Collaboration of Hugging Face & AWS SageMaker Brings Revolution to NLP Model Training

Overview

Natural Language Processing (NLP) is an ever-evolving discipline, and NLP models are expanding rapidly and in complexity. Due to sturdy ecosystem relationships with companies like Hugging Face and excellent distributed training features, Amazon SageMaker is one of the most accessible platforms to fast-train NLP models.

Define Natural Language Processing (NLP)

Natural Language Processing (NLP) is a branch of Artificial Intelligence that helps computers learn, analyze, and process human languages like Hindi and English so that they may study and deduce their meaning. In my previous blog, Amazon Machine Learning & Artificial Intelligence Services, I have discussed AI in detail.

What is Hugging Face?

Hugging Face is a firm that specializes in natural language processing (NLP). It provides model libraries that enable developers to create NLP models in a few lines of code with excellent accuracy. Hugging Face offers a variety of well-known models for their effectiveness and ease of execution. Here you can find more information about Hugging Face pre-trained models. 

AWS SageMaker

Amazon SageMaker is a machine learning service that Amazon wholly manages. Data scientists and developers can use SageMaker to construct and train machine learning models fast and efficiently, then deploy them directly into a production-ready hosted environment. You do not have to manage servers because it has an integrated Jupyter publishing notebook instance to access your data sources for exploration and analysis easily.

Hugging Face is collaborated with AWS SageMaker to make NLP easy & accessible for all. SageMaker makes easy to make & deploy Hugging Face models. Amazon SageMaker offers high-performance resources to train & use NLP Models.

Following are the advantages of collaboration –

• Hugging Face and Amazon unveiled new deep learning containers (DLCs) to make training the hugging face transformer in Amazon SageMaker easier than before.
• A hugging face extension has been added to the SageMaker Python SDK to help the data science team save time by lowering its time to set up and execute studies from days to minutes.
• Amazon SageMaker’s ability to automatically tune models. To adjust the training hyperparameters automatically and quickly improve model accuracy
• Training artifacts and experiments can be compared and tracked using Amazon SageMaker’s web-based IDE.
• Customers who use Hugging Face & AWS SageMaker Containers benefit from built-in performance optimization for TensorFlow and PyTorch.

Hugging Face deep learning containers are entirely integrated with Amazon SageMaker distributed training libraries, allowing the model to be trained quickly before employing the next generation of EC2 instances. Hugging Face & SageMaker also has an example gallery where you can get ready-to-use, high-quality embracing face scripts on Amazon SageMaker. On Hugging Face, SageMaker provides an integrated process for generating models. You can write your script in either a notebook or studio instance.

Using AWS SageMaker’s new hugging face estimators, you can quickly train, fine-tune, and optimize Hugging Face models created with TensorFlow and PyTorch.

Get more Insights into AWS SageMaker Studio and Its Popular Features.

Fine-tune the model and start a SageMaker training job

We will need a Hugging Face Estimator to do a SageMaker training job. The estimator oversees all aspects of Amazon SageMaker training and deployment. We specify which fine-tuning script should be used as an entry point, which instance type should be used, which hyperparameters should be sent in, and so on in an Estimator:

When data scientists call appropriate methods on estimators, the SageMaker control planes create an instance specified in the estimator. The Hugging Face deep learning container is copied into the training instance when the model is up. The data in S3 or EFS are copied into training instance training and then start near real-time logs files & metrics copies into AWS CloudWatch. Once training completes, model artifacts are copied into S3 & training instance gets destroyed.

NLP data sets could be massive, leading to a very long training time to speed up the training jobs. Hugging Face’s transformers library works with AWS SageMaker’s data parallelism library.

The distributed model parallel library automatically & effectively split the models into multiple GPU & instances.

Conclusion

SageMaker’s collaboration with Hugging Face has resulted in a significant shift in the NLP domain, making it easier to train NLP solutions. Businesses can more easily integrate the power of NLP into their services using Hugging Face pre-trained language models and AWS SageMaker resources. In the next blog, you will learn how to use the Hugging Face transformers library in PyTorch or TensorFlow and AWS SageMaker’s distributed training libraries to fast train an NLP model.

About CloudThat

CloudThat is the official AWS Advanced Consulting Partner, Microsoft Gold Partner, and Training partner helping people develop knowledge on the cloud and help their businesses aim for higher goals using best in industry cloud computing practices and expertise. We are on a mission to build a robust cloud computing ecosystem by disseminating knowledge on technological intricacies within the cloud space. Our blogs, webinars, case studies, and white papers enable all the stakeholders in the cloud computing sphere.

If you have any queries about Amazon SageMaker, Natural Language Processing, Hugging Face, or anything related to AWS services, feel free to drop in a comment. We will get back to you quickly. Visit our Consulting Page for more updates on our customer offerings, expertise, and cloud services.

FAQs

Q1. Is AWS SageMaker Python SDK necessary to use the Hugging Face Deep Learning Containers (DLCs)?

Ans: No, you can also use different SDKs like boto3 or AWS CLI (Command Line Interface) to deploy your Hugging Face DLC models. DLCs are also available via Amazon ECR and can be retrieved and utilized in any environment.

Q2. What are the advantages of using Hugging Face Deep Learning Containers (DLCs)?

Ans: The DLCs are thoroughly tested, efficient deep learning environments that do not need to be installed, configured, or maintained. The inference DLC (deep learning containers) consists of a pre-providing layer, reducing the technical barrier for DL serving dramatically.

Q3. How do models deploy for Inference?

You can launch an inference task with SageMaker using your trained Hugging Face model or one of the pre-trained Hugging Face models. You can deploy your trained and pre-trained models with SageMaker using this cooperation using just one line of code.

WRITTEN BY Shubham Dubey

Shubham Dubey works as a Sr. Research Associate at CloudThat. He has 3+ years of experience in AI/ML. He is highly passionate about learning new skills and technologies.