kh hddlZddlmZddlmZmZmZmZmZddl Z ddl m Z ddl m Z mZddlmZmZmZmZmZmZmZddlmZddlmZdd lmZdd lmZmZd gZ d Z!Gd de jDZ#Gdde jDZ$e jJdfde&de&dee jDde'de jPf dZ)Gdde jDZ*Gdde jDZ+Gdde jDZ,Gdde jDZ-Gd d!e jDZ.Gd"d#e jDZ/Gd$d%e jDZ0Gd&d'e jDZ1Gd(d)e jDZ2Gd*d e jDZ3dWd+Z4eid,e4d-.d/e4d-.d0e4d-.d1e4d-.d2e4d-.d3e4d-.d4e4d-d56d7e4d-.d8e4d-.d9e4d-.d:e4d-.d;e4d-.d<e4d-.d=e4d-d56d>e4d?d@d5dAdBdCDdEe4dFdGd5dAdBdCDdHe4dIdJd5dAdBdCDZ5dKZ6dXdLZ7edXdMZ8edXdNZ9edXdOZ:edXdPZ;edXdQZedXdTZ?edXdUZ@edXdVZAy)YN)partial)ListOptionalTupleTypeUnionIMAGENET_DEFAULT_MEANIMAGENET_DEFAULT_STD)DropPath trunc_normal_ create_conv2d ConvNormAct SqueezeExciteuse_fused_attnClassifierHead)build_model_with_cfg)feature_take_indices)checkpoint_seq)register_modelgenerate_default_cfgsFastVitc&|sy||zdk(sJ||zS)Nrr) group_sizechannelss O/var/www/teggl/fontify/venv/lib/python3.12/site-packages/timm/models/fastvit.py num_groupsrs( *$))):%%ceZdZdZddddddddej f dededed ed ed ed ed ededededeejddffd Z de jde jfdZ dZdee je jffdZdeej$ej&fdee je jffdZxZS)MobileOneBlocka#MobileOne building block. This block has a multi-branched architecture at train-time and plain-CNN style architecture at inference time For more details, please refer to our paper: `An Improved One millisecond Mobile Backbone` - https://arxiv.org/pdf/2206.04040.pdf rrFTin_chsout_chs kernel_sizestridedilationrinference_modeuse_seuse_actuse_scale_branchnum_conv_branches act_layerreturnNc tt| ||_t |||_||_||_||_||_ ||_ | |_ |r t|dntj|_|r"t!||||||j d|_nd|_||k(r|dk(rtj$|nd|_| dkDrstj(t+|jD cgc];} t-|j|j||j |j d =c} |_nd|_d|_|dkDr@| r>t-|j|jd|j |j d |_| r | |_ytj|_ycc} w) aConstruct a MobileOneBlock module. Args: in_chs: Number of channels in the input. out_chs: Number of channels produced by the block. kernel_size: Size of the convolution kernel. stride: Stride size. dilation: Kernel dilation factor. group_size: Convolution group size. inference_mode: If True, instantiates model in inference mode. use_se: Whether to use SE-ReLU activations. use_act: Whether to use activation. Default: ``True`` use_scale_branch: Whether to use scale branch. Default: ``True`` num_conv_branches: Number of linear conv branches. r) rd_divisorTr%r&r'groupsbiasN) num_featuresrFr%r&r2 apply_act)superr"__init__r(rr2r&r'r%r#r$r,rnnIdentityser reparam_conv BatchNorm2didentity ModuleListrangerconv_kxk conv_scaleact)selfr#r$r%r&r'rr(r)r*r+r,r-_ __class__s rr8zMobileOneBlock.__init__,s< nd,., V4    &  !2;A-A6bkkm  -'!{{!D !%D f$1F3 M!1$ "  %T%;%;< /   $/#{{#{{"'  / ! !% #DOQ#3"-KKLL !;;;;# ##*9;r{{}3 /s;AGxc|j/|j|j|j|Sd}|j|j|}d}|j|j |}||z}|j |j D] }|||z }|j|j|S)zApply forward pass.r)r<rCr;r>rBrA)rDrG identity_out scale_outoutrcs rforwardzMobileOneBlock.forwards    (88DGGD$5$5a$89: : == $==+L ?? &*I,& == $mm r!u  xx %%r c <|jy|j\}}t|j|j|j |j |j|jd|_||jj_ ||jj_ |jD]\}}d|vr |j|jd|jdt|dr|jdd|_y)aFollowing works like `RepVGG: Making VGG-style ConvNets Great Again` - https://arxiv.org/pdf/2101.03697.pdf. We re-parameterize multi-branched