Applications
The library is sleek. In 8 lines of code you can train a state of the art vision classifier on the Oxford IIT pets dataset. The running thread through all the applications is the familiarity between very different datasets. It all includes importing the library, pulling in the data, creating a learner, and training the model. These steps are unsurprisingly similar between applications because deep learning is applicable to in variety of cases.
Abstractions Another key area to the Fastai library is abstractions it provides. All models will share attributes, thus there is the Learner class for that. All models must train, so there are a variety of function calls which can be used for that, ie model.fit_one_cycle(4)
Best Practices A nice bonus to the library is the current best practices built into the library defaults. Instead of expecting a user to have full knowledge of what the current best practices are, the model will default to generically proven values which yield generally good results. This allows better models to be discovered faster, with less tuning on the code side.
The applications provided by the library are as follows:
- Vision
- Text
- Tabular
- Collaborative Filtering
Each follows a similar code structure and allow the engineer to use their own data to get excellent results quickly on a number of application areas, as well as encouraging engineers to tinker in sundry applications.