Skip to main content

Module

Trait Module 

Source
pub trait Module<T: Scalar, B: ComputeBackend + Default = MoiraiBackend> {
    // Required methods
    fn parameters(&self) -> Vec<Var<T, B>>;
    fn forward(
        &self,
        input: &Var<T, B>,
    ) -> Result<Var<T, B>, ModuleError<B::Error>>;

    // Provided methods
    fn named_parameters(&self) -> Vec<Parameter<T, B>> { ... }
    fn load_parameters(&mut self, _params: &[Var<T, B>]) { ... }
    fn load_named_parameters(
        &mut self,
        parameters: &[Parameter<T, B>],
    ) -> Result<(), ParameterLoadError> { ... }
    fn zero_grad(&self) -> Result<(), B::Error> { ... }
    fn train(&mut self, _mode: bool) { ... }
    fn eval(&mut self) { ... }
}
Expand description

Trait for neural network modules.

Each module owns Parameters, can contain sub-modules, and can support train/eval modes.

Required Methods§

Source

fn parameters(&self) -> Vec<Var<T, B>>

Collect all trainable parameters (including from sub-modules).

Source

fn forward(&self, input: &Var<T, B>) -> Result<Var<T, B>, ModuleError<B::Error>>

Evaluate this module.

§Errors

Returns a typed module or backend failure when the input violates the module contract, state cannot be borrowed, or backend execution fails.

Provided Methods§

Source

fn named_parameters(&self) -> Vec<Parameter<T, B>>

Collect trainable parameters with stable hierarchical names.

Leaf modules with the canonical zero-, one-, or two-parameter layout inherit [], [weight], or [weight, bias]. Modules with a wider layout or child modules must override this method so names express the owned field hierarchy instead of depending on flattened ordinals.

Source

fn load_parameters(&mut self, _params: &[Var<T, B>])

Write optimizer-updated parameter values back into this module’s own fields, consuming params in the same order parameters() enumerates them (i.e. params.len() == parameters().len()).

This exists because Var::clone() shares gradient state but not the value storage: coeus_optim::{SGD, Adam}::step mutates its own owned Vec<Var<T, B>> in place (via copy-on-write on first mutation), so a module that read parameters() to build an optimizer must call this after each step() to see the updated values, unless the module’s own fields are the optimizer’s Vars (as in a flat training loop that never separately clones them into a named struct).

Default is a no-op, which is correct for parameterless modules (activations, dropout, pooling, reshape/view layers, …). Any module whose parameters() returns a non-empty Vec must override this.

Source

fn load_named_parameters( &mut self, parameters: &[Parameter<T, B>], ) -> Result<(), ParameterLoadError>

Load optimizer-updated named parameters after validating the complete stable path inventory.

§Errors

Returns ParameterLoadError when count or path ordering differs from this module’s canonical named inventory.

Source

fn zero_grad(&self) -> Result<(), B::Error>

Zero all parameter gradients.

Source

fn train(&mut self, _mode: bool)

Set the training mode of the module and its sub-modules.

Source

fn eval(&mut self)

Set the module and its sub-modules to evaluation mode.

Dyn Compatibility§

This trait is dyn compatible.

In older versions of Rust, dyn compatibility was called "object safety".

Implementors§

Source§

impl<ScalarType: Scalar, B: ComputeBackend + Default, H: Module<ScalarType, B>, T: Module<ScalarType, B>> Module<ScalarType, B> for StaticSeq<H, T>

Source§

impl<T, B, M> Module<T, B> for Bidirectional<M>

Source§

impl<T, B: BackendOps<T> + Default, const DIM: usize> Module<T, B> for ConvTranspose<T, B, DIM>
where T: RealScalar + Float,

Source§

impl<T, B> Module<T, B> for DepthwiseConv3d<T, B>

Source§

impl<T: AttentionScalar + FloatElement + RealScalar, B: BackendOps<T> + AttentionOps<T> + Default, const H: usize, M: AttentionMask> Module<T, B> for MultiHeadAttention<T, B, H, M>

Self-attention Module impl (Q = K = V = input).

Source§

impl<T: AttentionScalar + FloatElement, B: BackendOps<T> + AttentionOps<T> + Default, M: AttentionMask> Module<T, B> for ScaledDotProductAttention<T, B, M>

Convenience Module impl — self-attention: Q = K = V = input.

