Skip to main content

coeus_nn/
lib.rs

1//! Neural network layer module system built on [`coeus_autograd`].
2//!
3//! # Module trait
4//! [`Module<T, B>`](module::Module) is the core abstraction. Its `forward`
5//! method returns a typed [`ModuleError`] so backend, input-contract, and
6//! module-state failures cannot be reported as successful execution.
7//!
8//! # Layer families
9//! - **Linear** — [`Linear`], weight + optional bias, Xavier/Kaiming init via [`init`].
10//! - **Convolution** — [`Conv1d`], [`Conv2d`], [`Conv3d`] with stride/padding/dilation.
11//! - **Normalization** — [`LayerNorm`], [`RMSNorm`], [`BatchNorm1d/2d/3d`](BatchNorm2d), [`GroupNorm`], [`InstanceNorm1d`], [`InstanceNorm2d`], [`InstanceNorm3d`].
12//! - **Pooling** — [`MaxPool2d`], [`AvgPool2d`], [`MaxPool3d`], [`AvgPool3d`].
13//! - **Attention** — [`MultiHeadAttention`], [`ScaledDotProductAttention`] with [`CausalMask`] / [`NullMask`].
14//! - **Transformer** — [`TransformerEncoder`], [`TransformerDecoder`], [`FeedForward`] blocks.
15//! - **Positional** — [`SinusoidalEncoding`], [`RotaryEmbedding`].
16//! - **Composites** — [`Sequential`], [`StaticSeq`], [`Dropout`], [`Embedding`], [`Softmax`].
17
18// ── Coeus NN ──
19// Neural network building blocks.
20#![deny(missing_docs)]
21
22/// Activation functions (ReLU, GeLU, SiLU, etc.).
23pub mod activation;
24/// Attention mechanisms (MHA, SDP, masks).
25pub mod attention;
26/// Bilinear interaction layer.
27pub mod bilinear;
28/// Convolution layers (1D/2D/3D and transposed variants).
29pub mod conv;
30/// Dropout regularization layer.
31pub mod dropout;
32/// Embedding lookup layer.
33pub mod embedding;
34/// EmbeddingBag aggregation layer.
35pub mod embeddingbag;
36/// Weight initialization utilities (Xavier, Kaiming).
37pub mod init;
38/// Spatial interpolation operations.
39pub mod interpolate;
40/// Fully-connected linear layer.
41pub mod linear;
42/// Loss functions (cross-entropy, MSE, etc.).
43pub mod loss;
44/// Core `Module` trait for neural network layers.
45pub mod module;
46/// Normalization layers (LayerNorm, BatchNorm, GroupNorm, etc.).
47pub mod normalization;
48/// Learnable parameter wrapper.
49/// Pooling layers (max, average, global).
50pub mod pool;
51/// Positional encoding layers (sinusoidal, RoPE).
52pub mod positional;
53/// Regularization layers (AlphaDropout, FeatureAlphaDropout, GaussianNoise, LocalResponseNorm).
54pub mod regularization;
55/// Recurrent layers (LSTM, GRU).
56pub mod rnn;
57/// Sequential and static-sequence module containers.
58pub mod sequential;
59/// Softmax layer and functional softmax.
60pub mod softmax;
61/// SwiGLU gated feed-forward unit.
62pub mod swiglu;
63/// Transformer encoder, decoder, and sub-layers.
64pub mod transformer;
65
66pub use activation::{
67    celu, elu, gelu, gelu_tanh, glu, hardshrink, hardsigmoid, hardswish, hardtanh, leaky_relu,
68    log_sigmoid, mish, prelu, relu, sigmoid, silu, softplus, softshrink, softsign, tanh,
69    tanhshrink, threshold, Celu, CeluOp, GeLU, GeLUTanh, Hardshrink, HardshrinkOp, Hardsigmoid,
70    HardsigmoidOp, Hardswish, HardswishOp, Hardtanh, HardtanhOp, LeakyReLU, LogSigmoid, Mish,
71    PReLU, ReLU, SiLU, Sigmoid, Softplus, Softshrink, SoftshrinkOp, Softsign, SoftsignOp, Tanh,
72    Tanhshrink, Threshold, ThresholdNode, ELU, GLU,
73};
74pub use attention::{
75    multi_head_attention_cross, AttentionMask, CausalMask, MhaProjectionParams, MultiHeadAttention,
76    NullMask, ScaledDotProductAttention,
77};
78pub use bilinear::{bilinear, Bilinear};
79pub use conv::{
80    Conv, Conv1d, Conv2d, Conv3d, ConvDim, ConvTranspose, ConvTranspose1d, ConvTranspose2d,
81    ConvTranspose3d, DepthwiseConv3d, Dim1D, Dim2D, Dim3D, Fold1d, Fold2d, Unfold1d, Unfold2d,
82};
83pub use dropout::Dropout;
84pub use embedding::Embedding;
85pub use embeddingbag::{EmbeddingBag, EmbeddingBagMode};
86pub use init::{kaiming_uniform, xavier_uniform};
87pub use interpolate::{interpolate_1d, interpolate_2d, ResizeMode};
88pub use linear::Linear;
89pub use loss::{
90    bce_with_logits, binary_cross_entropy, cosine_embedding_loss, cosine_similarity,
91    cross_entropy_loss, ctc_loss, gaussian_nll_loss, hinge_embedding_loss, huber_loss,
92    kl_divergence, l1_loss, margin_ranking_loss, mse_loss, multi_label_soft_margin_loss,
93    multi_margin, nanmean, nansum, nll_loss, pairwise_distance, poisson_nll, smooth_l1_loss,
94    soft_margin, triplet_margin_loss, triplet_margin_with_distance_loss,
95};
96pub use module::{Module, ModuleError, ModuleResult, ParameterLoadError};
97pub use normalization::{
98    group_norm, layer_norm, rms_norm, BatchNorm1d, BatchNorm2d, BatchNorm3d, GroupNorm,
99    InstanceNorm1d, InstanceNorm2d, InstanceNorm3d, LayerNorm, NormalizedShape, RMSNorm,
100};
101pub use pool::{
102    AdaptiveAvgPool1d, AdaptiveAvgPool2d, AdaptiveMaxPool1d, AdaptiveMaxPool2d, AvgPool1d,
103    AvgPool2d, AvgPool3d, GlobalAvgPool1d, GlobalAvgPool2d, GlobalAvgPool3d, GlobalMaxPool2d,
104    GlobalMaxPool3d, MaxPool1d, MaxPool2d, MaxPool3d,
105};
106pub use positional::{RotaryEmbedding, SinusoidalEncoding};
107pub use regularization::{AlphaDropout, FeatureAlphaDropout, GaussianNoise, LocalResponseNorm};
108pub use rnn::{Bidirectional, GRUCell, Gru, LSTMCell, Lstm, RNNCell, Rnn, RnnNonlinearity};
109pub use sequential::{ModuleExt, Sequential, StaticSeq};
110pub use softmax::{softmax, Softmax};
111pub use swiglu::SwiGlu;
112pub use transformer::{
113    feed_forward, transformer_decoder_layer, transformer_encoder_layer, FeedForward, Transformer,
114    TransformerDecoder, TransformerDecoderLayer, TransformerDecoderLayerParams, TransformerEncoder,
115    TransformerEncoderLayer, TransformerEncoderLayerParams,
116};