Embedder¶

siglip2_toolkit.embedder.Embedder ¶

Embedder(
    model,
    device="auto",
    dtype=None,
    trust_remote_code=False,
    model_id=None,
)

Generates text and image embeddings with a SigLIP 2 model.

model is a local model directory; model_id is the identifier recorded in logs and in the embedding-space manifest (defaults to the path). Load the model from the toolkit cache with siglip2_toolkit.models.resolve_model.

Embeddings are L2-normalised float32 numpy arrays. The instance is safe to share across threads (inference is serialised with an internal lock); it is not safe to share across processes.

device `property` ¶

device

Device string (e.g. "cuda"/"cpu"), or None if not loaded.

dimensions `property` ¶

dimensions

Embedding dimension, or None if the model has not been loaded.

load ¶

load()

Load the model and processor from the local directory (idempotent).

embedding_space ¶

embedding_space()

The embedding space manifest for the loaded model.

embed_texts ¶

embed_texts(texts, batch_size=32, max_length=64)

Embed a batch of texts. Returns a (n, dimensions) float32 array.

embed_text ¶

embed_text(text)

Embed a single text. Returns a (dimensions,) float32 array.

embed_images ¶

embed_images(
    images,
    batch_size=32,
    on_error="raise",
    return_indices=False,
)

Embed a batch of images.

images items may be paths, bytes, file-like objects or PIL images. With on_error="skip" unreadable items are logged and skipped; the kept positions are returned alongside the matrix when return_indices=True.

embed_image ¶

embed_image(image)

Embed a single image. Returns a (dimensions,) float32 array.

zero_shot_logits ¶

zero_shot_logits(
    images, labels, template="{}", batch_size=32
)

Raw zero-shot classification logits, shaped (n_images, n_labels).

Scores follow the SigLIP recipe: cosine similarity between normalised embeddings, scaled by the model's learned logit_scale and shifted by logit_bias. template wraps each label ({} is replaced by the label). When the model has no learned scale (non-SigLIP checkpoints), a CLIP-style temperature of 100 is used.

classify ¶

classify(
    images, labels, template="{}", top_k=None, batch_size=32
)

Zero-shot image classification.

Returns, for each image, (label, probability) pairs sorted by decreasing softmax probability over the candidate labels (mutually exclusive assumption). Use zero_shot_logits when you need the raw scores for custom decision rules.