If you're tired of rewriting the same boilerplate code of your training pipelines in PyTorch, I've found a pretty neat solution that could make your life easier. Don't worry, it's not a heavy library that'll change your way of doing things.
It's rather a lightweight wrapper that encapsulates your training logic in a single class. It's built on top of PyTorch, it's quite recent but I've tested it and I think it does what it promises so far.
It's called Tez and we'll see it today in action on a fun multi-label text classification problem. Let's jump right in.
- Using the Datasets library to load and manipulate go_emotions data
- Defining the training pipeline with Tez
- Training a SqueezeBert lightweight model for a multi-label classification problem and reaching +0.9 AUC on validation and test data
- Deploying the model
- Crafting a small UI with React or Streamlit