o پiA@sdZddlmZmZmZddlmZddlmZm Z m Z m Z m Z ddl Z ddlmZddlmZmZddlmZmZmZmZmZmZdd lmZdd lmZmZdd lm Z m!Z!d gZ"eGd ddZ#ddZ$eGdddZ%eGdddZ&         dddZ'Gdddej(Z)Gdddej(Z*Gdd d ej(Z+Gd!d"d"ej(Z,Gd#d$d$ej(Z-Gd%d&d&ej(Z.d'd(d'd)d*d*ej/ej0df d+d,Z1d-d.Z2d/d0Z3d1d2Z4d3e&d4e5d5e6d6e e7effd7d8Z8Gd9d d ej(Z9dd:d;Z:e;didd?d@dAe%dBdCdDdEdFdGdHdIdJe&e#dKd'd?d@dAe%dBdCdDdEdFddGdLdIdMe&e#dKd'd?d@dAe%dBdNdDddOdFdGdLdIdPe&e#d=d>d?d@dAe%dBdNdDd(dddFdGdQdIdRe&e#d(d'dd*dAe%dSdTd)dUdVdWdGddXdIdYe&e#d(d'dd*dAe%dZdTd[d\d]ddd^dId_e&e#d(d'dd*dAe%d`dTd[d\d]ddd^dIdae&e#d(d'dd*dAe%d`dTd)dFddbdddcdIdde&e#d(d'dd*dAe%dSdTd)dFdddd^dIdee&e#d(d'dd*dAe%dSdTd)dFddGdddfdIdge'dFdFdhdie'djdkdhdle'dme'dndodhdpe'dFdFdGdqdre'djdkdGdqdse'dGdtdue'dndodGdqdve'dbe;dOdwdxdye'dbdndodzd{e&e#d|d'd)d*dAe%d}dNd)d~dFdFdbdddde'dndodddde'dndodddbe;dOdwdZe!ide>dddde>de>de>dddde>dddde>de>de>de>ddddddde>dddddde>ddde>ddddddde>ddddddde>dddddddde>dddddddde>ddddddde>dddddde>dde>dddddde>dddddde>dde>dddddde>ddddddddZ?e dde9fdd<Z@e dde9fddJZAe dde9fddMZBe dde9fddPZCe dde9fddRZDe dde9fddYZEe dde9fdd_ZFe dde9fddaZGe dde9fdddZHe dde9fddeZIe dde9fddgZJe dde9fddiZKe dde9fddlZLe dde9fddmZMe dde9fddpZNe dde9fddrZOe dde9fddsZPe dde9fdduZQe dde9fddvZRe dde9fddyZSe dde9fdd{ZTe dde9fddZUe dde9fddZVdS)aPyTorch CspNet A PyTorch implementation of Cross Stage Partial Networks including: * CSPResNet50 * CSPResNeXt50 * CSPDarkNet53 * and DarkNet53 for good measure Based on paper `CSPNet: A New Backbone that can Enhance Learning Capability of CNN` - https://arxiv.org/abs/1911.11929 Reference impl via darknet cfg files at https://github.com/WongKinYiu/CrossStagePartialNetworks Hacked together by / Copyright 2020 Ross Wightman ) dataclassasdictreplace)partial)AnyDictOptionalTupleUnionNIMAGENET_DEFAULT_MEANIMAGENET_DEFAULT_STD)ClassifierHead ConvNormActDropPathget_attncreate_act_layermake_divisible)build_model_with_cfg) named_applyMATCH_PREV_GROUP)register_modelgenerate_default_cfgsCspNetc@sveZdZUdZeeeedffed<dZeeeedffed<dZ eed<dZ eee fed <dZ e e ed <d S) CspStemCfg .out_chsstride kernel_sizepaddingpoolN)__name__ __module__ __qualname__rr intr __annotations__rr!r#strr$rr+r+F/home/ubuntu/.local/lib/python3.10/site-packages/timm/models/cspnet.pyrs  rcCsNt|ttfs |f}t|}||}|dkr|d|St||df|S)Nr) isinstancetuplelistlen)xncurr_npad_nr+r+r,_pad_arg(s r6c@seZdZUdZeedfed<dZeedfed<dZe eeedffed<dZ e eeedffed <d Z