qat_config Module

TrainingMethod

Select a QAT training method:

class model_compression_toolkit.qat.TrainingMethod(value)

An enum for selecting a QAT training method

STE - Standard straight-through estimator. Includes PowerOfTwo, symmetric & uniform quantizers

DQA - DNN Quantization with Attention. Includes a smooth quantization introduces by DQA method

LSQ - Learned Step size Quantization. Includes PowerOfTwo, symmetric & uniform quantizers: https://arxiv.org/pdf/1902.08153.pdf


QATConfig

Class to configure the quantization process of the model when quantizing using Quantization-aware Training (QAT):

class model_compression_toolkit.qat.QATConfig(weight_training_method=TrainingMethod.STE, activation_training_method=TrainingMethod.STE, weight_quantizer_params_override=None, activation_quantizer_params_override=None)

QAT configuration class.

Parameters:

  • weight_training_method (TrainingMethod) – Training method for weight quantizers

  • activation_training_method (TrainingMethod) – Training method for activation quantizers:

  • weight_quantizer_params_override – A dictionary of parameters to override in weight quantization quantizer instantiation. Defaults to None (no parameters)

  • activation_quantizer_params_override – A dictionary of parameters to override in activation quantization quantizer instantiation. Defaults to None (no parameters)