Skip to main content

Linear

Linear (Fully-Connected Layer)

Applies a linear transformation to the incoming data: y = xW^T + b This is a shape-aware wrapper around torch.nn.Linear that validates tensor dimensions according to NeuroScript shape algebra.

Parameters:

  • in_dim: Size of each input sample (last dimension)
  • out_dim: Size of each output sample (last dimension)

Shape Contract:

  • Input: [*, in_dim] where * means any number of leading dimensions
  • Output: [*, out_dim] with same leading dimensions as input

Notes:

  • Preserves all leading dimensions (broadcasting-compatible)
  • Weight shape: [out_dim, in_dim]
  • Includes learnable bias by default

Signature​

neuron Linear(in_dim, out_dim)

Ports​

Inputs:

  • default: [*, in_dim]

Outputs:

  • default: [*, out_dim]

Implementation​

External { kwargs: [] }