LatentViewer

LatentViewer is a visualisation tool to inspect image embeddings. It uses principal component analysis to display the embedding vectors as a point cloud. Individual points can be selected to display an image, as well as its nine closest neighbours.

Additionally, there is the possibility to train an SVM classifier through an active learning method. This is particularly useful when dealing with embeddings of an unlabeled data set.

Installation

Install the latent viewer with pip:

pip install latent-viewer

Usage

After installation, the latent-viewer can be invoked with lv. For using, it is important to specify both a file with the embeddings, as well as an HDF5 image archive.

lv -e embeddings.csv -a image_archive.hdf5