MechaCritter/pyvisim-demo-models

pyvisim demo stores

Image stores used by the pyvisim demo Space. Each file is an InMemoryImageEmbeddingStore saved with pyvisim 1.0.0. It holds an embedder and the 7169 train and validation images of Oxford Flowers 102, indexed with brute-force search or with HNSW. The file names follow the pattern <embedder>_<index>.safetensors. The Space README lists the parameters of every embedder and index.

Load a store with:

from pyvisim.retrieval.image_store import InMemoryImageEmbeddingStore

store = InMemoryImageEmbeddingStore.load_from_disk("clip_hnsw.safetensors")

Keep in mind that:

• The stores save bare image file names such as image_00042.jpg. The images are in the demo's Oxford Flowers dataset repository.
• The CLIP stores were saved with device="cuda". pyvisim loads them onto the CPU on a machine without a CUDA GPU.

Trained networks

The networks/ folder holds the trained weights of the networks in the Contrastive Siamese and Triplet stores. The Space README describes how they were trained.

File	Class
networks/contrastive_siamese.safetensors	ContrastiveSiameseNetwork
networks/triplet_batch_hard.safetensors	TripletNeuralNetwork

Load a network with:

from pyvisim.neural_networks import TripletNeuralNetwork

network = TripletNeuralNetwork.load_from_disk("triplet_batch_hard.safetensors")