Hardsigmoid

Quantized lucid.nn.Hardsigmoid — hardsigmoid(x) then fake-quant.

The inference-time replacement for a calibrated float nn.Hardsigmoid, installed by lucid.quantization.convert / from_float. Hardsigmoid is the piecewise-linear approximation of the sigmoid gate used in MobileNetV3's squeeze-and-excitation blocks; like Sigmoid its output is bounded to $[0, 1]$, so it is a natural fixed-qparams activation. Under the sidecar representation (design B) the quantized layer is the float activation followed by a fake-quantize to the calibrated output grid.

Because the range is fixed to $[0, 1]$, pair it with a lucid.quantization.FixedQParamsObserver (scale $1/256$, zero-point $0$) to skip calibrating a range known a priori.

with $\operatorname{fake\_quant}(t) = (\operatorname{clamp}( \operatorname{round}(t/S) + Z, q_{\min}, q_{\max}) - Z)\, S$ and, for the fixed grid, $S = 1/256,\ Z = 0$.