SqnrEncodingAnalyzer

class aimet_torch.v2.quantization.encoding_analyzer.SqnrEncodingAnalyzer(shape, num_bins=2048, *, asymmetric_delta_candidates=17, symmetric_delta_candidates=101, offset_candidates=21, max_parallelism=64, gamma=3.0)[source]

Encoding Analyzer for SQNR Calibration technique

Parameters:

  • shape (tuple) – Shape of calculated encoding

  • num_bins (int) – number of bins to use per histogram

  • asymmetric_delta_candidates – number of delta values to search over in asymmetric mode

  • symmetric_delta_candidates – number of delta values to search over in symmetric mode

  • offset_candidates – number of offset values to search over in asymmetric mode

  • max_parallelism – maximum number of encodings to process parallely (higher number results in higher memory usage but faster computation)

  • gamma – weighting factor on clipping noise (higher value results in less clipping noise)

compute_encodings_from_stats(stats, num_steps, is_symmetric)[source]

Searches for encodings which produce the lowest expected SQNR based on the histograms in stats

Parameters:

  • stats (List[_Histogram]) – A list of _Histogram objects with length equal to the number of encodings to compute

  • num_steps (int) – The number of bins the quantized range is split into

  • is_symmetric (bool) – If True, computes symmetric encodings, else computes asymmetric encodings

Return type:

Tuple[Optional[Tensor], Optional[Tensor]]

Returns:

Tuple of computed encodings (min, max) as tensors with shape self.shape