o p i @s<dZddlmZddlmZddlmZGdddeZd S) z Layers used for QAT. ) functional)Layer)ConvertibleQuantedLayercsFeZdZdZdeffdd ZddZddZd d Zd d Z Z S) QuantedConv2Dz The computational logic of QuantizedConv2D is the same as Conv2D. The only difference is that its inputs are all fake quantized. layercst|j|_|j|_|j|_|j|_|jdkr|j|_|j|_|j|_|j |_ |j |_ d|_ d|_ |j dur@|j ||_ |jdurN|j ||_ dSdS)Nzeros)super__init___groups_stride_padding _padding_mode _reversed_padding_repeated_twice _dilation _data_formatweightbiasweight_quanteractivation_quanterZ _instanceZ activation)selfrZq_config __class___/home/app/PaddleOCR-VL/.venv_paddleocr/lib/python3.10/site-packages/paddle/nn/quant/qat/conv.pyr s&    zQuantedConv2D.__init__cCs@|}|j}|jdur||}|jdur||j}|||S)N)rrr _conv_forward)rinputZ quant_inputZ quant_weightrrrforward6s     zQuantedConv2D.forwardc CsN|jdkrtj||j|j|jd}d|_tj|||j|j|j|j |j |jdS)Nr )mode data_formatr)rpaddingZstrideZdilationgroupsr ) rFpadrrrZconv2drr rr )rZinputsweightsrrrr?s$ zQuantedConv2D._conv_forwardcCdgS)N)rrrrrrrweights_to_quantersTz!QuantedConv2D.weights_to_quanterscCr&)Nrrr'rrractivation_quantersWr)z!QuantedConv2D.activation_quanters) __name__ __module__ __qualname____doc__rr rrr(r* __classcell__rrrrrs rN) r.Z paddle.nnrr#Z layer.layersrformatrrrrrrs