CLIP ViT-B/32 multilingual text encoder, one ONNX graph
The text side of content search in Usva, a personal photo vault that runs on your own computer. The app downloads these files when you turn the feature on. Searches and photos stay on your computer.
This is an ONNX export of
sentence-transformers/clip-ViT-B-32-multilingual-v1,
which is Apache-2.0. The weights are unchanged. The only difference is packaging: the published ONNX file outputs
last_hidden_state, while this export includes the mean pooling and the 768→512 projection in one graph.
| File | What it is |
|---|---|
textual.onnx |
Inputs input_ids and attention_mask; output a 512-dimensional vector, not normalised |
vocab.txt |
The cased WordPiece vocabulary of DistilBERT multilingual |
The vectors match those of sentence-transformers (cosine 1.0, checked at export) and share a space with the CLIP ViT-B/32
image encoder, so text can be compared with images. Normalise the vector to length 1 before comparing.
SHA-256:
textual.onnx 5713755215ce0fc89fac58c0ca1c2895886b2341f29d9d4bc721ff5783a171ff
vocab.txt fe0fda7c425b48c516fc8f160d594c8022a0808447475c1a7c6d6479763f310c
Licence: Apache-2.0, as the original model. Exported by TaitoLabs Oy.