Struct SparseConv2d 

pub struct SparseConv2d { /* private fields */ }

Sparse 2D Convolution layer

Implements 2D convolution operations optimized for sparse inputs and/or sparse kernels. This can significantly reduce computation when either the input or kernel has many zeros.

Mathematical Formulation

For standard convolution: y[i,j] = Σ_m Σ_n w[m,n] * x[i+m, j+n] For sparse convolution: Only compute terms where w[m,n] ≠ 0

Use Cases

• Computer vision with sparse activations (post-ReLU)
• Medical imaging with sparse features
• Pruned neural networks
• Specialized domains with naturally sparse data

Implementations

impl SparseConv2d

pub fn new( in_channels: usize, out_channels: usize, kernel_size: (usize, usize), stride: Option<(usize, usize)>, padding: Option<(usize, usize)>, dilation: Option<(usize, usize)>, sparsity: f32, use_bias: bool, ) -> TorshResult<Self>

Create a new sparse 2D convolution layer

Arguments

• in_channels - Number of input channels
• out_channels - Number of output channels
• kernel_size - Convolution kernel size (height, width)
• stride - Stride for convolution (default: (1, 1))
• padding - Padding for convolution (default: (0, 0))
• dilation - Dilation for convolution (default: (1, 1))
• sparsity - Kernel sparsity level (0.0 = dense, 1.0 = fully sparse)
• use_bias - Whether to include learnable bias

Returns

• TorshResult<Self> - New sparse 2D convolution layer or error

Example

use torsh_sparse::nn::convolution::SparseConv2d;

// Create sparse conv: 3->64 channels, 3x3 kernel, 90% sparse
let conv = SparseConv2d::new(
    3, 64, (3, 3),
    Some((1, 1)), Some((1, 1)), None,
    0.9, true
).unwrap();

pub fn forward(&self, input: &Tensor) -> TorshResult<Tensor>

Forward pass for sparse convolution

Arguments

• input - Input tensor (batch_size, in_channels, height, width)

Returns

• TorshResult<Tensor> - Output tensor (batch_size, out_channels, out_height, out_width)

pub fn num_parameters(&self) -> usize

Get the number of parameters

pub fn kernel_sparsity(&self) -> f32

Get kernel sparsity

pub fn in_channels(&self) -> usize

Get input channels

pub fn out_channels(&self) -> usize

Get output channels

pub fn kernel_size(&self) -> (usize, usize)

Get kernel size

pub fn stride(&self) -> (usize, usize)

Get stride

pub fn padding(&self) -> (usize, usize)

Get padding

pub fn dilation(&self) -> (usize, usize)

Get dilation

Trait Implementations

impl Clone for SparseConv2d

fn clone(&self) -> SparseConv2d

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for SparseConv2d

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.