Skip to content

tenzr::nn

Classes

Name
class tenzr::nn::BatchNorm
Batch normalization module over the channel dimension.
class tenzr::nn::BCELoss
Callable wrapper around the binary cross-entropy loss.
class tenzr::nn::Conv2D
2-D convolution module with configurable kernel, stride, and padding.
class tenzr::nn::CrossEntropyLoss
Callable wrapper around the cross-entropy loss over logits.
class tenzr::nn::Dropout
Dropout regularization module active only in training mode.
class tenzr::nn::Embedding
Lookup-table module mapping integer indices to dense vectors.
class tenzr::nn::GELU
Gaussian error linear unit activation module.
class tenzr::nn::LayerNorm
Layer normalization module over the last dimension.
class tenzr::nn::Linear
Fully connected affine module applying a weight and optional bias.
class tenzr::nn::LogSoftmax
Log-softmax module normalizing over a configurable axis.
class tenzr::nn::Module
Base class for neural-network layers holding parameters, submodules, and training state.
class tenzr::nn::MSELoss
class tenzr::nn::ReLU
Rectified linear unit activation module.
class tenzr::nn::RMSNorm
Root-mean-square normalization module.
class tenzr::nn::Sequential
Container module that applies its layers in order.
class tenzr::nn::Sigmoid
Logistic sigmoid activation module.
class tenzr::nn::SiLU
SiLU (swish) activation module.
class tenzr::nn::Softmax
Softmax module normalizing over a configurable axis.
class tenzr::nn::Tanh
Hyperbolic tangent activation module.

Functions

Name
Array softmax(const Array & input, int64_t axis =-1)
Array log_softmax(const Array & input, int64_t axis =-1)
Tensor MseLoss(const Tensor & pred, const Tensor & target, std::string reduction ="mean")
Tensor BceLoss(const Tensor & pred, const Tensor & target, double eps =1.0e-7, std::string reduction ="mean")
Tensor cross_entropy_loss(const Tensor & logits, const Tensor & targets, double eps =1.0e-7, std::string reduction ="mean")

Functions Documentation

function softmax

Array softmax(
    const Array & input,
    int64_t axis =-1
)

Parameters:

  • input Input tensor.
  • axis Axis to normalize over.
  • input Input tensor.
  • axis Reduction axis, supporting negative indices.

Return:

  • Softmax probabilities.
  • Softmax-normalized tensor.

Numerically stable softmax along axis. Numerically stable softmax along axis.

function log_softmax

Array log_softmax(
    const Array & input,
    int64_t axis =-1
)

Parameters:

  • input Input tensor.
  • axis Axis to normalize over.
  • input Input tensor.
  • axis Reduction axis, supporting negative indices.

Return:

  • Log-softmax values.
  • Log-softmax tensor.

Numerically stable log-softmax along axis. Numerically stable log-softmax along axis.

function MseLoss

Tensor MseLoss(
    const Tensor & pred,
    const Tensor & target,
    std::string reduction ="mean"
)

Parameters:

  • pred Predictions.
  • target Ground-truth values.
  • reduction "mean", "sum", or "none".
  • pred Predicted tensor.
  • target Reference tensor with the same shape.
  • reduction One of "mean", "sum", or "none".

Return:

  • Scalar (or unreduced) loss.
  • Scalar loss, or the elementwise squared error when reduction == "none".

Mean squared error between pred and target. Mean squared error between pred and target.

function BceLoss

Tensor BceLoss(
    const Tensor & pred,
    const Tensor & target,
    double eps =1.0e-7,
    std::string reduction ="mean"
)

Parameters:

  • pred Predicted probabilities.
  • target Binary ground-truth values.
  • eps Clamp epsilon guarding log(0).
  • reduction "mean", "sum", or "none".
  • pred Predicted probabilities in [0, 1].
  • target Reference labels with the same shape.
  • eps Clamp bound keeping logarithms finite.
  • reduction One of "mean", "sum", or "none".

Return:

  • Scalar (or unreduced) loss.
  • Scalar loss, or the elementwise loss when reduction == "none".

Binary cross-entropy between probabilities pred and targets in [0, 1]. Binary cross-entropy between probabilities pred and targets target.

function cross_entropy_loss

Tensor cross_entropy_loss(
    const Tensor & logits,
    const Tensor & targets,
    double eps =1.0e-7,
    std::string reduction ="mean"
)

Parameters:

  • logits Unnormalized scores of shape (..., num_classes).
  • targets Integer class indices.
  • eps Clamp epsilon for the log-softmax.
  • reduction "mean", "sum", or "none".
  • logits Unnormalized scores of shape [B, C].
  • targets Class indices [B] or one-hot probabilities [B, C].
  • eps Unused; kept for interface symmetry.
  • reduction One of "mean", "sum", or "none".

Return:

  • Scalar (or unreduced) loss.
  • Scalar loss, or the per-sample loss when reduction == "none".

Cross-entropy between unnormalized logits and integer class targets. Softmax cross-entropy from rank-2 logits.