fovi.arch.architectures

fovi.arch.architectures.alexnet2023(cfg, device='cuda')

Build a standard (non-foveated) AlexNet-2023 model from configuration.

Parameters:

• cfg – Configuration object containing model and training parameters.

• device (str, optional) – Device to build the model on. Defaults to ‘cuda’.

Returns:

The configured AlexNet-2023 model.

Return type:

nn.Module

fovi.arch.architectures.fovi_alexnet2023(cfg, device='cuda')

Build a foveated KNN-based AlexNet-2023 model from configuration.

This model uses KNN-based convolutions instead of standard convolutions, enabling foveated processing with non-uniform spatial sampling.

Parameters:

• cfg – Configuration object containing model, saccades, and training parameters.

• device (str, optional) – Device to build the model on. Defaults to ‘cuda’.

Returns:

The configured foveated AlexNet-2023 model wrapped with projector.

Return type:

nn.Module

fovi.arch.architectures.resnet18(*args, **kwargs)

fovi.arch.architectures.resnet34(*args, **kwargs)

fovi.arch.architectures.resnet50(*args, **kwargs)

fovi.arch.architectures.fovi_resnet(cfg, in_conv_stride=2, in_pool_stride=2, depthwise_sep_conv=False, layers=[2, 2, 2, 2], device='cuda')

Build a foveated KNN-based ResNet model from configuration.

This model uses KNN-based convolutions instead of standard convolutions, enabling foveated processing with non-uniform spatial sampling.

Parameters:

• cfg – Configuration object containing model, saccades, and training parameters.

• in_conv_stride (int, optional) – Stride for the initial convolution. Defaults to 2.

• in_pool_stride (int, optional) – Stride for the initial pooling. Defaults to 2.

• depthwise_sep_conv (bool, optional) – Whether to use depthwise separable convolutions. Defaults to False.

• layers (list, optional) – Number of blocks in each layer. Defaults to [2, 2, 2, 2].

• device (str, optional) – Device to build the model on. Defaults to ‘cuda’.

Returns:

The configured foveated ResNet model wrapped with projector.

Return type:

nn.Module

fovi.arch.architectures.fovi_resnet9(*args, **kwargs)

fovi.arch.architectures.fovi_resnet9_lowres(*args, **kwargs)

fovi.arch.architectures.fovi_resnet18(*args, **kwargs)

fovi.arch.architectures.fovi_resnet18_lowres(*args, **kwargs)

fovi.arch.architectures.fovi_resnet9_dwsep(*args, **kwargs)

fovi.arch.architectures.fovi_resnet9_dwsep_lowres(*args, **kwargs)

fovi.arch.architectures.fovi_resnet18_dwsep(*args, **kwargs)

fovi.arch.architectures.fovi_resnet18_dwsep_lowres(*args, **kwargs)

fovi.arch.architectures.fovi_vit(cfg, embed_dim, num_heads, device='cuda')

Build a foveated KNN-based Vision Transformer from configuration.

This model uses KNN-based tokenization instead of standard patch embedding, creating tokens based on spatial relationships in the foveated coordinate system.

Parameters:

• cfg – Configuration object containing model, saccades, and training parameters.

• embed_dim (int) – Embedding dimension for transformer.

• num_heads (int) – Number of attention heads.

• device (str, optional) – Device to build the model on. Defaults to ‘cuda’.

Returns:

The configured foveated ViT model wrapped with projector.

Return type:

nn.Module

fovi.arch.architectures.fovi_vit_base(*args, **kwargs)

Foveated ViT-Base configuration (l layers, 768 dim, 12 heads)

fovi.arch.architectures.fovi_vit_small(*args, **kwargs)

Foveated ViT-Small configuration (l layers, 384 dim, 6 heads)

fovi.arch.architectures.fovi_vit_tiny(*args, **kwargs)

Foveated ViT-Tiny configuration (l layers, 192 dim, 3 heads)

fovi.arch.architectures.vit(cfg, embed_dim, num_heads, device='cuda')

Build a standard Vision Transformer from configuration.

This model uses the classic ViT architecture with standard patch embedding.

Parameters:

• cfg – Configuration object containing model and training parameters.

• embed_dim (int) – Embedding dimension for transformer.

• num_heads (int) – Number of attention heads.

• device (str, optional) – Device to build the model on. Defaults to ‘cuda’.

Returns:

The configured ViT model wrapped with projector.

Return type:

nn.Module

fovi.arch.architectures.vit_base(*args, **kwargs)

ViT-Base configuration (l layers, 768 dim, 12 heads)

fovi.arch.architectures.vit_small(*args, **kwargs)

ViT-Small configuration (l layers, 384 dim, 6 heads)

fovi.arch.architectures.vit_tiny(*args, **kwargs)

ViT-Tiny configuration (l layers, 192 dim, 3 heads)

fovi.arch.architectures.vit_custom(cfg, device='cuda')

Custom non-foveated ViT configuration

fovi.arch.architectures.fovi_vit_custom(cfg, device='cuda')

Custom Foveated ViT configuration.

fovi.arch.architectures.fovi_dinov3(cfg, device='cuda')

pre-trained dinov3 – architectural differences in the dinov3 model are specified via cfg.pretrained_model.path

fovi.arch.architectures.arch_wrapper(backbone, cfg, device='cuda')

Build a SaccadeNet-ready architecture from an arbitrary backbone network.

Wraps a backbone network with MLP projection heads for downstream tasks and optionally for self-supervised learning.

Parameters:

• backbone (nn.Module) – The backbone network to wrap.

• cfg – Configuration object containing model and training parameters.

• device (str, optional) – Device to place the model on. Defaults to ‘cuda’.

Returns:

The wrapped model ready for SaccadeNet training.

Return type:

BackboneProjectorWrapper

class fovi.arch.architectures.ArchitectureRegistry

Bases: object

Registry for architecture builder functions.

This registry stores builder functions that can construct architecture instances from a configuration object. This allows external repositories to register custom architectures without modifying the core architecture code.

__init__()

register(name, builder_fn)

Register an architecture builder function.

Parameters:

• name (str) – Architecture name to register

• builder_fn (callable) – Function that takes cfg and device and returns an architecture instance

get(name)

Get an architecture builder by name.

Parameters:

name (str) – Architecture name to retrieve

Returns:

Builder function for the architecture

Return type:

callable

Raises:

ValueError – If architecture name is not found in registry

has(name)

Check if an architecture is registered.

Parameters:

name (str) – Architecture name to check

Returns:

True if architecture is registered, False otherwise

Return type:

bool

list_architectures()

List all registered architecture names.

Returns:

Sorted list of registered architecture names

Return type:

list