Source§

impl<T: Float + FloatElement + Neg<Output = T> + RealScalar, B: BackendOps<T> + Default> Module<T, B> for LocalResponseNorm

Source§

impl<T: Float + FloatElement + RealScalar, B: BackendOps<T> + Default> Module<T, B> for AlphaDropout

Source§

impl<T: Float + FloatElement + RealScalar, B: BackendOps<T> + Default> Module<T, B> for FeatureAlphaDropout

Source§

impl<T: Float + FloatElement + RealScalar, B: BackendOps<T> + Default> Module<T, B> for GaussianNoise

Source§

impl<T: Float + FloatElement, B: BackendOps<T> + Default> Module<T, B> for Dropout

Source§

impl<T: Float + RealScalar, B: BackendOps<T> + Default> Module<T, B> for Bilinear<T, B>

Source§

impl<T: Float + RealScalar, B: BackendOps<T> + Default> Module<T, B> for FeedForward<T, B>

Source§

impl<T: Float + RealScalar, B: BackendOps<T> + Default> Module<T, B> for GLU

Source§

impl<T: Float + RealScalar, B: BackendOps<T> + Default> Module<T, B> for GRUCell<T, B>

Source§

impl<T: Float + RealScalar, B: BackendOps<T> + Default> Module<T, B> for Gru<T, B>

Source§

impl<T: Float + RealScalar, B: BackendOps<T> + Default> Module<T, B> for LSTMCell<T, B>

Source§

impl<T: Float + RealScalar, B: BackendOps<T> + Default> Module<T, B> for Lstm<T, B>

Source§

impl<T: Float + RealScalar, B: BackendOps<T> + Default> Module<T, B> for PReLU<T, B>

Source§

impl<T: Float + RealScalar, B: BackendOps<T> + Default> Module<T, B> for RNNCell<T, B>

Source§

impl<T: Float + RealScalar, B: BackendOps<T> + Default> Module<T, B> for Rnn<T, B>

Source§

impl<T: Float + RealScalar, B: BackendOps<T> + Default> Module<T, B> for Softmax

Source§

impl<T: Float + RealScalar, B: BackendOps<T> + Default> Module<T, B> for SwiGlu<T, B>

Source§

impl<T: Float + RealScalar, B: BackendOps<T> + Default> Module<T, B> for Tanhshrink

Implements the crate::module::Module interface for Tanhshrink.

Source§

impl<T: Float + RealScalar, B: ElementwiseOps<T> + ReductionOps<T> + Default> Module<T, B> for SinusoidalEncoding<T, B>

Source§

impl<T: Float + RealScalar, B> Module<T, B> for RotaryEmbedding<T, B>
where B: RotateHalfOps<T> + BackendOps<T> + Default,

Source§

impl<T: Float, B: BackendOps<T> + Default, D: ConvDim> Module<T, B> for Conv<T, B, D>

Source§

impl<T: Float, B: BackendOps<T> + Default> Module<T, B> for AvgPool1d<T, B>

Source§

impl<T: Float, B: BackendOps<T> + Default> Module<T, B> for AvgPool2d<T, B>

Source§

impl<T: Float, B: BackendOps<T> + Default> Module<T, B> for AvgPool3d<T, B>

Source§

impl<T: Float, B: BackendOps<T> + Default> Module<T, B> for Celu

Source§

impl<T: Float, B: BackendOps<T> + Default> Module<T, B> for ELU

Source§

impl<T: Float, B: BackendOps<T> + Default> Module<T, B> for GeLU

Implements the crate::module::Module interface for GeLU.

Source§

impl<T: Float, B: BackendOps<T> + Default> Module<T, B> for GeLUTanh

Source§

impl<T: Float, B: BackendOps<T> + Default> Module<T, B> for Hardshrink

Source§

impl<T: Float, B: BackendOps<T> + Default> Module<T, B> for Hardsigmoid

Implements the crate::module::Module interface for Hardsigmoid.

Source§

impl<T: Float, B: BackendOps<T> + Default> Module<T, B> for Hardswish

Implements the crate::module::Module interface for Hardswish.

Source§

impl<T: Float, B: BackendOps<T> + Default> Module<T, B> for Hardtanh

Source§

impl<T: Float, B: BackendOps<T> + Default> Module<T, B> for LeakyReLU

Source§

impl<T: Float, B: BackendOps<T> + Default> Module<T, B> for LogSigmoid

Implements the crate::module::Module interface for LogSigmoid.

Source§

impl<T: Float, B: BackendOps<T> + Default> Module<T, B> for MaxPool1d<T, B>

Source§

impl<T: Float, B: BackendOps<T> + Default> Module<T, B> for MaxPool2d<T, B>

Source§

impl<T: Float, B: BackendOps<T> + Default> Module<T, B> for MaxPool3d<T, B>

Source§

impl<T: Float, B: BackendOps<T> + Default> Module<T, B> for Mish

Implements the crate::module::Module interface for Mish.

Source§

impl<T: Float, B: BackendOps<T> + Default> Module<T, B> for SiLU

Implements the crate::module::Module interface for SiLU.

Source§

impl<T: Float, B: BackendOps<T> + Default> Module<T, B> for Sigmoid

Implements the crate::module::Module interface for Sigmoid.

