Embedder Base Class

geoembed.models.base

Abstract base class for all embedding models.

Embedder(device=None)

Bases: ABC

Abstract base class for all embedding models.

Ensures a consistent interface for the pipeline regardless of the underlying neural network architecture (e.g., ResNet, DOFA, Clay).

Subclasses must implement the embed method with model-specific inference logic.

Initialise the embedder, automatically selecting the best available hardware (CUDA, MPS or CPU) if no device is specified.

Parameters:

Name	Type	Description	Default
device	str | None	Explicit device string (e.g., "cuda:0"). If None, auto-detects.	None

Source code in src/geoembed/models/base.py

def __init__(self, device: str | None = None):
    """
    Initialise the embedder, automatically selecting the best
    available hardware (CUDA, MPS or CPU) if no device is specified.

    Args:
        device: Explicit device string (e.g., "cuda:0"). If None, auto-detects.
    """
    self.device = device or get_default_device()

embedding_dim abstractmethod property

Dimensionality of the output embedding vector.

embed(batch) abstractmethod

Run the model to generate embeddings.

Parameters:

Name	Type	Description	Default
batch	Tensor	A tensor of image chips. Shape (B, C, H, W). B = Batch Size C = Channels (typically 3 for RGB) H, W = Height, Width (typically 224)	required

Returns:

Type	Description
Tensor	A tensor of embeddings. Shape (B, D).
Tensor	D = Embedding Dimension (e.g., 768 for ViT-Base).

Source code in src/geoembed/models/base.py

@abstractmethod
def embed(self, batch: torch.Tensor) -> torch.Tensor:
    """
    Run the model to generate embeddings.

    Args:
        batch: A tensor of image chips. Shape (B, C, H, W).
               B = Batch Size
               C = Channels (typically 3 for RGB)
               H, W = Height, Width (typically 224)

    Returns:
        A tensor of embeddings. Shape (B, D).
        D = Embedding Dimension (e.g., 768 for ViT-Base).
    """
    ...