e e ee dffed <d Z e e ee dffed <d Ze eeedffed<dZee eeedffed<dZee eeefed<dZe eeefed<dZe eeefed<d Ze e ee dffed<d Ze eeedffed<d Ze eeedffed<ddZdS) CspStagesCfgr r r.depthrrrrgroups? block_ratio bottle_ratioFavg_downN attn_layer attn_kwargscsp stage_typebottle block_type expand_ratio cross_linear down_growthcCst|j}t|j|ksJt|j||_t|j||_t|j||_t|j||_t|j||_t|j ||_ t|j ||_ t|j ||_ t|j ||_ t|j ||_ t|j||_t|j||_dSN)r1r:rr6rr@rBrCrDrErFrHrJrKrLrM)selfr3r+r+r, __post_init__Fs zCspStagesCfg.__post_init__)r%r&r'r:r r(r)rrr r@rBfloatrCrDboolrErr*rFrrHrJrKrLrMrPr+r+r+r,r73s   r7c@sReZdZUeed<eed<dZeed<dZe ed<dZ e ed<d Z e e ed <d S) CspModelCfgstemstagesTzero_init_last leaky_relu act_layer batchnorm norm_layerNaa_layer) r%r&r'rr)r7rVrRrXr*rZr[rr+r+r+r,rSXs    rSrAFsiludarkc s|rttdddddd} nttfdddDd ddd } t| ttfd dd Dtfd ddDd|d|||d|d |dS)N@rr")rr!rr#r$cg|]}t|qSr+r.0cwidth_multiplierr+r, sz_cs3_cfg..rr^r rr!rr$cr`r+rarbrer+r,rgxrhr;cr`r+)r()rcd)depth_multiplierr+r,rgyrh)r r_ r ?cs3) rr:rrCrBrDrErFrHrJrTrUrX)rrr/rSr7) rfrlrDrXfocusrErFrCrJstem_cfgr+)rlrfr,_cs3_cfgbs2   rsc sHeZdZdZdddejejddddf fdd Zd d Zd d Z Z S) BottleneckBlockz ResNe(X)t Bottleneck Block r?FNc stt|tt||} t||d} | duo|}| duo!| }t|| fddi| |_t| | fd||| d| |_|rE| | |dnt |_ t| |fddd| |_ |r_| ||dnt |_ | rjt| nt |_t||_dS) NrXrZr!rr r!dilationr@ drop_layerrXFr! apply_act)superrt__init__r(rounddictrconv1conv2nnIdentityattn2conv3attn3r drop_pathract3)rOin_chsrryrCr@rXrZ attn_lastrE drop_blockrmid_chsckwargs attn_first __class__r+r,rs&  zBottleneckBlock.__init__cCtj|jjjdSrN)rinitzeros_rbnweightrOr+r+r,rVzBottleneckBlock.zero_init_lastcCsR|}||}||}||}||}||}|||}||}|SrN)rrrrrrrrOr2shortcutr+r+r,forwards      zBottleneckBlock.forward r%r&r'__doc__rReLU BatchNorm2drrVr __classcell__r+r+rr,rtsrtcFeZdZdZdddejejdddffdd Zdd Zd d Z Z S) DarkBlockz DarkNet Block rrnNrvc stt|tt||} t||d} t|| fddi| |_|dur+|| |dnt |_ t| |fd||| d| |_ | rHt | |_ dSt |_ dS)Nrwr!rr{r rx)r~rrr(rrrrrrattnrrr rOrrryrCr@rXrZrErrrrrr+r,rs   zDarkBlock.__init__cCrrNrrrrrrrr+r+r,rVrzDarkBlock.zero_init_lastcC4|}||}||}||}|||}|SrNrrrrrr+r+r,r    zDarkBlock.forwardrr+r+rr,rrcr) EdgeBlockzZ EdgeResidual / Fused-MBConv / MobileNetV1-like 3x3 + 1x1 block (w/ activated output) rrnNrvc stt|tt||} t||d} t|| fd||| d| |_|dur.|| |dnt |_ t| |fddi| |_ | rHt | |_ dSt |_ dS)Nrwr rxr{r!r)r~rrr(rrrrrrrrrrrrr+r,rs   zEdgeBlock.