tenzr::nn::Conv2D¶

2-D convolution module with configurable kernel, stride, and padding.

#include <conv2d.h>

Inherits from tenzr::nn::Module

Public Functions¶

Additional inherited members¶

Public Functions inherited from tenzr::nn::Module

Protected Attributes inherited from tenzr::nn::Module

Public Functions Documentation¶

function Conv2D¶

Conv2D(
    int64_t in_channels,
    int64_t out_channels,
    std::pair< int64_t, int64_t > kernel_size,
    std::pair< int64_t, int64_t > stride,
    std::pair< int64_t, int64_t > padding,
    bool use_bias,
    DType dtype,
    std::optional< uint64_t > seed =std::nullopt
)

Parameters:

• in_channels Number of input channels.
• out_channels Number of output channels.
• kernel_size (height, width) of the kernel.
• stride (height, width) stride.
• padding (height, width) zero padding.
• use_bias Add a bias term when true.
• dtype Parameter dtype.
• seed Optional PRNG seed.
• in_channels Number of input channels.
• out_channels Number of output filters.
• kernel_size Kernel height and width.
• stride Convolution stride.
• padding Zero padding.
• use_bias Whether to allocate a bias term.
• dtype Parameter dtype.
• seed Optional seed for weight initialization.

Construct with an explicit runtime dtype. Create a 2D convolution with Kaiming-uniform weights.

function Conv2D¶

template <typename T  =float>
inline Conv2D(
    int64_t in_channels,
    int64_t out_channels,
    std::pair< int64_t, int64_t > kernel_size,
    std::pair< int64_t, int64_t > stride ={1, 1},
    std::pair< int64_t, int64_t > padding ={0, 0},
    bool use_bias =true,
    std::optional< uint64_t > seed =std::nullopt
)

Construct with the parameter dtype inferred from T.

function Conv2D¶

template <typename T >
inline explicit Conv2D(
    std::type_identity< T > ,
    int64_t in_channels,
    int64_t out_channels,
    std::pair< int64_t, int64_t > kernel_size,
    std::pair< int64_t, int64_t > stride ={1, 1},
    std::pair< int64_t, int64_t > padding ={0, 0},
    bool use_bias =true,
    std::optional< uint64_t > seed =std::nullopt
)

Construct from an explicit std::type_identity dtype tag.

function Forward¶

virtual Tensor Forward(
    const Tensor & input
) override

Parameters:

• input Input of shape (N, in_channels, H, W).
• input Input tensor in NCHW layout.

Return:

• Output of shape (N, out_channels, H_out, W_out).
• Output tensor of shape [N, out_channels, out_h, out_w].

Reimplements: tenzr::nn::Module::Forward

Apply the convolution to input. Convolve a rank-4 [N, C, H, W] input with the layer's filters, adding bias.

function Weight¶

inline Tensor & Weight()

Return: Mutable reference to the weight tensor.

function Weight¶

inline const Tensor & Weight() const

Return: Const reference to the weight tensor.

function Bias¶

inline Tensor & Bias()

Return: Mutable reference to the bias tensor.

function Bias¶

inline const Tensor & Bias() const

Return: Const reference to the bias tensor.

function InChannels¶

inline int64_t InChannels() const

Return: Number of input channels.

function OutChannels¶

inline int64_t OutChannels() const

Return: Number of output channels.

function KernelSize¶

inline std::pair< int64_t, int64_t > KernelSize() const

Return: (height, width) kernel size.

function Stride¶

inline std::pair< int64_t, int64_t > Stride() const

Return: (height, width) stride.

function Padding¶

inline std::pair< int64_t, int64_t > Padding() const

Return: (height, width) zero padding.

function HasBias¶

inline bool HasBias() const

Return: true when the layer has a bias term.

function Name¶

inline virtual std::string Name() const override

Return: "Conv2D".

Reimplements: tenzr::nn::Module::Name

function Create¶

template <typename T  =float>
static inline Conv2D Create(
    int64_t in_channels,
    int64_t out_channels,
    std::pair< int64_t, int64_t > kernel_size,
    std::pair< int64_t, int64_t > stride ={1, 1},
    std::pair< int64_t, int64_t > padding ={0, 0},
    bool use_bias =true,
    std::optional< uint64_t > seed =std::nullopt
)

Static factory mirroring the templated constructor.