HuggingFace

HuggingFace was originally developed as a chatbot app for teenagers in 2017, but it has since grown into a comprehensive platform that allows users to host AI models, train them, and collaborate with their team on projects. From coding to deploying AI in live applications or services, Hugging Face offers all the necessary tools. Additionally, users can explore and discover models created by other individuals, search data sets, and try out demo projects.

One of the standout features of HuggingFace is the ability to build your own AI models. These models can be hosted on the platform where you can add more information, upload essential files, and keep track of different versions. Moreover, you have full control over whether your models are public or private, giving you complete autonomy over when and how you share them with others.

Another convenient aspect of HuggingFace is its built-in discussion feature on model pages. This facilitates easy collaboration and allows contributors to suggest modifications through pull requests. Once your model is ready for use, there’s no need to find another hosting platform—you can simply utilize it directly on HuggingFace by making requests or pulling results for the apps you’re developing.

When you become a member of HuggingFace, you gain access to a hosted repository based on Git. This repository allows you to store your Models, Datasets, and Spaces. In the field of machine learning, the training of an AI model involves using a dataset that contains labeled instances. These labels provide instruction to the model on how to interpret each sample.

As the model processes the dataset, it begins to learn the relationship between examples and labels, identifying patterns and word frequencies. Eventually, you can test the model by providing it with a prompt that was not included in the original dataset. By doing so, you can observe how well the trained model performs. The output generated will be informed by the knowledge gained during training.

Creating a high-quality dataset requires significant time and effort due to its necessity for accurately reflecting reality. Without this accuracy, there is a higher chance of hallucinations or undesired consequences in the model’s outputs.

Fortunately, Hugging Face simplifies the training process by offering access to over 30k datasets that can be utilized with your models. Additionally, as an open-source community platform, you have opportunities both to upload your own datasets and explore more accurate ones published by others.

As with other AI models such as natural language processing, computer vision or audio, HuggingFace also includes datasets to be trained for each specific task. The contents change based on the task: text data, for example, is used by natural language processing, images by computer vision, and audio by audio data.

HuggingFace has recently introduced Inference Endpoints, a solution that addresses the challenges of deploying transformers in production. This managed service enables users to deploy their models on HuggingFace Hub and various cloud platforms like AWS and Azure (with GCP support coming soon).

Additionally, it offers flexibility in choosing instance types, including GPU options. The company itself is transitioning some of its ML models from CPU-based inference to this new service. In this blog post, we will explore the reasons behind this decision and discuss why you may want to consider leveraging Inference Endpoints as well.

The company recently transitioned to using Inference Endpoints for managing our models. Previously, we had internally managed them on AWS Elastic Container Service (ECS), supported by AWS Fargate. This allowed them to have a serverless cluster for running container-based tasks.

They have found HuggingFace Inference Endpoints to be a simple and convenient solution for deploying transformer and sklearn models into an endpoint that can be easily accessed by applications. While it does come at a slightly higher cost compared to our previous ECS approach, the time saved on deployment is well worth it.