[Feature]: Anomaly Detection / One Class Classification

[jeffpicard]

Problem statement

Flair doesn't have an Anomaly Detection / One Class Classification model. While this question describes alternate solutions, some use cases could benefit from this dedicated model type.

Solution

Add a new AnomalyDetection class to flair.models. The implementation would train an autoencoder on a single class of data. At inference time, the autoencoder's reconstruction error of the input would be calculated and compared to a threshold to determine whether the input is in-class.

Additional Context

I plan to pull request code for this soon.

[alanakbik]

Hello @jeffpicard that sounds great, let us know if you need any help!

[jeffpicard]

Thanks @alanakbik! I put up a first stab (linked above) if you're interested.