How to Build Better AI Models with a Production-Aware Approach and NAS

Given the impact of your hardware and inference environment on the performance of your model, a production-aware approach in the model selection and development is crucial.

?? What is production-aware model development?

Production-aware model development actively considers the different types of inputs, production settings, and performance targets throughout the development process. By designing your model for the target inference hardware and production environment, the success rate in production increases.

Unfortunately, going through model selection manually is challenging. It can lead to a long development cycle and a lot of manual work.

This brings us to neural architecture search.

?? Neural architecture search: What is it and its limitations??

Neural architecture search, or NAS, is a technique that can help you discover the best model for a given problem, hardware, and task. It’s an algorithm that can automate the model design and selection of deep neural networks, ensuring better accuracy and performant speed than manually designing architectures.

Now, if it's so great, why isn't everyone using NAS?

NAS is very time-consuming. It can take running several high-end GPUs for weeks at a time, resulting in very extensive computation. Even if you have the time, money to spend, and access to all these GPUs, you still need a high level of expertise to run NAS correctly. This makes traditional NAS difficult and inaccessible.

?

?? What’s the solution to the challenges of using NAS?

AutoNAC is Deci’s proprietary engine for neural architectural construction.

What makes AutoNAC unique is that it does NAS very fast. It can run one problem between two to three days, making it much more affordable and commercially accessible.

How does it work? You come with three main inputs: task, inference environment and hardware, and data characteristics. AutoNAC doesn't require data because data is private and usually stays on your side, but rather the characteristics of the data. If it’s object detection, for example, what is the distribution of the bounding boxes in the images? All these different types of data characteristics are used to create a proxy data set which is then used in the engine.

The AutoNAC then runs, 10 to the power of 19 different potential architectures. At the end of the process, it not only chooses one correct architecture, but also takes parts and building blocks from different types of architecture and generates a completely new architecture that has never been created before.?

The output is a PyTorch file, which is a model without any weights, and then it's trained completely on your premise.

?

?? How to get started with building your model with AutoNAC?

There are two ways that AutoNAC is served to the public: foundation models and custom modes.

Foundation models are state-of-the-art models created with AutoNAC. They come with a baseline recipe for the given task so that lowers all the risk of hyperparameter tuning. If it's appropriate for your use case, you can take it off the shelf and just run it for your environment.

Examples of foundation models that are the fastest and most accurate in their respective spaces:

• YOLO-NAS, object detection on NVIDIA GPUs.
• YOLO-NAS Pose, pose estimation on Intel Xeon CPU & NVIDIA Jetson.
• DeciSegs, semantic segmentation on NVIDIA GPUs.

Custom models, on the other hand, are for specific use cases that are not supported by the existing foundation models. The AutoNAC works with your task on your hardware. It supports any type of hardware. All it needs is access to the hardware, whether it's remote or direct access. You also get a custom recipe, in addition to your architecture and model.

Interested in learning more? Talk with our experts.

?? Get ahead with the latest deep learning content

• 苹果 researchers introduce AIM, a collection of vision models pre-trained with an autoregressive generative objective. It demonstrates that autoregressive pre-training of image features shows comparable scaling characteristics to its textual equivalent, such as LLMs.
• Researchers introduce InRanker models, which are distilled rankers designed to enhance the effectiveness of zero-shot retrieval. The primary concept revolves around utilizing large models to generate synthetic data extensively from the collections that will be employed during inference.
• An International Monetary Fund report shares recommendations for achieving an inclusive AI-driven world. Advanced economies should focus on AI innovation and integration with robust regulatory frameworks. Emerging markets and developing economies should prioritize building a strong foundation through investments in digital infrastructure and a digitally competent workforce.
• Runway , a generative AI startup, enhances its Gen-2 foundation model. The update includes a new feature called Muti Motion Brush, enabling creators to incorporate multiple directions and types of motion into their AI-generated video content.
• LangChain Benchmarks for evaluating the practical abilities of LLMs. The framework includes basic tool usage, complex information extraction, and advanced retrieval-augmented generation tasks.

??? Save the date

[Live Webinar] How to Ship Computer Vision Models to Production Faster with Better Performance | Jan 30th

Fast and efficient inference plays a key role in the success of computer vision-based applications, especially when there’s a strict performance requirement such as in autonomous vehicles and IoT-enabled and mobile devices. In this webinar, learn how to achieve real-time inference to deliver the best user experience.

Save your spot!

[Technical Session] Fine-tuning LLMs with Hugging Face SFT ?? | Jan 31st

While LLMs have showcased exceptional language understanding, tailoring them for specific tasks can pose a challenge. In this session, discover the nuances of supervised fine-tuning, instruction tuning, and the powerful techniques that bridge the gap between model objectives and user-specific requirements.

Save your spot!

?? Quick Deci updates

The past months have been quite eventful for us at Deci!

• Launch of DeciCoder-6B and DeciDiffusion 2.0. Both models are optimized for the Qualcomm Cloud AI 100 solution. DeciCoder-6B has fewer parameters in comparison to its counterparts, resulting in a reduced memory footprint and freeing up an extra 2GB of memory compared to CodeGen 2.5 7B and other 7B parameter models. Meanwhile, DeciDiffusion 2.0 generates superior image quality with 40% fewer iterations and employs a smaller and faster U-Net component compared to Stable Diffusion 1.5.

• Deci’s models on Microsoft Azure AI Studio. Thrilled to be working with Microsoft and the Azure AI team on enabling more people and organizations to use our highly accurate and cost-efficient models – DeciLM, DeciCoder, and DeciDiffusion.

Enjoyed these deep learning tips? Help us make our newsletter bigger and better by sharing it with your colleagues and friends!