Value Transforms¶
Abstract base class for value transforms. |
|
A simple invertible log-transform. |
This module contains some useful value transforms. These are functions
that can be used to rescale or warp the returns for more a more robust training
signal, see e.g. coax.value_transforms.LogTransform
.
Object Reference¶
- class coax.value_transforms.ValueTransform(transform_func, inverse_func)[source]¶
Abstract base class for value transforms. See
coax.value_transforms.LogTransform
for a specific implementation.- property inverse_func¶
The inverse transformation function \(y\mapsto x=f^{-1}(y)\).
- Parameters:
y (ndarray) – The values in their transformed representation.
- Returns:
x (ndarray) – The values in their original representation.
- property transform_func¶
The transformation function \(x\mapsto y=f(x)\).
- Parameters:
x (ndarray) – The values in their original representation.
- Returns:
y (ndarray) – The values in their transformed representation.
- class coax.value_transforms.LogTransform(scale=1.0)[source]¶
A simple invertible log-transform.
\[x\ \mapsto\ y\ =\ \lambda\,\text{sign}(x)\, \log\left(1+\frac{|x|}{\lambda}\right)\]with inverse:
\[y\ \mapsto\ x\ =\ \lambda\,\text{sign}(y)\, \left(\text{e}^{|y|/\lambda} - 1\right)\]This transform logarithmically supresses large values \(|x|\gg1\) and smoothly interpolates to the identity transform for small values \(|x|\sim1\) (see figure below).
- Parameters:
scale (positive float, optional) – The scale \(\lambda>0\) of the linear-to-log cross-over. Smaller values for \(\lambda\) translate into earlier onset of the cross-over.
- property inverse_func¶
The inverse transformation function \(y\mapsto x=f^{-1}(y)\).
- Parameters:
y (ndarray) – The values in their transformed representation.
- Returns:
x (ndarray) – The values in their original representation.
- property transform_func¶
The transformation function \(x\mapsto y=f(x)\).
- Parameters:
x (ndarray) – The values in their original representation.
- Returns:
y (ndarray) – The values in their transformed representation.