SushiAI::Autograd::GradCheckOptions

double tolerance = 1e-5

Relative-error ceiling for a coordinate to count as matching.

double step_scale = 6.1e-6

Step size relative to the coordinate's magnitude.

Near eps^(1/3) for binary64, which is where the truncation and cancellation terms of a central difference cross.

double min_step = 1e-8

Absolute floor on the step, for coordinates at or near zero.

double error_floor = 1e-6

Denominator floor in the relative-error ratio.

Below this the comparison falls back to an absolute one, so a gradient that is legitimately zero does not fail on a 1e-12 difference divided by itself.

double kink_margin = 10.0

How many step sizes a coordinate must clear a kink by.

std::uint64_t seed = 0x9E3779B97F4A7C15ull

Seed for the directional mode's random directions.