NgG ddlZddlmZddlmZmZmZmZddlZddl m Z ddl m Z m Z ddlmZmZmZmZmZmZmZddlmZddlmZdd lmZdd lmZmZd gZd Z Gd de j!Z"Gdde j!Z#e j$dfde%de%de j!de&de j'f dZ(Gdde j!Z)Gdde j!Z*Gdde j!Z+Gdde j!Z,Gd d!e j!Z-Gd"d#e j!Z.Gd$d%e j!Z/Gd&d'e j!Z0Gd(d)e j!Z1Gd*d e j!Z2dXd,Z3eid-e3d./d0e3d./d1e3d./d2e3d./d3e3d./d4e3d./d5e3d.d67d8e3d./d9e3d./d:e3d./d;e3d./d<e3d./d=e3d./d>e3d.d67d?e3d@dAd6dBdCdDEdFe3dGdHd6dBdCdDEdIe3dJdKd6dBdCdDEZ4dLZ5dYdMZ6edYdNZ7edYdOZ8edYdPZ9edYdQZ:edYdRZ;edYdSZedYdVZ?edYdWZ@dS)ZN)partial)ListOptionalTupleUnionIMAGENET_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*|sdS||zdksJ||zS)Nrr) group_sizechannelss O/var/www/html/ai-engine/env/lib/python3.11/site-packages/timm/models/fastvit.py num_groupsrs2 &q*$)))):%%c2eZdZdZddddddddejf dededed ed ed ed ed ededededejddffd Z de j de j fdZ dZ dee j e j ffdZdeejejfdee j e j ffdZxZS)MobileOneBlocka#MobileOne building block. This block has a multi-branched architecture at train-time and plain-CNN style architecture at inference time For more details, please refer to our paper: `An Improved One millisecond Mobile Backbone` - https://arxiv.org/pdf/2206.04040.pdf rrFTin_chsout_chs kernel_sizestridedilationrinference_modeuse_seuse_actuse_scale_branchnum_conv_branches act_layerreturnNc .tt|_t ||_|_|__|_ |_ | _ |rt|dntj_|r!t!||||jd_nd_||kr|dkrtj|nd_| dkr9tjfdt+j D_nd_d_dkr0| r.t1j j djjd _| r | ntj_dS) aConstruct a MobileOneBlock module. Args: in_chs: Number of channels in the input. out_chs: Number of channels produced by the block. kernel_size: Size of the convolution kernel. stride: Stride size. dilation: Kernel dilation factor. group_size: Convolution group size. inference_mode: If True, instantiates model in inference mode. use_se: Whether to use SE-ReLU activations. use_act: Whether to use activation. Default: ``True`` use_scale_branch: Whether to use scale branch. Default: ``True`` num_conv_branches: Number of linear conv branches. r) rd_divisorTr$r%r&groupsbiasN) num_featuresrc bg|]+}tjjjjd,S)Fr$r%r1 apply_act)rr"r#r%r1).0_r$selfs r z+MobileOneBlock.__init__..msT / / /   $/#{#{"'  / / /rFr5)superr!__init__r'rr1r%r&r$r"r#r+rnnIdentityser reparam_conv BatchNorm2didentity ModuleListrangeconv_kxk conv_scaleract)r9r"r#r$r%r&rr'r(r)r*r+r, __class__s` ` rr<zMobileOneBlock.__init__,s< nd##,,..., V44    &  !2;AS-A6666bkmm -  -'!{!!!D  !%D f$$1F3333 M!1$$ " / / / / / %T%;<< / / / ! ! !