Sunday, August 27, 2023
TensorFlow is an open-source platform for creating and using machine learning models. It implements many of the common algorithms and patterns needed for machine learning, saving you from needing to learn all the underlying math and logic and enabling you to just to focus on your scenario. It’s aimed at everyone from hobbyists, to professional developers, to researchers pushing the boundaries of artificial intelligence. Importantly, it also supports the deployment of models to the web, cloud, mobile, and embedded systems
The process of creating machine learning models is called training. This is where a computer uses a set of algorithms to learn about inputs and what distinguishes them from each other. So, for example, if you want a computer to recognize cats and dogs, you can use lots of pictures of both to create a model, and the computer will use that model to try to figure out what makes a cat a cat, and what makes a dog a dog. Once the model is trained, the process of having it recognize or categorize future inputs is called inference.
TensorFlow there are three main ways to do this: you can code everything by hand, where you figure out the logic for how a computer can learn and then implement that in code (not recommended); you can use built-in estimators, which are already-implemented neural networks that you can customize; or you can use Keras, a high-level API that allows you to encapsulate common machine learning paradigms in code.
There are many ways to train a model. For the most part, you’ll probably just use a single chip, be it a central processing unit (CPU) , a graphics processing unit (GPU), or something new called a tensor processing unit (TPU). In more advanced working and research environments, parallel training across multiple chips can be used, employing a distribution strategy where training spans multiple chips. TensorFlow supports this too.
Beyond creating models, you’ll also need to be able to get them into people’s hands where they can be used. To this end, TensorFlow includes APIs for serving, where you can provide model inference over an HTTP connection for cloud or web users. For models to run on mobile or embedded systems, there’s TensorFlow Lite, which provides tools for model inference on Android and iOS as well as Linux-based embedded systems such as a Raspberry Pi. A fork of TensorFlow Lite, called TensorFlow Lite Micro (TFLM), also provides inference on microcontrollers through an emerging concept known as TinyML. Finally, if you want to provide models to your browser or Node.js users, TensorFlow.js offers the ability to train and execute models in this way.
