AWS AI Machine Learning Podcast Episode 16 AWS news

Transcript

Hi, this is Julien from Arcee. Welcome to episode 16 of my podcast. Please subscribe to my channel to be notified of future videos. I hope you're doing okay in these difficult times. I hope you're safe, you have plenty of food and entertainment. As you can see, everything's perfectly okay here. I made some new friends. Not much of a conversation with these guys, but they're generally pretty friendly so far. So let's see how it goes. Anyway, this week I'm going to go through some AWS news and announcements. So let's get started. What do we have this week? Let's start with a new feature in Amazon Textract. Textract is a high-level service that lets you accurately extract text and structure from forms, documents, etc. Now it's even more accurate for checkboxes and selection elements. Let's look at an example. This is the kind of document you would want to use with Textract. It's actually a health insurance form, so it has a lot of boxes as you can see. The purpose is to detect whether a box has been ticked or not. Textract can find those boxes and figure out if a box has been selected or not. This will reflect in the information you get from the JSON API. This is a cool feature because this is a super popular use case for Textract. Okay, let's look at the next one. The next one is one of my favorite services, Amazon Polly. Polly is text-to-speech. A while ago, we launched a feature called the Newscaster style. Let me explain in case you missed it. The Newscaster style is the ability to apply a news style to the speech generated from text. This is made possible by a new text-to-speech engine in Polly called the Neural Engine, where the waveform you listen to is generated by a deep learning network. Since it's generated, we can apply styles as well. A while ago, we launched a newscaster style for English, and now it's available for US Spanish. Let me show you how to use it. It's super simple. You would go to the Polly console or use the API and need to use SSML syntax. The syntax is `` and then the `` tag, saying we want to apply the style of a newscaster. We select the neural engine and the language. The supported languages are British English, US English, Brazilian Portuguese, and US Spanish. Make sure you use the SSML syntax and close the tags, and use the neural engine. Let's try an example: "En esencia, algo pareció cambiar en Lionel Messi el verano pasado en la Copa América. No cambió el resultado y Argentina se marchó de Brasil sin el título. Pero el niño tímido y silencioso, ese que se ponía nervioso en una charla frente a sus compañeros y que por rebeldía o pereza no cantaba el himno, se descubrió como un tipo cercano." I don't speak Spanish. This is a news article from El País, a leading Spanish newspaper. As you can hopefully hear, not only is the speech extremely natural and lifelike, but it's also dynamic, something you would expect to hear on the radio or TV because we have this style applied to the text. That's the new Polly text-to-speech feature for Spanish. Pretty cool. Let's move on. The next thing I want to talk about is a whole bunch of new tutorials for D-Plants. Remember D-Plants? Here's one. Did you forget about D-Plants? It's a really cool device. It's a tiny computer vision device with an Intel board and a camera, connected to AWS. You can train a computer vision model in the cloud, maybe on SageMaker, and easily deploy it to the camera using Greengrass. Then, you can run predictions on the video stream captured by the camera, all happening locally. You can send information through IoT or another service back to the cloud to say, "Hey, these are the predictions I made, these are the objects I detected," etc. DeepLens has been out for a while, and I've spent quite a bit of time writing about it and presenting it. Now, we have a new website for DeepLens where we show you how to get started and added a whole bunch of new projects, from simple ones to more advanced projects like worker safety, coffee counter, and a trash sorter. You can go and experiment with these new projects. I guess a lot of us have some time to do that at the moment. If you have a DeepLens camera and haven't looked at it or played with it for a while, it's a good opportunity to spend more time with D-Plants and learn more about computer vision and architectures that work well for these kinds of problems. I think that's pretty cool. Let's move on. Of course, we're going to talk about SageMaker. The news on SageMaker this week is that you can now train using G4DN and C5N instances. There's a long list of instances that SageMaker supports for training and deployment, and we just added these two fairly recent instance families. G4 is a GPU family, powered by the T4 GPU from NVIDIA. The D thing means it has fast local storage in the form of NVMe SSD storage. If you have datasets and train and copy those datasets locally to the training instances, which is the nominal scenario for SageMaker, the local I/O on the instance is going to be extremely fast. I've benchmarked those NVMe SSDs a while ago, and they are blazingly fast. If you want to save time, they're a good option. The N thing means enhanced networking. On the larger GeForce, this means you can go all the way up to 100 gigabits of network bandwidth. If you do distributed training or stream the dataset to the training instance using pipe mode, that's a pretty sweet network bandwidth to work with. So, G4DN, T4 GPU, local NVMe storage, up to 100 gigabit networking. These pack quite a bit of a punch. The C5N instances are the latest evolution of the C5 family, compute-optimized with Intel Skylake chips. Again, the N extension means up to 100 gigabit networking. A good fit if you use CPU instances for distributed training, you'll save quite a bit of time thanks to the increased bandwidth. What do we have next? Well, of course, we have yet another update to the deep learning containers. These containers are AWS-maintained containers for PyTorch, TensorFlow, and Apache MXNet. They come in a CPU configuration or a GPU configuration. You can use these containers as is to train or predict and use them on SageMaker. No need to maintain your own containers; we do that for you. I strongly encourage you to use these. They are a nice time-saver and free to use. You just pay for the compute instances you train or predict on, but the containers are free. Here, we're updating for PyTorch 1.4 and MXNet 1.6, and I'm sure we'll be back at some point with PyTorch 1.5 and MXNet 1.7. That race never stops. That's it for this episode. Again, please subscribe to my channel for future videos. Plenty of really cool stuff coming in the next few weeks. Please stay safe wherever you are. As for me, well, I'm ready for anything. If those guys in the background start creating trouble, I'm kind of ready. All right. Enough comedy. Stay safe, and I'll see you soon. Until then, keep rocking.