Working with Models from Hugging Face Hub

Working with Hugging Face

Jacob H. Marquez

Lead Data Engineer

Introduction to the Transformers Library

Simplifies working with pre-trained models

Transformers GitHub repo

¹ https://github.com/huggingface/transformers

How Pre-Trained Models Are Used

Analyze text, images, and audio
Summarize documents and recognize speech
Enable autonomous driving by detecting objects

Text, images, and audio

autonomous driving.jpg

Navigating the Hub

Hugging Face Hub top navigation

¹ https://huggingface.co/

Searching for models

Hugging Face Model Hub section

¹ https://huggingface.co/models

Searching for models

Hugging Face Model Hub highlighting the filter

¹ https://huggingface.co/models

Searching for models

Hugging Face Model Hub highlighting the language filters

¹ https://huggingface.co/models

Searching for models

Hugging Face Model Hub highlighting the libraries filter section

¹ https://huggingface.co/models

Model cards

Model card for the Open AI Whisper model

¹ https://huggingface.co/openai/whisper-large-v3

Running a Basic Pipeline

from transformers import pipeline


my_pipeline = pipeline(
    "text-classification", 
    model="distilbert-base-uncased-finetuned-sst-2-english"
)


print(my_pipeline("DataCamp is awesome!"))

[{'label': 'POSITIVE', 'score': 0.99}]

Adjusting Pipeline Parameters

from transformers import pipeline

my_pipeline = pipeline("text-generation", model="gpt2")


results = my_pipeline("What if AI", max_length=10, num_return_sequences=2)


for result in results:
    print(result['generated_text'])

What if AI had never existed?  
What if AI could be really smarter than us?

When and why to save models

Save models locally for:
- Offline access
- Custom modifications
- Large-scale deployments

💾 Check model size on the Hub before saving!

¹ https://huggingface.co/distilbert/distilbert-base-uncased-finetuned-sst-2-english/tree/main

Saving a model locally

my_pipeline = pipeline("text-classification", 
                        model="distilbert-base-uncased-finetuned-sst-2-english")


my_pipeline.save_pretrained("/path/to/saved_model_directory")


reloaded_pipeline = pipeline("text-classification", 
                              model="/path/to/saved_model_directory")

print(reloaded_pipeline("Hugging Face is great!"))

[{'label': 'POSITIVE', 'score': 0.9998798370361328}]

Let's practice!

Working with Hugging Face