demo-qdrant-fiftyone

FiftyOne is a tool designed to support creating high-quality visual datasets. It allows inspecting the data points and monitoring the performance of your models to allow reacting much faster if you spot a degraded quality of the predictions.

Qdrant is a vector database designed to perform an approximate nearest neighbour search (ANN) on dense neural embeddings. Those embeddings have some interesting capabilities, as they are able to capture the semantics of the data points. This is especially useful for unstructured data like images and videos, so you can encode not only pixel similarities but also some more complex relationships.

The neural search might be a great strategy for performing a classification, based on the similarity to the examples from the training set, in a way similar to k-NN, but with much better performance.

This repository contains a working example of integrating Qdrant with FiftyOne. The example uses the MNIST dataset and performs a classification of the test samples using Approximate Neural Neighbours in the embedding space, using cosine distance.

Running the example

The example requires some dependencies to be installed:

pip install -r requirements.txt

Qdrant has to be launched using Docker:

bash run_qdrant.sh

Finally, the script performing the classification and running FiftyOne in web browser can be launched via:

python main.py

As a result, the script should open the default web browser with the FiftyOne UI showing the misclassified examples.