Transform¶

class paddle.distribution. Transform [source]

Base class for the transformations of random variables.

Transform can be used to represent any differentiable and injective function from the subset of \(R^n\) to subset of \(R^m\), generally used for transforming a random sample generated by Distribution instance.

Suppose \(X\) is a K-dimensional random variable with probability density function \(p_X(x)\). A new random variable \(Y = f(X)\) may be defined by transforming \(X\) with a suitably well-behaved function \(f\). It suffices for what follows to note that if f is one-to-one and its inverse \(f^{-1}\) have a well-defined Jacobian, then the density of \(Y\) is

\[p_Y(y) = p_X(f^{-1}(y)) |det J_{f^{-1}}(y)|\]

where det is the matrix determinant operation and \(J_{f^{-1}}(y)\) is the Jacobian matrix of \(f^{-1}\) evaluated at \(y\). Taking \(x = f^{-1}(y)\), the Jacobian matrix is defined by

\[\begin{split}J(y) = \begin{bmatrix} {\frac{\partial x_1}{\partial y_1}} &{\frac{\partial x_1}{\partial y_2}} &{\cdots} &{\frac{\partial x_1}{\partial y_K}} \\ {\frac{\partial x_2}{\partial y_1}} &{\frac{\partial x_2} {\partial y_2}}&{\cdots} &{\frac{\partial x_2}{\partial y_K}} \\ {\vdots} &{\vdots} &{\ddots} &{\vdots}\\ {\frac{\partial x_K}{\partial y_1}} &{\frac{\partial x_K}{\partial y_2}} &{\cdots} &{\frac{\partial x_K}{\partial y_K}} \end{bmatrix}\end{split}\]

A Transform can be characterized by three operations:

forward Forward implements \(x \rightarrow f(x)\), and is used to convert one random outcome into another.

inverse Undoes the transformation \(y \rightarrow f^{-1}(y)\).

log_det_jacobian The log of the absolute value of the determinant of the matrix of all first-order partial derivatives of the inverse function.

Subclass typically implement follow methods:

_forward

_inverse

_forward_log_det_jacobian

_inverse_log_det_jacobian (optional)

If the transform changes the shape of the input, you must also implemented:

_forward_shape

_inverse_shape

forward ( x: Tensor ) → Tensor forward¶

Forward transformation with mapping \(y = f(x)\).

Useful for turning one random outcome into another.

Parameters: x (Tensor) – Input parameter, generally is a sample generated from Distribution.
Returns: Outcome of forward transformation.
Return type: Tensor

inverse ( y: Tensor ) → Tensor inverse¶

Inverse transformation \(x = f^{-1}(y)\). It’s useful for “reversing” a transformation to compute one probability in terms of another.

Parameters: y (Tensor) – Input parameter for inverse transformation.
Returns: Outcome of inverse transform.
Return type: Tensor

forward_log_det_jacobian ( x: Tensor ) → Tensor forward_log_det_jacobian¶

The log of the absolute value of the determinant of the matrix of all first-order partial derivatives of the inverse function.

Parameters: x (Tensor) – Input tensor, generally is a sample generated from Distribution
Returns: The log of the absolute value of Jacobian determinant.
Return type: Tensor

inverse_log_det_jacobian ( y: Tensor ) → Tensor inverse_log_det_jacobian¶

Compute \(log|det J_{f^{-1}}(y)|\). Note that forward_log_det_jacobian is the negative of this function, evaluated at \(f^{-1}(y)\).

Parameters: y (Tensor) – The input to the inverse Jacobian determinant evaluation.
Returns: The value of \(log|det J_{f^{-1}}(y)|\).
Return type: Tensor

forward_shape ( shape: Sequence[int] ) → Sequence[int] forward_shape¶

Infer the shape of forward transformation.

Parameters: shape (Sequence[int]) – The input shape.
Returns: The output shape.
Return type: Sequence[int]

inverse_shape ( shape: Sequence[int] ) → Sequence[int] inverse_shape¶

Infer the shape of inverse transformation.

Parameters: shape (Sequence[int]) – The input shape of inverse transformation.
Returns: The output shape of inverse transformation.
Return type: Sequence[int]