tenzr::nn::Module¶

Base class for neural-network layers holding parameters, submodules, and training state.

#include <module.h>

Inherited by tenzr::nn::BatchNorm, tenzr::nn::Conv2D, tenzr::nn::Dropout, tenzr::nn::Embedding, tenzr::nn::GELU, tenzr::nn::LayerNorm, tenzr::nn::Linear, tenzr::nn::LogSoftmax, tenzr::nn::RMSNorm, tenzr::nn::ReLU, tenzr::nn::Sequential, tenzr::nn::SiLU, tenzr::nn::Sigmoid, tenzr::nn::Softmax, tenzr::nn::Tanh

Public Functions¶

Protected Attributes¶

Public Functions Documentation¶

function ~Module¶

virtual ~Module() =default

function Forward¶

virtual Tensor Forward(
    const Tensor & input
) =0

Parameters:

• input Input tensor.

Return: Module output.

Reimplemented by: tenzr::nn::Embedding::Forward, tenzr::nn::ReLU::Forward, tenzr::nn::Sigmoid::Forward, tenzr::nn::Tanh::Forward, tenzr::nn::SiLU::Forward, tenzr::nn::GELU::Forward, tenzr::nn::Softmax::Forward, tenzr::nn::LogSoftmax::Forward, tenzr::nn::Conv2D::Forward, tenzr::nn::Dropout::Forward, tenzr::nn::Linear::Forward, tenzr::nn::LayerNorm::Forward, tenzr::nn::RMSNorm::Forward, tenzr::nn::BatchNorm::Forward, tenzr::nn::Sequential::Forward

Run the forward pass.

function operator()¶

inline Tensor operator()(
    const Tensor & input
)

function Train¶

inline virtual void Train(
    bool mode
)

Parameters:

• mode true for training, false for evaluation.

Set training mode and propagate it to all submodules.

function Train¶

inline void Train()

Switch to training mode.

function Eval¶

inline void Eval()

Switch to evaluation mode.

function IsTraining¶

inline bool IsTraining() const

Return: true when the module is in training mode.

function RegisterParameter¶

inline Tensor & RegisterParameter(
    std::string name,
    Tensor param
)

Parameters:

• name Parameter name.
• param Parameter tensor.

Return: Reference to the stored parameter.

Register a named trainable parameter.

function RegisterSubmodule¶

template <typename T ,
typename... Args>
inline std::shared_ptr< T > RegisterSubmodule(
    std::string name,
    Args &&... args
)

Parameters:

• name Submodule name.
• args Forwarded constructor arguments.

Return: Shared pointer to the new submodule.

Construct and register a submodule of type T.

function RegisterSubmodule¶

inline std::shared_ptr< Module > RegisterSubmodule(
    std::string name,
    std::shared_ptr< Module > submod
)

Parameters:

• name Submodule name.
• submod Submodule to register.

Return: The registered submodule.

Register an existing submodule instance.

function Parameters¶

inline std::vector< Tensor > Parameters(
    bool recurse =true
) const

Parameters:

• recurse Include submodule parameters when true.

Return: Flat list of parameter tensors.

Collect all registered parameters.

function NamedParameters¶

inline std::unordered_map< std::string, Tensor > NamedParameters(
    std::string prefix ="",
    bool recurse =true
) const

Parameters:

• prefix Prefix prepended to each name.
• recurse Include submodule parameters when true.

Return: Map from name to parameter tensor.

Collect parameters keyed by dotted fully-qualified name.

function StateDict¶

inline std::unordered_map< std::string, Tensor > StateDict() const

Return: Map from name to cloned parameter tensor.

Build a deep copy of all named parameters.

function LoadStateDict¶

inline void LoadStateDict(
    const std::unordered_map< std::string, Tensor > & state
)

Parameters:

• state Map from name to parameter tensor.

Copy values from state into matching parameters.

function Name¶

inline virtual std::string Name() const

Return: Human-readable module name.

Reimplemented by: tenzr::nn::ReLU::Name, tenzr::nn::Sigmoid::Name, tenzr::nn::Tanh::Name, tenzr::nn::SiLU::Name, tenzr::nn::GELU::Name, tenzr::nn::Softmax::Name, tenzr::nn::LogSoftmax::Name, tenzr::nn::Conv2D::Name, tenzr::nn::Dropout::Name, tenzr::nn::Embedding::Name, tenzr::nn::Linear::Name, tenzr::nn::LayerNorm::Name, tenzr::nn::RMSNorm::Name, tenzr::nn::BatchNorm::Name, tenzr::nn::Sequential::Name

Protected Attributes Documentation¶

variable parameters_¶

std::unordered_map< std::string, Tensor > parameters_;

variable submodules_¶

std::unordered_map< std::string, std::shared_ptr< Module > > submodules_;

variable training_¶

bool training_ {true};