Source§

impl<T: Float, B: BackendOps<T> + Default> Module<T, B> for Softplus

Implements the crate::module::Module interface for Softplus.

Source§

impl<T: Float, B: BackendOps<T> + Default> Module<T, B> for Softshrink

Source§

impl<T: Float, B: BackendOps<T> + Default> Module<T, B> for Softsign

Implements the crate::module::Module interface for Softsign.

Source§

impl<T: Float, B: BackendOps<T> + Default> Module<T, B> for Tanh

Implements the crate::module::Module interface for Tanh.

Source§

impl<T: Float, B: BackendOps<T> + Default> Module<T, B> for Threshold

Source§

impl<T: FloatElement + AttentionScalar + RealScalar, B: BackendOps<T> + AttentionOps<T> + Default, const H: usize, M: AttentionMask> Module<T, B> for TransformerEncoderLayer<T, B, H, M>

Source§

impl<T: FloatElement + AttentionScalar + RealScalar, B: BackendOps<T> + AttentionOps<T> + Default, const H: usize, const N: usize, M: AttentionMask> Module<T, B> for TransformerEncoder<T, B, H, N, M>

Source§

impl<T: FloatElement + Float + RealScalar, B: BackendOps<T> + Default, const DIM: usize> Module<T, B> for BatchNorm<T, B, DIM>

Source§

impl<T: FloatElement + Float + RealScalar, B: BackendOps<T> + Default, const DIM: usize> Module<T, B> for InstanceNorm<T, B, DIM>

Implements the Module interface for every InstanceNorm rank.

Source§

impl<T: FloatElement + Float + RealScalar, B: BackendOps<T> + Default, const G: usize> Module<T, B> for GroupNorm<T, B, G>

Implements the crate::module::Module interface for GroupNorm.

Source§

impl<T: FloatElement + Float + RealScalar, B: BackendOps<T> + Default> Module<T, B> for AdaptiveAvgPool1d<T, B>

Source§

impl<T: FloatElement + Float + RealScalar, B: BackendOps<T> + Default> Module<T, B> for AdaptiveAvgPool2d<T, B>

Source§

impl<T: FloatElement + Float + RealScalar, B: BackendOps<T> + Default> Module<T, B> for AdaptiveMaxPool1d<T, B>

Source§

impl<T: FloatElement + Float + RealScalar, B: BackendOps<T> + Default> Module<T, B> for AdaptiveMaxPool2d<T, B>

Source§

impl<T: FloatElement + Float + RealScalar, B: BackendOps<T> + Default> Module<T, B> for EmbeddingBag<T, B>

Source§

impl<T: FloatElement + Float + RealScalar, B: BackendOps<T> + Default> Module<T, B> for GlobalAvgPool1d<T, B>

Source§

impl<T: FloatElement + Float + RealScalar, B: BackendOps<T> + Default> Module<T, B> for GlobalAvgPool2d<T, B>

Source§

impl<T: FloatElement + Float + RealScalar, B: BackendOps<T> + Default> Module<T, B> for GlobalAvgPool3d<T, B>

Source§

impl<T: FloatElement + Float + RealScalar, B: BackendOps<T> + Default> Module<T, B> for GlobalMaxPool2d<T, B>

Source§

impl<T: FloatElement + Float + RealScalar, B: BackendOps<T> + Default> Module<T, B> for GlobalMaxPool3d<T, B>

Source§

impl<T: FloatElement + Float + RealScalar, B: BackendOps<T> + Default> Module<T, B> for LayerNorm<T, B>

Implements the crate::module::Module interface for LayerNorm.

Source§

impl<T: FloatElement + Float + RealScalar, B: BackendOps<T> + Default> Module<T, B> for RMSNorm<T, B>

Implements the crate::module::Module interface for RMSNorm.

Source§

impl<T: Scalar + RealScalar, B: BackendOps<T> + Default> Module<T, B> for Linear<T, B>

Source§

impl<T: Scalar, B: BackendOps<T> + Default> Module<T, B> for Embedding<T, B>

Source§

impl<T: Scalar, B: BackendOps<T> + Default> Module<T, B> for Fold1d<T, B>

Source§

impl<T: Scalar, B: BackendOps<T> + Default> Module<T, B> for Fold2d<T, B>

Source§

impl<T: Scalar, B: BackendOps<T> + Default> Module<T, B> for ReLU

Implements the crate::module::Module interface for ReLU.

Source§

impl<T: Scalar, B: BackendOps<T> + Default> Module<T, B> for Unfold1d<T, B>

Source§

impl<T: Scalar, B: BackendOps<T> + Default> Module<T, B> for Unfold2d<T, B>

Source§

impl<T: Scalar, B: ComputeBackend + Default> Module<T, B> for Sequential<T, B>