architecture used at training time to obtain a plain CNN-like structure for inference. NT) in_channels out_channelsr%r&r'r2r3r<rArBr>)r<_get_kernel_biasrr#r$r%r&r'r2weightdatar3named_parametersdetach_ __delattr__hasattrr()rDkernelr3nameparas rreparameterizezMobileOneBlock.reparameterizes    ( ,,. ) ((;;]];; )/  %&*#//1 JD$% LLN  $ & 4 $   Z ("r cd}d}|j[|j|j\}}|jdz}tjj j |||||g}d}d}|j|j|j\}}d}d}|jEt|jD]-}|j|j|\} } || z }|| z }/||z|z} ||z|z} | | fS)zMethod to obtain re-parameterized kernel and bias. Reference: https://github.com/DingXiaoH/RepVGG/blob/main/repvgg.py#L83 Returns: Tuple of (kernel, bias) after fusing branches. r) rB_fuse_bn_tensorr%torchr9 functionalpadr>rAr@r,) rD kernel_scale bias_scalerakernel_identity bias_identity kernel_conv bias_convix_kernel_bias kernel_final bias_finals rrQzMobileOneBlock._get_kernel_biass"  ?? &'+';';DOO'L $L*""a'C 88..22<#sCQTAUVL == $-1-A-A$---P *O]  == $D223 #!%!5!5dmmB6G!Hw& U"  # #\1OC +m; Z''r branchct|tr|jj}|jj }|jj }|jj}|jj}|jj}n2t|tjsJt|ds|j|jz}tj|j||j |j f|jj"|jj$} t'|jD](} d| | | |z|j dz|j dzf<*| |_|j(}|j }|j }|j}|j}|j}||zj+} || z j-dddd} || z|||z| z z fS)aMethod to fuse batchnorm layer with preceding conv layer. Reference: https://github.com/DingXiaoH/RepVGG/blob/main/repvgg.py#L95 Args: branch: Sequence of ops to be fused. Returns: Tuple of (kernel, bias) after fusing batchnorm. id_tensordtypedevicerr]) isinstancerconvrRbn running_mean running_varr3epsr9r=rWr#r2r_zerosr%rqrrr@rosqrtreshape) rDrmrXrwrxgammabetary input_dim kernel_valueistdts rr^zMobileOneBlock._fuse_bn_tensors fk *[[''F!9911L ))//KII$$E99>>D))--Cfbnn5 554- KK4;;6 ${{[[)T-=-=t?O?OP ----!==//  t{{+A!1y=$*:*:a*?AQAQUVAVV".^^F!..L ,,KMME;;D**CS &&( S[ ! !"aA .z4,"6"<<<#(  %)%&)+Z=Z=Z= Z=  Z=  Z=Z=!Z=Z=Z=#Z= #Z=BIIZ= Z=x&&%,,&2!#F!(% ell(B"C!(F)=BMM2>>9:)= u||U\\) *)=r r"cleZdZdZ ddedededededeed ed eejd ed dffd Z de jd e jfdZ d e e je jffdZddZedej"dej$d e e je jffdZxZS)ReparamLargeKernelConvzBuilding Block of RepLKNet This class defines overparameterized large kernel conv block introduced in `RepLKNet `_ Reference: https://github.com/DingXiaoH/RepLKNet-pytorch Nr#r$r%r&r small_kernelr)r-r(r.c Ltt| ||_t |||_||_||_||_||_ | r"t||||d|j d|_ nid|_ t||||j|j d|_ |6||ksJdt||||j|j d|_|r t|d nt!j"|_| ||_yt!j"|_y) a!Construct a ReparamLargeKernelConv module. Args: in_chs: Number of input channels. out_chs: Number of output channels. kernel_size: Kernel size of the large kernel conv branch. stride: Stride size. Default: 1 group_size: Group size. Default: 1 small_kernel: Kernel size of small kernel conv branch. act_layer: Activation module. Default: ``nn.GELU`` inference_mode: If True, instantiates model in inference mode. Default: ``False`` rTr1NFr5zDThe kernel size for re-param cannot be larger than the large kernel!g?)rd_ratio)r7rr8r&rr2r#r$r%rrr<r large_conv small_convrr9r:r;rC) rDr#r$r%r&rrr)r-r(rFs rr8zReparamLargeKernelConv.__init__s0 $d46  V4   &(  -'{{!D !%D )'{{{{ DO' K/ZYZ/"- ,;;;;# #