Module nn 

Neural network modules for deep learning.

This module provides PyTorch-compatible neural network building blocks following the API design described in Paszke et al. (2019).

Architecture

The nn module is organized around the Module trait, which defines the interface for all neural network layers:

• Layers: Linear, Conv1d, Conv2d, Flatten
• Pooling: MaxPool1d, MaxPool2d, AvgPool2d, GlobalAvgPool2d
• Activations: ReLU, Sigmoid, Tanh, GELU
• Normalization: BatchNorm1d, LayerNorm, GroupNorm, InstanceNorm, RMSNorm
• Regularization: Dropout, Dropout2d, AlphaDropout
• Containers: Sequential, ModuleList, ModuleDict

Example

use aprender::nn::{Module, Linear, ReLU, Sequential};
use aprender::autograd::Tensor;

// Build a simple MLP
let model = Sequential::new()
    .add(Linear::new(784, 256))
    .add(ReLU::new())
    .add(Linear::new(256, 10));

// Forward pass
let x = Tensor::randn(&[32, 784]);  // batch of 32
let output = model.forward(&x);     // [32, 10]

// Get all parameters for optimizer
let params = model.parameters();

References

• Paszke, A., et al. (2019). PyTorch: An imperative style, high-performance deep learning library. NeurIPS.
• Glorot, X., & Bengio, Y. (2010). Understanding the difficulty of training deep feedforward neural networks. AISTATS.
• He, K., et al. (2015). Delving deep into rectifiers. ICCV.

Re-exports

pub use functional as F;

pub use gnn::AdjacencyMatrix;

pub use gnn::GATConv;

pub use gnn::GCNConv;

pub use gnn::MessagePassing;

pub use gnn::SAGEAggregation;

pub use gnn::SAGEConv;

pub use loss::BCEWithLogitsLoss;

pub use loss::CrossEntropyLoss;

pub use loss::L1Loss;

pub use loss::MSELoss;

pub use loss::NLLLoss;

pub use loss::Reduction;

pub use loss::SmoothL1Loss;

pub use optim::Adam;

pub use optim::AdamW;

pub use optim::Optimizer;

pub use optim::RMSprop;

pub use optim::SGD;

pub use scheduler::CosineAnnealingLR;

pub use scheduler::ExponentialLR;

pub use scheduler::LRScheduler;

pub use scheduler::LinearWarmup;

pub use scheduler::PlateauMode;

pub use scheduler::ReduceLROnPlateau;

pub use scheduler::StepLR;

pub use scheduler::WarmupCosineScheduler;

Modules

functional

Functional interface for neural network operations.

generation

Sequence generation and decoding algorithms.

gnn

Graph Neural Network layers for learning on graph-structured data.

loss

Differentiable loss functions for neural network training.

optim

Gradient-based optimizers for neural network training.

quantization

Quantization-Aware Training (QAT) module.

scheduler

Learning rate schedulers for training neural networks.

self_supervised

Self-Supervised Learning Pretext Tasks.

serialize

Neural network model serialization.

vae

Variational Autoencoder (VAE) module.

Structs

ALiBi

ALiBi (Attention with Linear Biases) (Press et al., 2022).

AlphaDropout

Alpha Dropout for SELU activations.

AvgPool2d

Average Pooling 2D.

BatchNorm1d

Batch Normalization for 1D inputs (Ioffe & Szegedy, 2015).

Bidirectional

Bidirectional RNN wrapper.

Conv1d

1D Convolution layer.

Conv2d

2D Convolution layer.

DropBlock

DropBlock regularization (Ghiasi et al., 2018).

DropConnect

DropConnect regularization (Wan et al., 2013).

Dropout

Dropout regularization layer.

Dropout2d

2D Dropout (Spatial Dropout).

Flatten

Flatten layer.

GELU

Gaussian Error Linear Unit (GELU) activation.

GRU

Gated Recurrent Unit (GRU) layer.

GlobalAvgPool2d

Global Average Pooling 2D.

GroupNorm

Group Normalization (Wu & He, 2018).

GroupedQueryAttention

Grouped Query Attention (GQA).

InstanceNorm

Instance Normalization.

LSTM

Long Short-Term Memory (LSTM) layer.

LayerNorm

Layer Normalization (Ba et al., 2016).

LeakyReLU

Leaky ReLU activation: LeakyReLU(x) = max(negative_slope * x, x)

Linear

Fully connected layer: y = xW^T + b

LinearAttention

Linear Attention with kernel feature maps.

MaxPool1d

Max Pooling 1D.

MaxPool2d

Max Pooling 2D.

ModuleDict

Dictionary of named modules with string-based access.

ModuleList

List of modules with index-based access.

MultiHeadAttention

Multi-Head Attention (Vaswani et al., 2017).

PositionalEncoding

Sinusoidal Positional Encoding (Vaswani et al., 2017).

RMSNorm

Root Mean Square Layer Normalization (Zhang & Sennrich, 2019).

ReLU

Rectified Linear Unit activation: ReLU(x) = max(0, x)

RotaryPositionEmbedding

Rotary Position Embedding (RoPE) (Su et al., 2021).

Sequential

Sequential container for chaining modules.

Sigmoid

Sigmoid activation: σ(x) = 1 / (1 + exp(-x))

Softmax

Softmax activation: softmax(x)_i = exp(x_i) / Σ_j exp(x_j)

Tanh

Tanh activation: tanh(x) = (exp(x) - exp(-x)) / (exp(x) + exp(-x))

TransformerDecoderLayer

Transformer Decoder Layer.

TransformerEncoderLayer

Transformer Encoder Layer.

Traits

Module

Base trait for all neural network modules.

Functions

generate_causal_mask

Generate causal (triangular) attention mask.

kaiming_normal

Kaiming normal initialization (He et al., 2015).

kaiming_uniform

Kaiming uniform initialization (He et al., 2015).

xavier_normal

Xavier normal initialization (Glorot & Bengio, 2010).

xavier_uniform

Xavier uniform initialization (Glorot & Bengio, 2010).