The Hugging Face Toolkit An Introduction

Transcript

The main project is the Transformers project from Hugging Face. It's one of the fastest-growing open-source projects ever. You can see the GitHub stars for Transformers, which is the yellow line on the left-hand side. We're all very humbled by this. We thank the community for this and really intend to keep the good stuff coming. It's not just a community; we see analysts catching up too. If you read the latest State of AI report, you probably saw that transformers are emerging as a general-purpose architecture for ML. It's really not just NLP or deep learning anymore; it's becoming the de facto standard for ML. This is confirmed by the Kaggle survey, which shows that traditional models like convolutional neural networks and recurrent neural networks are increasingly less popular, while transformers are increasingly more popular. We have over 1 million model downloads from the Hugging Face Hub every day. Lots of folks use us, and we're very happy and humbled. This shows the value of the transformer architecture. We have a whole portfolio of projects that we work on. We host datasets and models on the Hugging Face Hub, our website, which you will see in a minute. From there, you can use different solutions for training, either on your own machine using the Transformers library or with our open-source library called Optimum for hardware acceleration on training and inference. We have an AutoML product called AutoTrain. You can then move your models to Spaces, a cool way to build and demo your machine learning models in web applications. We also have Infermware, a managed hosting service for all models. If you prefer to work in the cloud, we have a partnership with AWS, which makes it easy to train and deploy models on Amazon SageMaker. As of a few weeks ago, we also launched a service on Microsoft Azure, where you can one-click deploy any NLP model from the hub to managed infrastructure on Azure. This is the family picture. If you're curious about the other stuff, feel free to visit our website, huggingface.co. The Hugging Face Hub lives at huggingface.co. You can sign up in a minute by providing a username, email address, and password. It's all free to use. On the Hub, we have models for a wide range of tasks, from natural language processing to computer vision, audio, and even reinforcement learning. Each model is managed in a Git repository, which you can clone using the Git workflow or use with our open-source libraries. The model card provides information on what the model is, what it's been trained on, metrics, and other relevant details. We also have similar information for datasets. The Hub includes tags that help you find models for specific tasks, like summarization in French. The Hub is user-friendly; just create an account and try it out. Imagine you need to demo this to a marketing director or a business person with no technical background. They might be confused and say, "All I wanted was a simple web application where we can paste customer content and get results." They don't want to see a Jupyter notebook full of code and numbers. This is where Spaces comes in. Spaces lets you build a web app using a web framework like Gradio, which is now part of Hugging Face, or Streamlit. You build a simple app, push it to a Git repo, and it fires up an application on managed infrastructure. You don't need to worry about servers or anything else. For example, a simple UI can be created in just 15 lines of code. You import Gradio, the pipeline API, create a text classification pipeline using a fine-tuned model, and build a basic interface with a text input, text output, and a button to trigger the prediction. The most complicated part is often extracting the label and converting it into a user-friendly format, like displaying stars instead of floating-point numbers. This is what it took: 15 lines of code, creating a space, cloning a repository, adding and committing your app file, and pushing it. Now, when you show this to stakeholders, it looks friendlier and more relatable. You can make it look much nicer and more complex, with multiple models collaborating to build a richer UI. This approach helps you quickly get feedback and iterate on the original business problem. For example, given a product description in English, you can easily get a result on customer satisfaction. We started with a dataset and a model, prepared the dataset using the datasets library, trained the model, deployed it locally and with the inference API, and built a small web app in 15 lines of code. The time it takes to do this depends on your familiarity with the libraries. On your first try, it might take a week, but on your second try, it could take just a day because you can reuse a lot of code. Training time will be the longer bit, but if you work with off-the-shelf models, you can skip training and focus on building and deploying your app to get user and customer feedback. We're building a collection of tools that lets you iterate quickly and efficiently. Machine learning projects should be done in days, not months or years. They should be highly iterative. For more information, visit huggingface.co. We have a Transformers course that goes deep into the libraries I showed today. You can ask questions on our forums, and we also offer commercial support for private deployments and expert assistance. Feel free to reach out if you have questions or want to know more.