ATan Derivative

class src.model.SurrogateGradient.ATan(*args, **kwargs)

Bases: Function

Class that implements the SuperSpike function as the surrogate gradient for the backward pass.

static backward(ctx, grad_output: Tensor) → Tensor

Backward pass using surrogate gradient function:

static forward(ctx, input_: Tensor) → Tensor

Forward pass using 'Heaviside step function'

spiking_mode = 'binary'

surrogate_scale = 2.0