U Kcm@sddlZddlmZddlmmmZddlZddlmm m Z ej ej hZ ej ej hZddZddZddZGdd d e jZedfd d ZdS) NcCst||S)zQ Given name of submodule, this function grabs the submodule from given model )dict named_modules)modelnamerG/tmp/pip-unpacked-wheel-gikjz4vx/torch/ao/quantization/_correct_bias.py get_module srcCs8|dd}t|dkr$d|dfS|d|dfSdS)zYSplits full name of submodule into parent submodule's full name and submodule's name .rN)rsplitlen)rZ split_namerrrparent_child_namess   rcCs"t||d}t|r|S|SdS)z Sometimes the weights/bias attribute gives you the raw tensor, but sometimes gives a function that will give you the raw tensor, this function takes care of that logic N)getattrcallable)moduleattrparamrrr get_params rcs0eZdZdZfddZddZddZZS)MeanShadowLoggerzA logger for a Shadow module whose purpose is to record the rolling mean of the data passed to the floating point and quantized models cs8tt|d|jd<d|jd<d|_d|_d|_dSNfloat quantizedr)superr__init__statscount float_sum quant_sumself __class__rrr's   zMeanShadowLogger.__init__cCs|jr|}|jd7_|jddkr<||jd<||_n |j|7_|j|j|jd<|jddkr|||jd<||_n |j|7_|j|j|jd<dS)z The inputs x,y are output data from the quantized and floating-point modules. x is for the quantized module, y is for the floating point module r rNr)Z is_quantizedZ dequantizerrrr)r xyrrrforward/s  zMeanShadowLogger.forwardcCs*d|jd<d|jd<d|_d|_d|_dSr)rrrrrrrrclearEs   zMeanShadowLogger.clear)__name__ __module__ __qualname____doc__rr%r& __classcell__rrr!rr#s rcCs.t||tti}|D]\}}t||kr|||<q|D]}t||} t| d} | dk r>d} |D]$} || d| d7} | |krfqqft|} t |\}}| |dd}| |dd}||}t t | }| dt||}| j|}|| _|D]\}}t|tr|qq>dS)ah Using numeric suite shadow module, the expected output of the floating point and quantized modules is recorded. Using that data the bias of supported modules is shifted to compensate for the drift caused by quantization Paper reference: https://arxiv.org/pdf/1906.04721.pdf (Section 4.2) Args: float_model: a trained model that serves as a reference to what bias correction should aim for quantized_model: quantized form of float_model that bias correction is to applied to img_data: calibration data to estimate the expected output (used to find quantization error) target_modules: specifies what submodules in quantized_model need bias correction (can be extended to unquantized submodules) neval_batches: a cap to the number of batches you want to be used for estimating the expected output biasNrr z.statsrr)nsZprepare_model_with_stubs_supported_modulesrrtyperrZget_logger_dictrlistrangeZdimremovetorchZmeandata isinstancer&)Z float_modelZquantized_modelZimg_dataZtarget_modulesZ neval_batchesZuncorrected_modulesr submoduleZuncorrected_moduleZquantized_submoduler,rr4Zob_dict parent_name_Z float_dataZ quant_dataZquantization_errorZdimsZexpected_errorZ updated_biasrrrbias_correctionLs8           r9)r3Ztorch.nnnnZtorch.ao.nn.quantizedZaorZnnqZtorch.ao.quantizationZtorch.ao.ns._numeric_suiter-Z_numeric_suiteZLinearZConv2dr.Z_supported_modules_quantizedrrrLoggerrr9rrrrs     )