This project was developed to support an app designed for labeling triplets of data to fine-tune a 3D similarity model. Learn more about this project here. Specifically, a reference item is displayed in the center, with two options presented alongside it. The user can swipe to select the option that appears most similar to the reference.
Architecture choices inspired by this FastAPI tutorial, Netflix Dispatch project.
A similarity service powered by pgvector has been developed.
The app will deal with two kind of triplets, which are stored in two different tables:
- The triplets located in the "triplets" table. They are the triplets coming from a single model that need to be labeled by the user of the app, to know which one of the proposition is closest to the anchor proposition, which is located at the center. They are the default triplets.
- The triplets located in the "validation triplets" table. Validation triplets are used to compare 2 iterations of our similarity model.
When we want to reference to both triplets and validation triplets, we use the reference 'all triplets'.
Triplets are loaded from the application. An example of datapack can be found in data_example/test_data.zip.