% #DOQ#3"-KL !;;# ####*<99;;;r{}}rxc|j;||||Sd}|j||}d}|j||}||z}|j|jD]}|||z }|||S)zApply forward pass.Nr)r@rGr?rBrFrE)r9rI identity_out scale_outoutrcs rforwardzMobileOneBlock.forwards   (88DGGD$5$5a$8$899:: : = $==++L ? &**I,& = $m  rr!uu xx %%%rc |jdS|\}}t|j|j|j|j|j|jd|_||jj _ ||jj _ | D]\}}d|vr | |d|dt|dr|dd|_dS)aFollowing works like `RepVGG: Making VGG-style ConvNets Great Again` - https://arxiv.org/pdf/2101.03697.pdf. We re-parameterize multi-branched architecture used at training time to obtain a plain CNN-like structure for inference. NT) in_channels out_channelsr$r%r&r1r2r@rErFrB)r@_get_kernel_biasr r"r#r$r%r&r1weightdatar2named_parametersdetach_ __delattr__hasattrr')r9kernelr2nameparas rreparameterizezMobileOneBlock.reparameterizes   ( F,,.. ) (;];   )/ %&*#//11  JD$%% LLNNNN $$$ &&& 4 $ $ )   Z ( ( ("rcd}d}|jP||j\}}|jdz}tjj|||||g}d}d}|j||j\}}d}d}|jDt|j D]/}||j|\} } || z }|| z }0||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. rN) rF_fuse_bn_tensorr$torchr= functionalpadrBrErDr+) r9 kernel_scale bias_scalerckernel_identity bias_identity kernel_conv bias_convix_kernel_bias kernel_final bias_finals rrSzMobileOneBlock._get_kernel_biass  ? &'+';';DO'L'L $L*"a'C 8.22<#sCQTAUVVL = $-1-A-A$--P-P *O]  = $D233 # #!%!5!5dmB6G!H!Hw& U" "\1OC +m; Z''rbranchct|trI|jj}|jj}|jj}|jj}|jj}|jj}nt|tj sJt|ds|j |j z}tj|j ||j|jf|jj|jj} t'|j D]} d| | | |z|jdz|jdzf<| |_|j}|j}|j}|j}|j}|j}||z} || z dddd} || z|||z| z z fS)a Method to fuse batchnorm layer with preceeding conv layer. Reference: https://github.com/DingXiaoH/RepVGG/blob/main/repvgg.py#L95 Args: branch: Sequence of ops to be fused. Returns: Tuple of (kernel, bias) after fusing batchnorm. id_tensordtypedevicerr_) isinstancerconvrTbn running_mean running_varr2epsr=rArYr"r1razerosr$rsrtrDrqsqrtreshape) r9rorZryrzgammabetar{ input_dim kernel_valueistdts rr`zMobileOneBlock._fuse_bn_tensors fk * * ['F!91L )/KI$E9>D)-CCfbn55 5 5 54-- . K4;6 ${[)T-=t?OP --!=/    t{++A!1y=$*:a*?AQUVAVV".^F!.L ,KME;D*CS &&(( S[ ! !"aA . .z4,"6"<<<9:)= u|U\) *)=)=)=)=)=)=)=)=rr!c0eZdZdZ ddededededed eed ed eejd ed dffd Z de j d e j fdZ d e e j e j ffdZddZedejdejd e e j e j ffdZxZS)ReparamLargeKernelConvzBuilding Block of RepLKNet This class defines overparameterized large kernel conv block introduced in `RepLKNet `_ Reference: https://github.com/DingXiaoH/RepLKNet-pytorch NFr"r#r$r%r small_kernelr(r,r'r-c ltt|||_t |||_||_||_||_||_ | r!t||||d|jd|_ nad|_ t||||j|jd|_ |4||ks Jdt||||j|jd|_|rt|d nt!j|_| |nt!j|_dS) a!Construct a ReparamLargeKernelConv module. Args: in_chs: Number of input channels. out_chs: Number of output channels. kernel_size: Kernel size of the large kernel conv branch. stride: Stride size. Default: 1 group_size: Group size. Default: 1 small_kernel: Kernel size of small kernel conv branch. act_layer: Activation module. Default: ``nn.GELU`` inference_mode: If True, instantiates model in inference mode. Default: ``False`` rTr0NFr5zDThe kernel size for re-param cannot be larger than the large kernel!g?)rd_ratio)r;rr<r%rr1r"r#r$rr r@r large_conv small_convrr=r>r?rG) r9r"r#r$r%rrr(r,r'rHs rr<zReparamLargeKernelConv.__init__s^0 $d++44666  V44   &(   -'{!!!D  !%D )'{{ DO' K///Y0//"- ,;;# ###