__init__cCrrNrrr+r+r,rVrzEdgeBlock.zero_init_lastcCrrNrrr+r+r,rrzEdgeBlock.forwardrr+r+rr,rrrc >eZdZdZdddddddddef fdd Zdd ZZS) CrossStagez Cross Stage.rArNFc stt|| p |} | r|n|}tt|||_}tt||}t|d|dd}|dd}|dks>| |krq| r^t |dkrKt dnt t ||fdd| d||_nt ||fd|| | |d ||_|}nt |_|}t ||fd| d ||_|d}t |_t|D]!}|jt||d||||| |dur||nd d ||}qt ||dfd di||_t ||fd di||_dSNrXrZrwr[rrr!rr@r r!rryr@r[r|rvrrryrCr@rr!r+)r~rrr(r expand_chsrgetpopr Sequential AvgPool2drr conv_downconv_expblocksrange add_moduler*conv_transition_bconv_transitionrOrrrryr:rBrCrKr@first_dilationrDrMrL block_dprblock_fn block_kwargsdown_chsexp_chs block_out_chs conv_kwargsr[prev_chsirr+r,r sR      zCrossStage.__init__cCs`||}||}|j|jddd\}}||}||}|tj ||gdd}|SNrr)dim) rrsplitrrr contiguousrtorchcat)rOr2xsxboutr+r+r,rMs   zCrossStage.forwardr%r&r'rrtrrrr+r+rr,r sArc r) CrossStage3z`Cross Stage 3. Similar to CrossStage, but with only one transition conv for the output. rArNFc stt|| p |} | r|n|}tt|||_}tt||}t|d|dd}|dd}|dks>| |krq| r^t |dkrKt dnt t ||fdd| d||_nt ||fd|| | |d ||_|}nd|_|}t ||fd| d ||_|d}t |_t|D]!}|jt||d||||| |dur||nd d ||}qt ||fd di||_dSr)r~rrr(rrrrrrrrrrrrrrrr*rrrr+r,r[sP     zCrossStage3.__init__cCsR||}||}|j|jddd\}}||}|tj||gdd}|Sr)rrrrrrrr)rOr2x1x2rr+r+r,rs   zCrossStage3.forwardrr+r+rr,rWs >rcs8eZdZdZdddddedffdd Zdd ZZS) DarkStagezDarkNet stage.rArNFc  stt|| p |} t| d| dd}| dd}| r=t|dkr*tdnt t ||fdd|d||_ nt ||fd|| ||d ||_ |}t t ||}t|_t|D]!}|jt|| d |||||| durw| |nd d | |}q`dS) NrXrZrwr[rrrr rrvrr+)r~rrrrrrrrrrrr(rrrrr*)rOrrrryr:rBrCr@rrDrrrrr[rrrrr+r,rsB    zDarkStage.__init__cC||}||}|SrN)rrrOr2r+r+r,r  zDarkStage.forwardrr+r+rr,rs .rr rrr"c  Cst} g} t|ttfs|g}t|} | sJ|dvsJd} |} | d}d}t|D]U\}}d|d}|dkr?|dksI||krK|dkrK|sKdnd}|dkrZ| durZ| | | |t | ||||dkrh|nd||d||9}|} t | |d d |gd } q,|r|dksJ| dur| | |dur| d tj d ddd | d|| ddd}n| d tj d ddd d }|d9}t | |d d |gd } | | | | fS)N)rrrconvrrr")rr#rXrZ.rTnum_chs reductionmoduler$r )r!rr#aa)channelsr) rrr.r/r0r1 enumerateappendrrrjoin MaxPool2d)in_chansrr!rr$r#rXrZr[rT feature_info stem_depth prev_featrlast_idx stem_striderchs conv_name conv_stride pool_namer+r+r,create_csp_stemsP  ,    rcCsn|d}|dvs J|dkr'|dd|dd|ddt}||fS|dkr1t}||fSt}||fS)NrH)r]rGror]rKrLrMrG)rrrr) stage_argsrHstage_fnr+r+r, _get_stage_fns     rcCs>|d}|dvs J|dkrt|fS|dkrt|fSt|fS)NrJ)r]edgerIr]r)rrrrt)rrJr+r+r, _get_block_fns  rcCsF|d}|ddp i}|durt|}|rt|fi|}||fS)NrErF)rrr)rrErFr+r+r, _get_attn_fn)s