Hi everybody, this is Julien from Arcee. Welcome to episode 14 of my podcast. I hope you're all safe. Stay home if you have to and use that time to watch all my videos and read all my blog posts. Don't forget to subscribe to my channel to be notified of future videos. In this episode, I'm going to go through AWS news, cover some of the latest announcements on machine learning, talk about Amazon Personalize and Elastic Inference, and a few more things. Okay, let's get started.
The first news is Amazon Forecast is now available in three new regions: Sydney, Mumbai, and Frankfurt. Forecast is a high-level service that lets you easily build forecasting models for time series data. If you've never looked at it, I recommend it. It's really super easy to use. Just upload your data in CSV format to S3, and then in just a few clicks or a few API calls, you can build forecasting models for supply chain, time series, or inventory metrics. Anything that's time series can be easily used with Forecast. So, a nice service and now customers in those three regions can use it locally. That's always nice.
The next news item is actually a combination of services. Now you can use Amazon Personalize with Amazon Pinpoint. I don't think I've ever talked about Pinpoint, so maybe I should explain first. Pinpoint is a service that lets you manage engagement campaigns over a variety of channels—email, SMS, voice notifications, etc. So you can push messages to your audience and measure engagement data. This service has been available for a while, and now you can actually seamlessly use Personalize with it. What this means is you can reference an Amazon Personalize campaign directly in Pinpoint. So, if you want to send personalized notifications to your audience with Pinpoint, you can easily do that now because you can build a personalization model with Amazon Personalize and plug it into your Amazon Pinpoint campaigns.
This is actually based on a solution. A solution is an architecture that has been designed and validated by our solution architects. You can easily deploy it using a CloudFormation template. So you can view the deployment guide and launch it directly. This CloudFormation template is going to create all of those resources inside your AWS account. It will provision Pinpoint resources, Personalize resources, Kinesis, etc. Solutions are really a super easy and fast way to deploy an architecture that solves a specific problem. You don't have to mess with anything; just review the template, run it, and all this will be ready in minutes. So you can read all about it. As a matter of fact, we have quite a few solutions out there. You can find them on the AWS website, and here I filtered for just the AI ML solutions. We can see the Pinpoint one and more. Some of them are 100% based on AWS services, and some are also based on partner solutions, such as fraud detection for machine learning, machine learning for telcos, etc. I would really recommend that you look at those because you might find that one of them is already really close to a problem you're trying to solve. You can just go and test it, and then, of course, you can adapt it to your own needs and customize it if you have to. But at least you won't be starting from a blank page. So solutions are nice, and of course, I will put the link to all of this in the description. That's it for Personalize and Pinpoint and all those nice AI ML solutions.
Now let's talk about SageMaker. There's always something happening on SageMaker. This week, we have new features on SageMaker Ground Truth. Ground Truth is a capability of SageMaker that makes it easy to annotate datasets at scale. I've covered Ground Truth in great detail in previous videos, so please go and check out that demo. The new feature here is that you can now launch multi-label tagging for images and texts. Previously, if you had to annotate a dataset with different labels, you had to go through multiple rounds to add the different labels. Now you can actually define the list of labels for images or text and get all that done in a single Ground Truth job. That's pretty cool. It's a great simplification for customers. If you've never seen Ground Truth, it lives in the SageMaker console. You can create labeling jobs, labeling datasets, and workforces, which could be private, third-party, or Mechanical Turk if you need to scale really high. And this is an example of a segmentation job. I think this is the one I used in my demo where I go through a bunch of images and do semantic segmentation on guitar players. If you want to know how to do this, please go and check out those other videos. With this new feature, when you create a new job, you'll see in the task type that you can create multi-label image classification jobs and multi-label text jobs as well. Very nice simplification. I really like Ground Truth. I think it's a really cool service. It solves a really hard problem, which is, "Hey, I've got thousands or tens of thousands of images to annotate. How do I get that done?" Well, Ground Truth is how. And it keeps improving, so that's great.
Okay, let's talk about frameworks. What do we have here? Well, I think last week we announced updated deep learning containers with the new frameworks. And, of course, not to be outdone, we have deep learning AMIs with the same updated frameworks. So the latest TensorFlow 2.1, latest PyTorch, and latest MXNet. Again, if you've never looked at the deep learning AMIs and you keep baking your own AMIs, please do yourself a favor and check the deep learning AMIs. We maintain them constantly, update them constantly, and just save you the hassle of building those AMIs and installing those deep learning frameworks, which are not always super easy to install, and the Nvidia drivers for GPU instances, etc. Deep learning AMIs and deep learning containers, which I've already discussed, are your friends. Give them a try. I think they're a huge time saver.
And well, we see this last bit here. Elastic Inference with PyTorch. This is actually the next feature I want to talk about. Elastic Inference was launched at re:Invent 2018, and I keep talking about it because I think it's a fantastic service. So let me explain once again, because I know some of you are not familiar with it, why it's so important. Some models are too heavy, too complicated to be deployed on CPU instances. Sure, you can deploy them, but they're really slow. Prediction is going to be slow. So what do you do? Well, of course, you deploy to GPU instances, and everything is super fast and fantastic. But let's be honest, GPU instances are a little more expensive than CPU instances. And sometimes you don't get the most bang for your buck because you use them by default, and if you monitor GPU usage, you realize maybe you're just keeping that GPU busy 10% or 20% of the time. Things are nice and fast, but you're paying for a full-fledged instance and not keeping it super busy because your model is not that complicated or not that crazy, or maybe you just don't send enough traffic to it. So it's not a great situation. You have to choose between performance and cost optimization. Well, that's exactly the problem that Elastic Inference is solving.
Elastic Inference lets you use fractional GPU acceleration. So you can pick between three sizes—medium, large, and xlarge—and you get a certain number of teraflops for each size. Now you can attach this Elastic Inference accelerator to any EC2 instance. It could be your EC2 instance, the one you manage, or a SageMaker managed instance. You have the choice. And now you'll be able to find the right combination of CPU and GPU acceleration for your application. You can run your benchmarks and find the sweetest spot for cost and performance. Elastic Inference is a really great feature. You can save up to 70% or 80% compared to using full-fledged GPU instances. So please, if you're using GPU instances today and you've never looked at Elastic Inference, or you've never benchmarked them, please give them a try. You might just be able to go back to your boss or your CFO and say, "Well, we just saved 70% on our GPU workloads." That's a very nice number, so give it a try. This was available at launch for TensorFlow and MXNet. We actually added extra APIs to those frameworks so that you can use them on your EC2 instances. And, of course, this was integrated into SageMaker. Now you can do the same with PyTorch. PyTorch is super popular. It is a really cool library. So there was no reason for PyTorch users to be left out of the Elastic Inference party. Now you can use Elastic Inference with PyTorch and save time and money and buy more beer. How cool is that?
All right, that's it for this episode. I hope you learned a few things. Until next time, please stay safe and keep rocking.