MovingAveragePerChannelMinMaxObserver

Per-channel observer with an EMA of the per-channel min / max.

The per-channel analogue of MovingAverageMinMaxObserver, and the outlier-tolerant sibling of PerChannelMinMaxObserver. It keeps an independent range per channel along ch_axis (reducing over every other axis) but updates each channel's [min, max] with an exponential moving average instead of a hard cumulative min/max — so a spike in one calibration batch nudges that channel's range only slightly and is relaxed back by later batches. calculate_qparams still returns a length-C vector of (scale, zero_point).

When to pick it. Use it for weights when weight calibration is noisy (e.g. running statistics observed during QAT rather than on static weights) and you want per-channel tightness and per-batch robustness. For static weights the plain PerChannelMinMaxObserver is usually enough; for activations, use the per-tensor MovingAverageMinMaxObserver or HistogramObserver.

With $x^{(c)}$ the values of channel $c$ in the current batch, the first batch seeds each channel directly and every later batch applies the per-channel EMA:

where $c$ is averaging_constant. The reduction to per-channel qparams is the (default symmetric, signed) map inherited from PerChannelMinMaxObserver, pinning each z_c = 0.