neuralset.extractors.image.HuggingFaceImage¶

pydantic model neuralset.extractors.image.HuggingFaceImage[source][source]¶

Compute image embeddings using transformer-based models obtained through HuggingFace API.

Parameters:

model_name (str, default="facebook/dinov2-base") – HuggingFace model identifier.

Fields:

batch_size (int)
event_types (Literal['Image', 'Video'])
frequency (float | Literal['native'])
hf_config (neuralset.extractors.image.HuggingFaceImageConfig)
imsize (int | None)
infra (exca.map.MapInfra)
model_name (str)

field event_types: Literal['Image', 'Video'] = 'Image'[source]¶

requirements: ClassVar[tuple[str, ...]] = ('transformers>=4.29.2', 'huggingface_hub>=0.27.0', 'transformers>=4.29.2', 'huggingface_hub>=0.27.0', 'torchvision>=0.15.2', 'transformers>=4.29.2', 'pillow>=9.2.0', 'pillow>=9.2.0')[source]¶

field model_name: str = 'facebook/dinov2-base'[source]¶

field hf_config: HuggingFaceImageConfig = HuggingFaceImageConfig(** { 'model_cls_name': 'AutoModel', 'model_kwargs': None, 'processor_cls_name': 'AutoProcessor', 'processor_kwargs': {'do_rescale': False}} )[source]¶

field infra: MapInfra = MapInfra(timeout_min=25, cpus_per_task=8, gpus_per_node=1, version='v6', min_samples_per_job=4096)[source]¶

field batch_size: int = 32[source]¶

field imsize: int | None = None[source]¶

field frequency: float | Literal['native'] = 0.0[source]¶

get_static(event: Image) → Tensor[source][source]¶

Return a single feature vector for the given event.

Override this method in subclasses to produce a static (non-temporal) embedding for one event. The returned tensor should have no time dimension — temporal wrapping is handled by BaseStatic automatically.

Parameters:: event (Event) – The event to extract a feature from.
Returns:: A tensor of shape (*feature_shape,) (no time axis).
Return type:: torch.Tensor

← Back to API reference