rrelu_(x: Tensor, lower: float = 1.0 / 8, upper: float = 1.0 / 3, training: bool = False, inplace: bool = True)In-place RReLU — randomised leaky ReLU written back into x.
During training the negative-slope is sampled per-element from a uniform ; at inference the midpoint is used.
Parameters
Input; modified in place.
lowerfloat= 1.0 / 8Lower bound of the uniform slope sampler. Default
1/8.upperfloat= 1.0 / 3Upper bound of the uniform slope sampler. Default
1/3.trainingbool= FalseWhen
False (default) uses the deterministic midpoint slope;
when True samples a fresh slope per element.inplacebool= TrueIgnored. Default
True.Returns
TensorThe same x tensor, now holding rrelu(x, lower, upper, training).