machine Learning for this scenario,
first we need to connect to the Azure Machine Learning workspace and again that’s the thing
that’s really bringing together my compute and data
and everything all together. I’ll create an experiment
so I can iterate over and train multiple models and keep track of their quality score,
and I’ll get access to some of the data that I have available, that telemetry data that I mentioned. And I can register those data sets in Azure Machine Learning to make them really easy for my data
scientists to find them. Then using standard approaches
in the Jupyter Notebook I can interactively explore
the data and get some intuition about what data might be useful to train a high quality model. – Cool, so this is now where
you might need to experiment with different models but
in classic machine learning with the experimentation that’s gonna take a lot of time, right? – It can, yeah, and typically
this could take days or maybe even weeks of experimentation. You’d have to take the
data set, across the age of the battery, the
temperatures, run time etc. and experiment with different algorithms and hyperparameters to train
the model and then repeat the process a bunch of times
because the scientific method is just kind of guess and check. – Right, right, right. – But now there’s a better way with Automated Machine Learning. – So can you explain
how automated ML works? – [Chris] Automated ML generates
different experiment runs using a combination of
different algorithms and hyperparameters and
trains models in parallel, returning a quality score for
each model after each run. Then based on what it learns, it’ll keep generating
different experiment runs with different combinations of algorithms and hyperparameters to try to
train an even better model. – So it’s actually then
using machine learning to select different experiments
and really automatically without machine learning
find the right model for you. – Exactly, so now as a data
scientist you can focus more on fine tuning the models,
versus a lot of the manual effort around trial and error. – But I can see where
this might get expensive from a compute perspective,
how does that work? – Well it depends. Some models use a lot of
compute cycles to train. However, this approach is more efficient because instead of your
data scientist manually trying different
combinations of algorithms and hyperparameters over and over, with automated machine
learning we can do this process in parallel in a fraction
of the time to save you even more money. So for most data scientists,
we’re coming from a world where everything’s running all the time and so I might run a job before I go home at night
for example and then the job might complete five
minutes after I leave– – Right and you’re leaving
that computer on all night. – Exactly, and so with AML
compute will automatically shut down the machines that
aren’t being used here. – So how’s this all working
then under the covers? – [Chris] When I run an
experiment, Azure Machine Learning will take these different experiment jobs and insert with into a queue, right? And then as the queue forms,
this kicks off automatic VM provisioning in the cluster. So the cluster starts of with
zero VMs in it, then it grows kind of on demand as
the workload increases. – [Jeremy] Makes sense. – [Chris] And so then we run as many jobs in parallel as we allow from
that compute quota limit perspective and then as
those VMs get saturated the remaining jobs stay in the queue, it’ll drain the queue
after jobs get complete, it’ll pick new jobs,
it’ll train the models all in parallel, and then
as we drain the queue completely then we’ll
automatically deprovision the virtual machines so
you don’t pay anymore. – So it’s truly elastic. That cluster had been paused effectively, it provisioned everything
out and then it contracted and paused again when you were done. – That’s right, and you still
have the full flexibility of configuring or choosing
what VMs queues you use, so you might use some
with GPUs for example to facilitate deep learning
jobs or pay for CPU VMs which are less expensive
if they’re more appropriate for your workload. – All right, so the
concepts make a lot of sense but can you show us this in action? – Yeah, so to run those
machine learning jobs in the AML compute cluster,
first I’ll grab a handle to this compute target which
has already been created. But I can create it from the
Python SDK here in Jupyter if appropriate, and then
I’ll create an automated ML config file that specifies
that this is a regression task, because I’m using
that data to predict how far in the future the
battery is likely to fail. And I can set some constraints
to control my costs, like how many models or
iterations do I wanna run and how many of those
should run in parallel, how long each of them should run for so I can control my cost. – [Jeremy] It’s like your
config parameters file, yeah. – [Chris] Yeah, exactly, and
then I’ll run this by calling it experiment.submit and
then I’ll grab a handle to the run object and then
I’ll visualize it directly here in the Jupyter Notebook and I can see that Azure Machine Learning ran a
bunch of these in parallel. I can see the quality scores and details about each one of them with a nice visualization. I can see that there’s a clear winner here based on the
metrics that I’ve chosen. – Right, right and the darker
green meant that was the best model run so now how do I go
about taking it to the next step where I wanna package
them up and get them deployed? – So this is one of the best parts of the Azure Machine Learning service. It’s super easy to
operationalize your models and deploy them in production. Here I can take this machine learning run, I can grab the best one of
those iterations automatically and then I can publish that
to the Azure Machine Learning model registry just by
calling it register model and I can give it some
tags and descriptions and then when I do that
I’ll actually see that in the workspace here, I
can see the list of models. And so this is the model
registry and so this is how I know what models my
team has and what models I might want to deploy and even see where some characteristics about each one of those and
see where they’re deployed. So in this case this one is deployed to an Azure Kubernetes service. I can even drill into the
details of that Kubernetes service deployed model
where I have a REST API. I can get the scoring URI,
I can recycle the keys for security purposes if I want
to, and that sort of thing. But let me show you how easy it is for a data scientist to
be able to deploy this. You simply create a scoring
file which runs the model and then I can package that
up with my dependencies that I need and this will
generate a docker container for me that I can then deploy
to my Kubernetes cluster by using the inference config. And just specify my scoring script and my dependencies yml file and then call model.deploy. Once it’s deployed,
then I can test this out and I can get the