rcfgdrop_path_rate output_stride stem_featc st|jt|jj}|sdg|nddtd|t|jj|jjDd<fddt D}t |j |j d}d}|d}|d } |} g} g} t |D]e\} }t|\}}t|\}}t|\}}|d }|dkrx| rx| | ||kr|dkr||9}d}||9}|d vrdnd }| || fi||||||j|d |g7} |d} t | |d| d} qR| | tj| | fS)NcSsg|]}|qSr+)tolist)rcr2r+r+r,rg<sz%create_csp_stages..rrcsg|] }tt|qSr+)rzipkeys)rcvaluescfg_dictr+r,rg=srwrrrrrrr)rrryrr[rErzstages.r)rrUr1r:rlinspacesumrrrrrXrZrrrrrrr[rr)rrrr num_stagesrrry net_striderrrrU stage_idxrrattn_fnrrr+rr,create_csp_stages3s\  (         r cseZdZdZ       ddeffd d Zejjd d d Z ejjd!ddZ ejjde j fddZ d"dedeefddZddZd defddZddZZS)#raCross Stage Partial base model. Paper: `CSPNet: A New Backbone that can Enhance Learning Capability of CNN` - https://arxiv.org/abs/1911.11929 Ref Impl: https://github.com/WongKinYiu/CrossStagePartialNetworks NOTE: There are differences in the way I handle the 1x1 'expansion' conv in this impl vs the darknet impl. I did it this way for simplicity and less special cases. r ravgrvTrc  st||_||_|dvsJt|fi| }t|j|j|jd} g|_ t |fit |j | \|_ } |j | ddt|||| dd\|_} | dd} |j | | |_|_t| |||d|_ttt|d|dS) a Args: cfg (CspModelCfg): Model architecture configuration in_chans (int): Number of input channels (default: 3) num_classes (int): Number of classifier classes (default: 1000) output_stride (int): Output stride of network, one of (8, 16, 32) (default: 32) global_pool (str): Global pooling type (default: 'avg') drop_rate (float): Dropout rate (default: 0.) drop_path_rate (float): Stochastic depth drop-path rate (default: 0.) zero_init_last (bool): Zero-init last weight of residual path kwargs (dict): Extra kwargs overlayed onto cfg )r)rXrZr[Nr-)rrrr) in_features num_classes pool_type drop_rate)rV)r~rrrrrrXrZr[rrrrTextendr rU num_featureshead_hidden_sizerheadrr _init_weights)rOrrrr global_poolrrrVkwargs layer_argsstem_feat_infostage_feat_inforrr+r,rrs:      zCspNet.__init__FcCs"td|rdnddtfdgd}|S)Nz^stem^stages\.(\d+))z^stages\.(\d+)\.blocks\.(\d+)Nz^stages\.(\d+)\..*transition)r)r)rTr)rr)rOcoarsematcherr+r+r, group_matcherszCspNet.group_matchercCs|rJddS)Nz$gradient checkpointing not supportedr+)rOenabler+r+r,set_grad_checkpointingszCspNet.set_grad_checkpointingreturncCs|jjSrN)rfcrr+r+r,get_classifierszCspNet.get_classifierNrrcCs||_|j||dSrN)rrreset)rOrrr+r+r,reset_classifierszCspNet.reset_classifiercCrrNrTrUrr+r+r,forward_featuresrzCspNet.forward_features