fovi.arch.knnresnet

class fovi.arch.knnresnet.KNNResNetBasicBlock(in_channels, out_channels, k, in_res, stride, fov, cmf_a, style='isotropic', conv_layer=KNNConvLayer, cart_res=None, norm_type='batch', arch_flag='', sample_cortex=True, device='cuda', auto_match_cart_resources=0, isotropic_plotting_type='v1like', fov_type='circular', ref_frame_mult=1)[source]

Bases: Module

Basic block for KNN-based ResNet architecture.

This block implements a residual connection with two KNN convolution layers, following the standard ResNet basic block design but using KNN convolutions instead of standard convolutions.

Parameters:

  • in_channels (int) – Number of input channels.

  • out_channels (int) – Number of output channels.

  • k (int) – Number of nearest neighbors for KNN convolution.

  • in_res (int) – Input resolution.

  • stride (int) – Stride for the first convolution layer.

  • fov (float) – Field of view parameter.

  • cmf_a (float) – Cortical magnification factor parameter.

  • style (str, optional) – Sampling style. Defaults to ‘isotropic’.

  • conv_layer (class, optional) – Convolution layer class to use. Defaults to KNNConvLayer.

  • cart_res (int, optional) – Cartesian resolution. Defaults to None.

  • norm_type (str, optional) – Normalization type. Defaults to ‘batch’.

  • arch_flag (str, optional) – Architecture flag. Defaults to ‘’.

  • sample_cortex (bool, optional) – Whether to sample cortex. Defaults to True.

  • device (str, optional) – Device to use. Defaults to ‘cuda’.

  • auto_match_cart_resources (int, optional) – Auto-match cartesian resources. Defaults to 0.

__init__(in_channels, out_channels, k, in_res, stride, fov, cmf_a, style='isotropic', conv_layer=KNNConvLayer, cart_res=None, norm_type='batch', arch_flag='', sample_cortex=True, device='cuda', auto_match_cart_resources=0, isotropic_plotting_type='v1like', fov_type='circular', ref_frame_mult=1)[source]

Initialize internal Module state, shared by both nn.Module and ScriptModule.

expansion = 1
forward(x)[source]

Forward pass through the basic block.

Parameters:

x (torch.Tensor) – Input tensor.

Returns:

Output tensor after applying residual connection.

Return type:

torch.Tensor

class fovi.arch.knnresnet.KNNResNetBottleneck(in_channels, out_channels, k, in_res, stride, fov, cmf_a, style='isotropic', conv_layer=KNNConvLayer, cart_res=None, norm_type='batch', arch_flag='', sample_cortex=True, device='cuda', auto_match_cart_resources=0, isotropic_plotting_type='v1like', fov_type='circular', ref_frame_mult=1)[source]

Bases: Module

Bottleneck block for KNN-based ResNet-50-style architectures.

This follows torchvision’s ResNet V1.5 layout: a 1x1 channel reduction, a 3x3 spatial convolution carrying the stride, and a 1x1 expansion. Only the spatial convolution uses a multi-point KNN neighborhood.

expansion = 4
__init__(in_channels, out_channels, k, in_res, stride, fov, cmf_a, style='isotropic', conv_layer=KNNConvLayer, cart_res=None, norm_type='batch', arch_flag='', sample_cortex=True, device='cuda', auto_match_cart_resources=0, isotropic_plotting_type='v1like', fov_type='circular', ref_frame_mult=1)[source]

Initialize internal Module state, shared by both nn.Module and ScriptModule.

forward(x)[source]

Define the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

class fovi.arch.knnresnet.KNNResNet(block=KNNResNetBasicBlock, layers=[2, 2, 2, 2], in_conv_stride=2, in_pool_stride=2, stem_kernel_size=7, fov=16, cmf_a=0.5, in_res=64, out_res=1, style='isotropic', conv_layer=KNNConvLayer, pool_layer=KNNPoolingLayer, norm_type='batch', arch_flag='', sample_cortex=True, device='cuda', auto_match_cart_resources=0, num_classes=None, isotropic_plotting_type='v1like', fov_type='circular', ref_frame_mult=1)[source]

Bases: Module

KNN-based ResNet architecture.

A ResNet implementation that uses KNN convolutions instead of standard convolutions, designed for foveated vision processing.

Parameters:

  • block (class, optional) – Block class to use for layers. Defaults to KNNResNetBasicBlock.

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

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

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

  • fov (float, optional) – Field of view parameter. Defaults to 16.

  • cmf_a (float, optional) – Cortical magnification factor parameter. Defaults to 0.5.

  • in_res (int, optional) – Input resolution. Defaults to 64.

  • out_res (int, optional) – Output resolution. Defaults to 1.

  • style (str, optional) – Sampling style. Defaults to ‘isotropic’.

  • conv_layer (class, optional) – Convolution layer class to use. Defaults to KNNConvLayer.

  • pool_layer (class, optional) – Pooling layer class to use. Defaults to KNNPoolingLayer.

  • norm_type (str, optional) – Normalization type. Defaults to ‘batch’.

  • arch_flag (str, optional) – Architecture flag. Defaults to ‘’.

  • sample_cortex (bool, optional) – Whether to sample cortex. Defaults to True.

  • device (str, optional) – Device to use. Defaults to ‘cuda’.

  • auto_match_cart_resources (int, optional) – Auto-match cartesian resources. Defaults to 0.

  • num_classes (int, optional) – Number of output classes for classification. If None, no classification head is added. Defaults to None.

__init__(block=KNNResNetBasicBlock, layers=[2, 2, 2, 2], in_conv_stride=2, in_pool_stride=2, stem_kernel_size=7, fov=16, cmf_a=0.5, in_res=64, out_res=1, style='isotropic', conv_layer=KNNConvLayer, pool_layer=KNNPoolingLayer, norm_type='batch', arch_flag='', sample_cortex=True, device='cuda', auto_match_cart_resources=0, num_classes=None, isotropic_plotting_type='v1like', fov_type='circular', ref_frame_mult=1)[source]

Initialize internal Module state, shared by both nn.Module and ScriptModule.

_make_layer(block, planes, blocks, stride=1)[source]

Create a layer with the specified number of blocks.

Parameters:

  • block (class) – Block class to use for the layer.

  • planes (int) – Number of output channels for the layer.

  • blocks (int) – Number of blocks in the layer.

  • stride (int, optional) – Stride for the first block. Defaults to 1.

Returns:

Sequential container with the layer blocks.

Return type:

nn.Sequential

forward(x)[source]

Forward pass through the KNN ResNet.

Parameters:

x (torch.Tensor) – Input tensor of shape (batch_size, 3, height, width).

Returns:

Output tensor. If num_classes is specified, returns

classification logits. Otherwise, returns feature embeddings.

Return type:

torch.Tensor