class

ELU

extends_QuantizedActivation

ELU(scale: Tensor, zero_point: Tensor, qdtype: QDtype = quint8)

source edit

Quantized lucid.nn.ELU — elu(x, alpha) then output fake-quant.

The inference-time replacement for a calibrated float nn.ELU, installed by lucid.quantization.convert / from_float. Under the sidecar representation (design B) the quantized layer is the float ELU followed by a fake-quantize to the calibrated output grid. Unlike the bounded activations, ELU is unbounded above and saturates to $-\alpha$ below, so its range is data-dependent and must be calibrated with a real observer rather than a lucid.quantization.FixedQParamsObserver.

The negative-branch scale $\alpha$ is a parameter of the activation, not a learned weight; from_float copies it from the source nn.ELU (via the internal _configure hook) so the quantized layer reproduces the exact float curve before snapping the output to the grid.

y = \operatorname{fake\_quant}\bigl(\operatorname{elu}(x)\bigr), \qquad \operatorname{elu}(x) = \begin{cases} x & x > 0 \\ \alpha\,(e^{x} - 1) & x \le 0 \end{cases}

with $\operatorname{fake\_quant}(t) = (\operatorname{clamp}( \operatorname{round}(t/S) + Z, q_{\min}, q_{\max}) - Z)\, S$ and the calibrated output (scale, zero_point) $S, Z$ .

Parameters

scaleTensor

Per-tensor output scale from the calibrating observer. Supplied by from_float; construct directly only with qparams in hand.

zero_pointTensor

Per-tensor output zero-point from the calibrating observer.

qdtypeQDtype= quint8

Quantized dtype bounding the grid. Defaults to lucid.quantization.quint8 ([0, 255]).

Attributes

alphafloat

The negative-branch scale

\alpha

\alpha\,(e^{x} - 1)

, copied from the source nn.ELU by from_float. Defaults to 1.0 on a bare instance.

Notes

Built via convert / from_float, which reads the output observer and copies alpha from the float module.
Range is data-dependent (unbounded above), so calibrate through the prepared model before convert; a fixed grid would clip the upper tail.
The forward is float elu(x, alpha) → fake-quant; only the output is snapped to the grid.

Examples

>>> import lucid, lucid.nn as nn
>>> import lucid.quantization as Q
>>> model = nn.Sequential(nn.ELU(alpha=1.5))
>>> model.qconfig = Q.get_default_qconfig()
>>> prepared = Q.prepare(model)               # attach the output observer
>>> _ = prepared(lucid.randn(32, 8))          # calibrate its range
>>> qelu = Q.convert(prepared)[0]             # from_float reads the observer
>>> type(qelu).__name__, qelu.alpha
('ELU', 1.5)

Used by 1

lucid.nn.quantized

>>> import lucid, lucid.nn as nn >>> import lucid.quantization as Q >>> model = nn.Sequential(nn.ELU(alpha=1.5)) >>> model.qconfig = Q.get_default_qconfig() >>> prepared = Q.prepare(model) # attach the output observer >>> _ = prepared(lucid.randn(32, 8)) # calibrate its range >>> qelu = Q.convert(prepared)[0] # from_float reads the observer >>> type(qelu).__name__, qelu.alpha ('ELU', 1.5)

ELU

Parameters

Attributes

Notes

Examples

See Also

Used by 1

ELU

Parameters

Attributes

Notes

Examples

See Also

Used by 1