:py:mod:`neural_compressor.adaptor.ox_utils.calibrator`
=======================================================

.. py:module:: neural_compressor.adaptor.ox_utils.calibrator

.. autoapi-nested-parse::

   Calibrator for onnx models.



Module Contents
---------------

Classes
~~~~~~~

.. autoapisummary::

   neural_compressor.adaptor.ox_utils.calibrator.CalibratorBase
   neural_compressor.adaptor.ox_utils.calibrator.MinMaxCalibrator
   neural_compressor.adaptor.ox_utils.calibrator.PercentileCalibrator
   neural_compressor.adaptor.ox_utils.calibrator.KLCalibrator
   neural_compressor.adaptor.ox_utils.calibrator.HistogramCollector



Functions
~~~~~~~~~

.. autoapisummary::

   neural_compressor.adaptor.ox_utils.calibrator.calib_registry
   neural_compressor.adaptor.ox_utils.calibrator.smooth_distribution



.. py:function:: calib_registry(calib_method)

   The class decorator used to register all Calibrator subclasses.


.. py:class:: CalibratorBase


   Base calibrator class.


.. py:class:: MinMaxCalibrator




   MinMax calibrator class.


.. py:class:: PercentileCalibrator(num_bins=2048, percentile=99.999)




   Percentile calibrator class.

   :param num_bins: number of bins to create a new histogram
                    for collecting tensor values. Defaults to 2048.
   :type num_bins: int, optional
   :param percentile: A float number between [0, 100]. Defaults to 99.999.
   :type percentile: float, optional


.. py:class:: KLCalibrator(num_bins=128, num_quantized_bins=128)




   KL calibrator class.

   :param num_bins: number of bins to create a new histogram
                    for collecting tensor values. Defaults to 128.
   :type num_bins: int, optional
   :param num_quantized_bins: number of quantized bins. Defaults to 128.
   :type num_quantized_bins: int, optional


.. py:class:: HistogramCollector(num_bins=2048)


   Histogram collctor class.


.. py:function:: smooth_distribution(p, eps=0.0001)

   Smooth distribution.

   Given a discrete distribution (may have not been normalized to 1),
   smooth it by replacing zeros with eps multiplied by a scaling factor
   and taking the corresponding amount off the non-zero values.
   Ref:
   http://hanj.cs.illinois.edu/cs412/bk3/KL-divergence.pdf
   https://github.com//apache/incubator-mxnet/blob/master/python/mxnet/contrib/quantization.py
   https://github.com/microsoft/onnxruntime/blob/main/onnxruntime/python/tools/quantization/calibrate.py

   :param p: distribution array
   :type p: array
   :param eps: a small probability. Defaults to 0.0001.
   :type eps: float, optional

   :returns: smoothed distribution
   :rtype: array