wi ddlZddlmcmZddlmZGddejZGddejZGddejZGd d ejZ Gd d ejZ ddZ dS)N)nnc(eZdZ dfd ZdZxZS) ConvBNActTc t||_tj|||||dz dz|d|_tj||_|jrtj|_ dSdS)NrF)paddinggroupsbias) super__init__use_actrConv2dconv BatchNorm2dbnReLUact)self in_channels out_channels kernel_sizestrider r __class__s }/var/www/development/aibuddy-work/election-extract/venv/lib/python3.11/site-packages/rapidocr/networks/backbones/rec_hgnet.pyr zConvBNAct.__init__s  I     1_*    ... < !wyyDHHH ! !c||}||}|jr||}|SN)rrrrrxs rforwardzConvBNAct.forwards> IIaLL GGAJJ <  Ar)rT__name__ __module__ __qualname__r r! __classcell__rs@rrrsOPT!!!!!!$rrc$eZdZfdZdZxZS) ESEModulecttjd|_tj||ddd|_tj|_dS)Nrr)rrrrr ) r r rAdaptiveAvgPool2davg_poolrrSigmoidsigmoid)rchannelsrs rr zESEModule.__init__"se ,Q// I !     z|| rc|}||}||}||}||zSr)r,rr.)rr identitys rr!zESEModule.forward.s@ MM!   IIaLL LLOO8|rr"r's@rr)r)!sG $ $ $ $ $rr)c(eZdZ dfd ZdZxZS)HG_BlockFc t||_tj|_|jt||ddt|dz D]-}|jt||dd.|||zz}t||dd|_ t||_ dS)Nrrrrr) r r r1r ModuleListlayersappendrrangeaggregation_convr)att) rr mid_channelsr layer_numr1_total_channelsrs rr zHG_Block.__init__7s    moo   ')       y1}%%  A K   ,!- !     %y<'?? )&% ! ! !  \**rc$|}g}|||jD]"}||}||#tj|d}||}||}|jr||z }|S)Nr)dim)r9r8torchcatr;r<r1)rr r1outputlayers rr!zHG_Block.forward_s a[  EaA MM!     If! $ $ $  ! !! $ $ HHQKK =  MAr)Fr"r's@rr3r36sR &+&+&+&+&+&+P       rr3c.eZdZdddgffd ZdZxZS)HG_StageTrrc ~t||_|rt||d||d|_g}|t ||||dt |dz D])} |t ||||d*tj||_ dS)Nr5F)rrrrr r)r1rT) r r downsamplerr9r3r:r Sequentialblocks) rrr=r block_numr>rJr blocks_listr?rs rr zHG_Stage.__init__os $  ''(" DO  [, iRW X X X   y1}%%  A    , iRV     m[1 rch|jr||}||}|Sr)rJrLrs rr!zHG_Stage.forwards1 ? #""A KKNNrr"r's@rrHrHnsY1v 2 2 2 2 2 2DrrHc6eZdZdZ dfd ZdZdZxZS) PPHGNeta PPHGNet Args: stem_channels: list. Stem channel list of PPHGNet. stage_config: dict. The configuration of each stage of PPHGNet. such as the number of channels, stride, etc. layer_num: int. Number of layers of HG_Block. use_last_conv: boolean. Whether to use a 1x1 convolutional layer before the classification layer. class_expand: int=2048. Number of channels for the last 1x1 convolutional layer. dropout_prob: float. Parameters of dropout, 0.0 means dropout is not used. class_num: int=1000. The number of classes. Returns: model: nn.Layer. Specific PPHGNet model depends on args. r5FNc t||_||ngd|_d|t jfdttdz D|_ |jrt j ddd|_ t j |_ g|_t|D]d\}}||\}} } } } } |j t#|| | | || | ||jvr|j| e|js|dd|_|dS) N)rrrr5rc dg|],}t||dzd|dkrdnd-S)rr5rrr6)r).0i stem_channelss r z$PPHGNet.__init__..s`  -a 0!.q1u!5 ! !Q11A rrr5r)rrr stage4)r r det out_indicesinsertrrKr:lenstem MaxPool2dpoolr7stagesr enumerater9rH _init_weights)rrV stage_configr>rrYrZblock_idkr=rrMrJrrs ` rr zPPHGNet.__init__s *5*A;;||| Q ,,,Ms=11A566    8 I 1aHHHDImoo $\22 7 7KHaQ  K       4+++!((666x : ,X 6q 9D  rc|D]}t|tjr%tj|jAt|tjrItj|jtj |j t|tj r$tj |j dSr) modules isinstancerrinitkaiming_normal_weightrones_zeros_r Linear)rms rrbzPPHGNet._init_weightss ' 'A!RY'' '''1111Ar~.. ' ah'''qv&&&&Ary)) 'qv&&& ' 'rc~||}|jr||}g}t|jD]5\}}||}|jr||jvr||6|jr|S|jrtj |ddg}ntj |ddg}|S)Nr(r5r) r]rYr_rar`rZr9trainingFadaptive_avg_pool2d avg_pool2d)rr outrUstages rr!zPPHGNet.forwards IIaLL 8  ! A!$+..  HAuaAx A!111 1 8 J = (%a!R11AA QA''Ar)r5FN)r#r$r%__doc__r rbr!r&r's@rrQrQsq  & ::::::x'''rrQFc gdgdgdgdd}ddddd d dggdd d dd dd ggd d dd d d dggddddd d dggd}tdgd|r|n|d|d|}|S)a{ PPHGNet_small Args: pretrained: bool=False or str. If `True` load pretrained parameters, `False` otherwise. If str, means the path of the pretrained model. use_ssld: bool=False. Whether using distillation pretrained model when pretrained=True. Returns: model: nn.Layer. Specific `PPHGNet_small` model depends on args. )rzrFr)r{rTr)r}rTr)rrTr)stage1stage2stage3rXrzr{rTrr|r}r~rrr)@rrz)rVrcr>rY)rQ) pretraineduse_ssldrYkwargsstage_config_detstage_config_recmodels r PPHGNet_smallrs/..------... S!TAq62S!TAq62S!TAq62T1dQF3   #mm),B%%2B        E Lr)FFF) rCtorch.nn.functionalr functionalrsModulerr)r3rHrQrrrrrs@  6 *55555ry555p'''''ry'''TfffffbifffR!!!!!!r