surrogate#

How to use surrogate#

class DynapSEtorch.surrogate.FastSigmoid(*args, **kwargs)[source]#

Bases: Function

Fast-sigmoid surrogated gradient Apply the fast-sigmoid gradient as a surrogated gradient for the heavyside step function.

\[\frac{\partial S}{\partial V} = \frac{1}{(\lambda \left|v\right| + 1.0)^2}\]

Where \(\lambda\) is a scale factor with default value 10.

static pseudo_derivative(v)[source]#

Compute the gradient of the fast-sigmoid function.

class DynapSEtorch.surrogate.Step(*args, **kwargs)[source]#

Bases: Function

Step function surrogated gradient Use the step function as a surrogated gradient of itself

static pseudo_derivative(V)[source]#

Compute the step function surrogate gradient.

class DynapSEtorch.surrogate.Triangular(*args, **kwargs)[source]#

Bases: Function

Triangular surrogated gradient Apply the triangular function as a surrogated gradient for the heavyside step function.

\[\frac{\partial S}{\partial V} = \lambda max(1 - \left|V\right|, 0)\]

Where \(\lambda\) is a scale factor with default value 0.3.

static pseudo_derivative(v)[source]#

Compute the triangular surrogate gradient.