ffnn

`RL4CRN.utils.ffnn`

Feed-forward neural network utilities.

This module defines a lightweight fully-connected neural network backbone used throughout RL4CRN (e.g., policy encoders, parameter heads, and other small MLP components). It provides a single FFNN class that builds an MLP with a configurable number of hidden layers and activations, intended as a simple, general-purpose function approximator.

`FFNN`

Bases: Module

Simple feed-forward neural network (MLP) backbone.

The network is a stack of fully-connected layers mapping an input vector to an output vector. It is primarily used as a reusable building block for encoders and heads in policies/value functions.

Architecture

Linear(input_size -> hidden_size) + ReLU
num_layers blocks of: Linear(hidden_size -> hidden_size) + Tanh
Linear(hidden_size -> output_size)

Notes

This class does not apply any output activation; callers should apply any required squashing (e.g., tanh/softplus) externally.
Input tensors are expected to be of shape (N, input_size), where N is the batch dimension.

PARAMETER	DESCRIPTION
`input_size`	Dimensionality of the input features. TYPE: `int`
`output_size`	Dimensionality of the output features. TYPE: `int`
`hidden_size`	Width of the hidden layers. TYPE: `int`
`num_layers`	Number of additional hidden blocks after the first layer (each block is Linear + Tanh). TYPE: `int`

ATTRIBUTE	DESCRIPTION
`input_size`	Stored input dimensionality. TYPE: `int`
`output_size`	Stored output dimensionality. TYPE: `int`
`hidden_size`	Stored hidden width. TYPE: `int`
`num_layers`	Stored number of hidden blocks. TYPE: `int`
`model`	The assembled PyTorch module. TYPE: `Sequential`

Example

net = FFNN(input_size=16, output_size=4, hidden_size=64, num_layers=2)
x = torch.randn(32, 16)
y = net(x)  # (32, 4)