pre_logitscCs|r |j||dS||S)N)r))r)rOr2r)r+r+r, forward_headszCspNet.forward_headcCrrN)r(r*rr+r+r,rrzCspNet.forward)r r rr rvrvTF)TrN)r%r&r'rrSrrjitignorerr!rModuler$r(rr*r&r(rRr*rrr+r+rr,rhs, >  cCst|tjr tjj|jddd|jdurtj|jdSdSt|tjr@tjj |jddd|jdur>tj|jdSdS|rMt |drO| dSdSdS)Nfan_outrelu)mode nonlinearityrvg{Gz?)meanstdrV) r.rConv2drkaiming_normal_rbiasrLinearnormal_hasattrrV)rnamerVr+r+r,rs     r cspresnet50r^rmaxrjr8r;r@rnT)r:rrrKrCrLr' cspresnet50d)rrr^)r:rrrKrCrBrL cspresnet50w)r=r>r?iru cspresnext50)r:rrr@rKrCrBrL cspdarknet53)rrr r r)r^r<r=r>r?)r?rA)rnrA)rArn)r:rrrKrCrBrMrJ darknet17)rrrrr)r)rn)rA)r:rrrCrBrHrJ darknet21)rrrrr sedarknet21se)r:rrrCrBrErHrJ darknet53 darknetaa53)r:rrrCrBrDrHrJ cs3darknet_s)rfrl cs3darknet_mg?gq= ףp? cs3darknet_l cs3darknet_xg?gHzG?cs3darknet_focus_s)rfrlrqcs3darknet_focus_mcs3darknet_focus_l)rqcs3darknet_focus_xcs3sedarknet_l)rd_ratio)rErFcs3sedarknet_x)rErfrlcs3sedarknet_xdwri)r r_ r)rrr=r>)r:rrr@rCrBrErp cs3edgenet_xg?r)rfrlrCrJcs3se_edgenet_x)rfrlrCrJrErFcKsP|ds |dr d}nd}|d|}tt||ft|td|dd|S) Ndarknet cspdarknet)rrrr rr9)rrrr r out_indicesT)flatten_sequentialr[) model_cfg feature_cfg) startswithrrr model_cfgsr)variant pretrainedrdefault_out_indicesr[r+r+r,_create_cspnets  rdc Ks|dddddttddd |S) Nr )r r=r=)r r gMb?bilinearzstem.conv1.convzhead.fc) urlr input_size pool_sizecrop_pct interpolationr3r4 first_conv classifierr )rfrr+r+r,_cfgsrmzcspresnet50.ra_in1kztimm/zlhttps://github.com/rwightman/pytorch-image-models/releases/download/v0.1-weights/cspresnet50_ra-d3e8d487.pth) hf_hub_idrfzcspresnet50d.untrainedzcspresnet50w.untrainedzcspresnext50.ra_in1kzqhttps://github.com/rwightman/pytorch-image-models/releases/download/v0.1-weights/cspresnext50_ra_224-648b4713.pthzcspdarknet53.ra_in1kzqhttps://github.com/rwightman/pytorch-image-models/releases/download/v0.1-weights/cspdarknet53_ra_256-d05c7c21.pthzdarknet17.untrainedzdarknet21.untrainedzsedarknet21.untrainedzdarknet53.c2ns_in1kzthttps://github.com/rwightman/pytorch-image-models/releases/download/v0.1-tpu-weights/darknet53_256_c2ns-3aeff817.pthbicubic)r rp)rnrfrjtest_input_size