Skip to main content
Version: 0.2

LayerQuantizationSettings​

CLASS - LayerQuantizationSettings(

  • weights_num_bits: int =  8
  • activations_num_bits: int =  8
  • skip_tail_quantization: bool =  True
  • automatic_skip_quantization: bool =  True
  • skip_quantization: bool =  False
  • skip_quantization_downstream: bool =  False
  • skip_quantization_until: Union[str, Tuple[str], List[str], None] =  None

)

Ancestors - (BaseQATQuantizationSettings, BaseQuantizationSettings)

Quantization Settings for a specific Layer that can be set in Settings.

Class Variables​

  • weights_num_bits (int) - Number of Bits to use for the Weights if applicable.
  • activations_num_bits (int) - Number of Bits to use for the Activations if applicable.
  • skip_tail_quantization (bool) - Whether or not to automatically skip Quantization for the Tail of the Model. It is better to keep it 'True' if unsure.
  • automatic_skip_quantization (bool) - Whether or not to automatically skip Quantization for layers that are too sensitive. It is better to keep it 'True' if unsure. It does not override explicitly marked 'skipped_quantization = True' layers.
  • skip_quantization (bool) - Whether or not to Skip Quantization for a specific Layer.
  • skip_quantization_downstream (bool) - Whether or not to Skip Quantization for the current Layer and everything below it in the Graph.
  • skip_quantization_until (Union[str, Tuple[str], List[str], None]) - Skip Quantization from the layer name the LayerQuantizationSettings is assigned to until another layer name