This section of the Autodistill documentation covers the low-level Autodistill API. This section may be useful for advanced users who want to understand the Autodisitll API in more depth.
Autodistill API, in a Nutshell¶
The Autodistill API consists of three concepts:
- Base models, which are used to auto-label data;
- Target models, which are trained on the auto-labeled data; and
- Ontologies, which tell Autodistill what you want to identify and what labels should be called in your dataset.
There are three different kinds of base models:
- Detection models, which identify objects in images and return bounding boxes and/or segmentation masks;
- Classification models, which classify images and return a class label; and;
- Embedding models, which return semantic embeddings for images. These are used with the
The base model you use depends on the type of data you want to label. You can also combine models using the ComposedDetectionModel API, which allows you to refine labels from detection models.
There are two different kinds of target models:
- Detection models, and;
- Classification models.
Autodistill does not support training embedding models.