Ritz here. You understand, I’ve been across the block a time or two in terms of working with object detection models. So once I heard about this hot latest thing called YOLO-NAS, I knew I had to provide it a whirl. And let me inform you, this bad boy doesn’t disappoint. It’s just like the Elon Musk of object detection models — daring, modern, and barely terrifying.

First up, we’d like to speak concerning the absolutely genius latest Quantization-Friendly Block they’ve cooked up. It’s like they checked out previous models and said, “These are cool, but what would make them even higher? A frickin’ tailor-made block for quantization!” After which they only went ahead and did it. It’s like they handed us the keys to a Lamborghini once we were expecting a Toyota Corolla.

Now, the training scheme for YOLO-NAS is just like the Rocky Balboa of object detection models — intense and seriously impressive. They decided to do some pre-training on the Object365 dataset, which is essentially the heavyweight champion of labeled objects. Then, they threw in some pseudo-labeled data like tossing some extra jalapenos in your nachos, simply to spice things up a bit. But they didn’t stop there. No, they took it a step further and decided to make use of knowledge distillation with a pre-trained teacher model. That’s like learning from the Yoda of object detection, and I’m all for it.

After the training, the YOLO-NAS team busted out their Post-Training Quantization (PTQ) magic like some form of tech wizards. They went ahead and converted the network to INT8, making it much more efficient. It’s like they took an already amazing cake and added an additional layer of delicious frosting only for good measure.

You is likely to be pondering, “Okay, Ritz, this all sounds great, but what concerning the architecture?” Well, my friend, let me introduce you to AutoNac. This bad boy optimizes the architecture space like a champ, and it does all of it while using the equivalent GPU time of coaching just five networks. It’s like they put the architectural optimization on a conveyor belt at a sushi restaurant — fast, efficient, and oh-so-tasty.

And as if that wasn’t enough, YOLO-NAS comes pre-trained on some top-tier datasets. We’re talking COCO, Objects365, and Roboflow 100 here. This implies you’re all set to crush it in downstream object detection tasks. Like, seriously, this model is able to rumble straight out of the box. It’s like getting a totally loaded gaming PC without having to spend hours assembling it yourself.

But wait, there’s more! YOLO-NAS has some wicked enhancements to detect small objects, improved localization accuracy, and the next performance-per-compute ratio. It’s like they took all the things we love about object detection and cranked it as much as 11, like turning the quantity up in your favorite song until the speakers are shaking.

And here’s the cherry on top: YOLO-NAS is ideal for real-time edge-device applications. So whether you’re working on a sick latest robot or simply attempting to make your phone do some crazy stuff (like, I don’t know, activate the blender from the opposite side of the room), this model has got you covered. It’s like having a multi-tool in your back pocket, able to tackle any challenge life throws at you.

Now, let’s talk numbers. In line with Deci, YOLO-NAS is around 0.5 mAP points more accurate and 10–20% faster than equivalent variants of YOLOv8 and YOLOv7. That’s right, folks. This model isn’t just a reasonably face; it’s got the stats to back it up. It’s like comparing a cheetah to a house cat — they’re each cool, but one is clearly faster and more powerful.

Here’s just a little rundown of the model’s mAP and latency stats on your viewing pleasure:

• YOLO-NAS S: 47.5 mAP, 3.21ms latency
• YOLO-NAS M: 51.55 mAP, 5.85ms latency
• YOLO-NAS L: 52.22 mAP, 7.87ms latency
• YOLO-NAS S INT-8: 47.03 mAP, 2.36ms latency
• YOLO-NAS M INT-8: 51.0 mAP, 3.78ms latency
• YOLO-NAS L INT-8: 52.1 mAP, 4.78ms latency

So, there you will have it. YOLO-NAS is the new latest object detection model that’s got everyone talking, and for good reason. If you happen to’re able to dive in and see what all of the fuss is about, just head over to GitHub, give the SuperGradients repo a star, and take a look at the starter notebook. It’s like unboxing the newest gadget and being the primary amongst your mates to point out off its awesomeness.

GitHub repo: https://lnkd.in/dpC8dnbA

Starter Notebook: https://lnkd.in/dqcrnDFH

AS-One Library — https://github.com/augmentedstartups/AS-One

In conclusion, YOLO-NAS is the article detection model you didn’t know you needed — until now. With its unique combination of modern features, optimized architecture, and incredible performance, it’s truly a game-changer on this planet of object detection. So, go ahead, give it a spin, and see what form of cool stuff you may create with the facility of YOLO-NAS. And remember, as the good Ritz at all times says, “When life gives you lemons, construct a lemonade-making robot using YOLO-NAS.” Or something like that.

So, now that you simply’re all overestimated about YOLO-NAS and able to jump in, let’s speak about a good cooler method to harness its power. If you happen to’re already acquainted with our YOLOv8 course, you’re in for a treat. With the upcoming release of our Modular AS-One library, you’ll give you the chance to simply swap out the YOLOv8 model for the shiny latest YOLO-NAS, taking your object detection projects to the following level. And guess what? It’s not nearly detection — the library will even cover training, so you may really fine-tune your projects like a mad scientist. But wait, there’s more! To be certain you don’t miss out on this incredible opportunity, head over to https://www.augmentedstartups.com/YOLO+SignUp and join to get notified when the magic happens. Trust me, you’ll wish to be the primary one to brag about your YOLO-NAS-powered creations. So, go on, embrace your inner Augmented Engineer, and let’s revolutionize the world of object detection together!

YOLOv8 course, which might be compatible with YOLO-NAS via the AS-One Library