response from the REST API end-point that we automatically generate. And then I can integrate that
with my Power BI dashboard. So instead of just a list of
potentially affected vehicles, then I actually can get
the predictions by using that REST API end plane. Not only can I get the
predictions but I can even get the list of the predicted features that explain why that model
chose those particular vehicles so when I call
the customer to bring their vehicles in, then
I can explain to them what characteristics
picked them if you will. – So this looks really
powerful but what if I’m not familiar with things like Jupyter, I can’t write Python. Are there things that
I can do that are a bit easier to get up and running? – Yeah, we hear that a lot and that’s why we’ve created the new user interface for the Automated Machine Learning experiments to get started. So here you can see I’ve got
a history of the recently run experiments but I
can also simply click create experiment and this will create a new experiment run. Either in the same experiment
or a new experiment. I’ll select the same one
I was using before here and the same compute cluster
that I was using as well. I’ll click next and then I’ll
be able to choose the data that I was using and then
I’ll be able to get a quick preview of it, I can
select the type of task, again this is a regression
task, and I can select the column that I’m trying to predict because I want to find out how long it’s likely to survive in days, right? – [Jeremy] Right. – [Chris] And so I can
click start and through that kind of wizard-like
approach then we can train new machine learning models. However, while this is
running I can even go into some of the details
of recently run experiments where we can see some
of the characteristics and we can even easily
click deploy the best model just like I did in the
Jupyter Notebook earlier, which will register the model in the machine learning model registry. – Okay, so you can track
then everything in one place but what helps then with
things like reproducibility. You mentioned earlier
that we’re using things like ML Pipelines, how do those actually come into play here? – Yeah, let me show you
how we can use the same sort of approach that I
just mentioned and automate the process of training
unique machine learning model for each different type of car year, make, and model that I have. So we can see I just created
a list of the different makes and models, then I’ll
create my automl config file specifying the type of task again. And then I can create
these different steps in my pipeline to run
the automl experiment, collect the best model and the key metrics out of that, and then
assemble them together in a pipeline where I
append a bunch of steps. You can run these in
serial or in parallel. And this is super
important because sometimes in a machine learning
experiment there can be 10s, 100s or even 1000s of
unique steps that I need to run. And so by creating a
pipeline that I publish and then I can just run it on a schedule or by calling a REST API then it abstracts all of the complexity inside the pipeline so that I can just run that for each loop for all the different makes
and models if you will. – So this all totally
makes sense in the context of the Notebook. – Yep, and from the
Azure portal I can also see the inventory, the list
of pipelines that I have. I can schedule then to
run, I can see the details about them as well like who created them, their REST API end-point
so that I could call them and see all the steps
that are in each pipeline. – So a nice visible and visual
view here that you’ve got. Now you mentioned earlier
that you can also integrate with things like Azure
DevOps, how does that work? – It works in two different ways. Build Pipelines and release Pipelines. The build Pipelines integrate with Azure Machine Learning Pipelines to facilitate training
and building the model. You can think of that model
as like a build artifact. So here I have a build
pipeline that then runs the ML pipeline that I mentioned earlier. And then once that machine learning model is built and registered in
the Azure Machine Learning model registry, then that will kick off the release pipeline. So we can see ACI can either run when you check in maybe a new scoring script or a new version of the
machine learning model. And then we can have multiple tasks in this test so it can run and say Azure Container Instance
and then it can run a bunch of different integration
tests with my application as well as quality tests to make sure that the model is good. And then once that’s done
then it can role this out to my Azure Kubernetes
service, which will again deploy that but not until, for example, we’ve incorporated a human in the loop for deployment approvals to make
sure for auditory compliance, regulatory type of reasons
that we know exactly what we’re deploying and
have proper approvals. – Right, so this is a
release pipeline you can kind of do anything with it
and basically you can even have things get deployed out to the edge if you wanted to with the
same release pipeline. Now one thing that’s great
for our data scientists that are watching, and we
mentioned earlier the integration of things like, for example,
deep learning models with TensorFlow, and you can also use machine learning then to go
beyond the predictive models that we just showed here. What would we do then in terms of taking it to the next
level and maybe using some different data input
types to do deep learning? – So for example if you had
a camera installed in the car you might use that video
telemetry stream to then train a model to identify whether drivers are distracted for example. And so we can take those images, train a model using PyTorch or TensorFlow in the Cloud, make sure that it’s good, then use IoT Hub to roll it out to the individual cars
and then you can maybe raise an alert in the
vehicle if the driver is doing something that’s unsafe. – Maybe that they shouldn’t be doing, so really good stuff. So this is an awesome intro
in terms of all the different things that you can do with
DevOps as well as MLOps and all the Azure
Machine Learning services that we’ve just rolled out. Where would people go to
learn more on how to start using all of this stuff
and really consuming it? – Try it yourself, that’s
the best way to get started. You’ll learn more at the link shown. – Really great tips and this is a space that continually evolves and changes and of course you’re
gonna wanna keep watching Microsoft Mechanics for the latest tech updates across Microsoft. Hit subscribe if you haven’t already and we’ll see you next time. (upbeat music)

1 thought on “What is Azure Machine Learning service and how data scientists use it

  • Do you have a link to a tutorial, github repo etc. to reproduce the steps shown here? Condensed a lot of information into a short succint format, thanks!

Leave a Reply

Your email address will not be published. Required fields are marked *