Embeddings and layouts

An embedding space holds one vector per sample for one model. A layout projects a space into 2D or 3D coordinates for the scatter panel. A dataset can hold several of each, so you can compare models and geometries side by side.

Compute embeddings

space = dataset.compute_embeddings(model="openai/clip-vit-base-patch32")

The provider is chosen from the model name:

• Hugging Face model IDs such as CLIP run through the built-in embed-anything provider.
• Names from hyper-models, such as hycoclip-vit-s or hyper3-clip-v1, produce hyperbolic embeddings.

Pass provider= to choose explicitly, for example open-clip, siglip, timm-image, sentence-transformers, or a hosted API such as openai, cohere, voyageai, jina or gemini-text. batch_size= fits the work to your hardware. Samples that already have embeddings for this model are skipped. hyperview.embeddings.list_embedding_providers() lists what is installed, and hv.register_provider adds your own.

Compute a layout

dataset.compute_visualization(space_key=space, layout="euclidean")

method="umap" (default) or "pca"; UMAP takes n_neighbors, min_dist and metric. Pass force=True to recompute an existing layout.

Why more than one geometry

Many embedding failures live in the shape of the space: collapsed classes, weak separation, a hierarchy squeezed into a flat plane, a long tail pushed to the edge. A single fixed projection hides some of these. Hyperbolic space has room for tree-like structure, so general concepts sit near the centre of the disk and specific ones near the boundary; a spherical view matches models trained with cosine similarity.

From the command line

With a workspace being served:

hyperview embeddings compute --workspace demo --dataset cifar100 \
  --model-id hycoclip-vit-s --layout poincare

hyperview layouts compute --workspace demo --dataset cifar100 \
  --space-key <space-key> --layout spherical

Both run as jobs in the server; hyperview jobs list shows their progress.