Function sparse_conv2d 

pub fn sparse_conv2d(
    input: &SparseTensor,
    weight: &Tensor,
    bias: Option<&Tensor>,
    stride: (usize, usize),
    padding: (usize, usize),
    dilation: (usize, usize),
) -> Result<SparseTensor>

Sparse 2D convolution

Performs 2D convolution on sparse input tensors with dense kernels. Optimized for sparse inputs by only processing non-zero elements.

Mathematical Formula

For input x and kernel w: y[b, o, h, w] = Σ(x[b, i, h + kh*dh - ph, w + kw*dw - pw] * w[o, i, kh, kw]) + bias[o] where b=batch, o=output channel, i=input channel, h,w=spatial positions, kh,kw=kernel positions

Arguments

• input - Sparse input tensor [batch_size, channels, height, width]
• weight - Dense weight tensor [out_channels, in_channels, kernel_height, kernel_width]
• bias - Optional bias tensor [out_channels]
• stride - Convolution stride (height, width)
• padding - Zero padding (height, width)
• dilation - Kernel dilation (height, width)

Returns

Sparse output tensor after convolution