Skip to content

Dense — Fully Connected Layer

The Dense layer applies a trainable affine transformation y = x @ W + b to the input.


Overview

Dense is MAGNET's standard fully connected layer. It learns a weight matrix W and optional bias vector b, supporting both current and older compatibility initializer APIs.


Key Features

  • Standard affine transformation on last input axis
  • Configurable in_features / out_features
  • Optional bias (use_bias=True)
  • Compatibility initializers (kernel_initializer, bias_initializer)
  • Custom backward function support

Parameters

Parameter Type Default Description
in_features int None Input dimension (alias n_in)
out_features int None Output dimension (alias n_out)
weight_init / kernel_initializer callable XavierUniform Weight initializer
bias_init / bias_initializer callable Zeros() Bias initializer
use_bias bool True Include bias term
name str | None None Layer name
verbose bool False Verbose logging

Example Usage

import numpy as np
from machinegnostics.magnet import Dense

d = Dense(in_features=3, out_features=2)
print(d(np.array([[1.0, 2.0, 3.0]])).shape)  # (1, 2)

Author: Nirmal Parmar