fastai: a deep learning library that balances high-level productivity with low-level research flexibility

What it solves

fastai is a deep learning library designed to make state-of-the-art results in standard deep learning domains accessible and easy to achieve. It bridges the gap between high-level components for practitioners who want rapid productivity and low-level components for researchers who need flexibility and hackability to build new approaches.

How it works

Built on top of PyTorch, fastai uses a layered architecture of decoupled abstractions. This allows users to choose the level of API they use—from high-level, concise code for standard tasks like image classification or text sentiment analysis, to low-level building blocks for custom research. Key technical features include a GPU-optimized computer vision library, a novel 2-way callback system for training modifications, a new data block API, and a custom type dispatch system for Python.

Who it’s for

It is designed for both deep learning practitioners who want to quickly build and deploy models, and researchers who need a highly configurable and configurable framework for experimenting with new deep learning techniques.

Highlights

• own a layered API that balances ease of use with deep hackability
• supports image classification, image segmentation, text sentiment, recommendation systems, and tabular models
• provides a GPU-optimized computer vision library
• features a flexible 2-way callback system to modify data, models, or optimizers during training
• integrates seamlessly with PyTorch and other PyTorch-based libraries