nerva_torch.softmax_functions

Softmax and log-softmax functions together with stable variants.

This module provides both function-only forms and simple callable classes.

Functions

log_softmax(X)	Row-wise log-softmax (numerically unsafe version).
log_softmax_jacobian(x)	Jacobian matrix of log_softmax for a single row vector.
softmax(X)	Row-wise softmax with explicit normalization (numerically unsafe).
softmax_jacobian(x)	Jacobian matrix of softmax for a single row vector.
stable_log_softmax(X)	Row-wise log-softmax with max-subtraction for stability.
stable_log_softmax_jacobian(x)	Jacobian matrix of stable log_softmax (same as log_softmax).
stable_softmax(X)	Row-wise softmax using max-subtraction for numerical stability.
stable_softmax_jacobian(x)	Jacobian matrix of stable softmax for a single row vector.

Classes

LogSoftmaxFunction()	Callable implementing row-wise log-softmax and its Jacobian.
SoftmaxFunction()	Callable implementing row-wise softmax and its Jacobian.
StableLogSoftmaxFunction()	Callable implementing numerically stable row-wise log-softmax and its Jacobian.
StableSoftmaxFunction()	Callable implementing numerically stable row-wise softmax and its Jacobian.

nerva_torch.softmax_functions.softmax(X: torch.Tensor) → torch.Tensor

Row-wise softmax with explicit normalization (numerically unsafe).

nerva_torch.softmax_functions.softmax_jacobian(x: torch.Tensor) → torch.Tensor

Jacobian matrix of softmax for a single row vector.

nerva_torch.softmax_functions.stable_softmax(X: torch.Tensor) → torch.Tensor

Row-wise softmax using max-subtraction for numerical stability.

nerva_torch.softmax_functions.stable_softmax_jacobian(x: torch.Tensor) → torch.Tensor

Jacobian matrix of stable softmax for a single row vector.

nerva_torch.softmax_functions.log_softmax(X: torch.Tensor) → torch.Tensor

Row-wise log-softmax (numerically unsafe version).

nerva_torch.softmax_functions.log_softmax_jacobian(x: torch.Tensor) → torch.Tensor

Jacobian matrix of log_softmax for a single row vector.

nerva_torch.softmax_functions.stable_log_softmax(X: torch.Tensor) → torch.Tensor

Row-wise log-softmax with max-subtraction for stability.

nerva_torch.softmax_functions.stable_log_softmax_jacobian(x: torch.Tensor) → torch.Tensor

Jacobian matrix of stable log_softmax (same as log_softmax).

class nerva_torch.softmax_functions.SoftmaxFunction

Bases: object

Callable implementing row-wise softmax and its Jacobian.

jacobian(X: torch.Tensor) → torch.Tensor

class nerva_torch.softmax_functions.StableSoftmaxFunction

Bases: object

Callable implementing numerically stable row-wise softmax and its Jacobian.

jacobian(X: torch.Tensor) → torch.Tensor

class nerva_torch.softmax_functions.LogSoftmaxFunction

Bases: object

Callable implementing row-wise log-softmax and its Jacobian.

jacobian(X: torch.Tensor) → torch.Tensor

class nerva_torch.softmax_functions.StableLogSoftmaxFunction

Bases: object

Callable implementing numerically stable row-wise log-softmax and its Jacobian.

jacobian(X: torch.Tensor) → torch.Tensor