test_crop_pctzdarknetaa53.c2ns_in1kzrhttps://github.com/rwightman/pytorch-image-models/releases/download/v0.1-tpu-weights/darknetaa53_c2ns-5c28ec8a.pth)rnrfrqrrzcs3darknet_s.untrained)rjzcs3darknet_m.c2ns_in1kzshttps://github.com/rwightman/pytorch-image-models/releases/download/v0.1-tpu-weights/cs3darknet_m_c2ns-43f06604.pthgffffff?zcs3darknet_l.c2ns_in1kzshttps://github.com/rwightman/pytorch-image-models/releases/download/v0.1-tpu-weights/cs3darknet_l_c2ns-16220c5d.pthzcs3darknet_x.c2ns_in1kzshttps://github.com/rwightman/pytorch-image-models/releases/download/v0.1-tpu-weights/cs3darknet_x_c2ns-4e4490aa.pth)rnrfrjrirqrrz&cs3darknet_focus_s.ra4_e3600_r256_in1k)rnrnrn)r @rs)rnr3r4rjrqrrzcs3darknet_focus_m.c2ns_in1kzyhttps://github.com/rwightman/pytorch-image-models/releases/download/v0.1-tpu-weights/cs3darknet_focus_m_c2ns-e23bed41.pthzcs3darknet_focus_l.c2ns_in1kzyhttps://github.com/rwightman/pytorch-image-models/releases/download/v0.1-tpu-weights/cs3darknet_focus_l_c2ns-65ef8888.pthzuhttps://github.com/rwightman/pytorch-image-models/releases/download/v0.1-tpu-weights/cs3sedarknet_l_c2ns-e8d1dc13.pthzuhttps://github.com/rwightman/pytorch-image-models/releases/download/v0.1-tpu-weights/cs3sedarknet_x_c2ns-b4d0abc0.pthzqhttps://github.com/rwightman/pytorch-image-models/releases/download/v0.1-tpu-weights/cs3edgenet_x_c2-2e1610a9.pthzvhttps://github.com/rwightman/pytorch-image-models/releases/download/v0.1-tpu-weights/cs3se_edgenet_x_c2ns-76f8e3ac.pth)zcs3darknet_focus_x.untrainedzcs3sedarknet_l.c2ns_in1kzcs3sedarknet_x.c2ns_in1kzcs3sedarknet_xdw.untrainedzcs3edgenet_x.c2_in1kzcs3se_edgenet_x.c2ns_in1kr"cKtdd|i|S)Nr<rb)r<rdrbrr+r+r,r<cKrt)Nr@rb)r@rurvr+r+r,r@rwcKrt)NrArb)rArurvr+r+r,rArwcKrt)NrBrb)rBrurvr+r+r,rBrwcKrt)NrCrb)rCrurvr+r+r,rCrwcKrt)NrDrb)rDrurvr+r+r,rDrwcKrt)NrErb)rErurvr+r+r,rErwcKrt)NrFrb)rFrurvr+r+r,rF rwcKrt)NrHrb)rHrurvr+r+r,rHrwcKrt)NrIrb)rIrurvr+r+r,rIrwcKrt)NrJrb)rJrurvr+r+r,rJrwcKrt)NrKrb)rKrurvr+r+r,rK!rwcKrt)NrLrb)rLrurvr+r+r,rL&rwcKrt)NrMrb)rMrurvr+r+r,rM+rwcKrt)NrNrb)rNrurvr+r+r,rN0rwcKrt)NrOrb)rOrurvr+r+r,rO5rwcKrt)NrPrb)rPrurvr+r+r,rP:rwcKrt)NrQrb)rQrurvr+r+r,rQ?rwcKrt)NrRrb)rRrurvr+r+r,rRDrwcKrt)NrTrb)rTrurvr+r+r,rTIrwcKrt)NrUrb)rUrurvr+r+r,rUNrwcKrt)NrWrb)rWrurvr+r+r,rWSrwcKrt)NrXrb)rXrurvr+r+r,rXXrw) rArAFr\FNNrAr]r+r+)r")Wr dataclassesrrr functoolsrtypingrrrr r rtorch.nnr timm.datar r timm.layersrrrrrr_builderr _manipulaterr _registryrr__all__rr6r7rSrsr.rtrrrrrrrrrrrrQr(r*r rrrr`rdrm default_cfgsr<r@rArBrCrDrErFrHrIrJrKrLrMrNrOrPrQrRrTrUrWrXr+r+r+r,s,     $  %3((MK8 6    5 n  $1>JWeq          "     $)-15 M