o پiF@sdZddlmZmZmZmZddlZddlmZddl m Z m Z ddl m Z mZmZmZmZmZmZddlmZddlmZdd lmZdd lmZmZd gZGd d d ejZGdddej Z!Gdddej Z"Gdd d ej Z#e$e$gdgdgddgddddde$gdgdgddgddddde$gdgdgddgd d!d!d"de$gdgdgddgd d!d!d"de$gdgdgddgd d!dd"de$gdgdgddgd d!dd"de$gdgdgddgdd!dd"de$gdgdgddgdd!dd"de$gdgdgddgd#d!dd"de$gdgdgddgdd!dd$dd% Z%e%d&e%d'<dKd(d)Z&dLd*d+Z'ee'dd,e'dd,e'dd,e'd-d.d/d0e'dd,e'd-d.d/d0e'd-d1d1d/d2d3d4d5d6e'dd,e'dd,e'dd,d7 Z(edKd8e#fd9d:Z)edKd8e#fd;d<Z*edKd8e#fd=d>Z+edKd8e#fd?d@Z,edKd8e#fdAdBZ-edKd8e#fdCd&Z.edKd8e#fdDdEZ/edKd8e#fdFdGZ0edKd8e#fdHdIZ1edKd8e#fdJd'Z2dS)Ma VoVNet (V1 & V2) Papers: * `An Energy and GPU-Computation Efficient Backbone Network` - https://arxiv.org/abs/1904.09730 * `CenterMask : Real-Time Anchor-Free Instance Segmentation` - https://arxiv.org/abs/1911.06667 Looked at https://github.com/youngwanLEE/vovnet-detectron2 & https://github.com/stigma0617/VoVNet.pytorch/blob/master/models_vovnet/vovnet.py for some reference, rewrote most of the code. Hacked together by / Copyright 2020 Ross Wightman )ListOptionalTupleUnionNIMAGENET_DEFAULT_MEANIMAGENET_DEFAULT_STD) ConvNormActSeparableConvNormActBatchNormAct2dClassifierHeadDropPath create_attncreate_norm_act_layer)build_model_with_cfg)feature_take_indices)checkpoint_seq)register_modelgenerate_default_cfgsVovNetcs<eZdZfddZdejdeejdejfddZZS)SequentialAppendListcstt|j|dSN)superr__init__)selfargs __class__F/home/ubuntu/.local/lib/python3.10/site-packages/timm/models/vovnet.pyrszSequentialAppendList.__init__x concat_listreturncCsNt|D]\}}|dkr|||q|||dqtj|dd}|S)Nrr)dim) enumerateappendtorchcat)rr!r"imodulerrr forward"s zSequentialAppendList.forward) __name__ __module__ __qualname__rr(Tensorrr, __classcell__rrrr rs (rcs4eZdZdddeejdffdd ZddZZS)OsaBlockFNc stt|||_||_t|| d} |} |jr,| |kr,|r Jt| |dfi| |_nd|_g} t|D] }|jrDt ||fi| }n t| |dfi| }|} | |q5t | |_ |||} t| |fi| |_ |rrt||nd|_| |_dS)N norm_layer act_layerr)rr2rresidual depthwisedictr conv_reductionranger r'rconv_mid conv_concatrattn drop_path)rin_chsmid_chsout_chslayer_per_blockr8r9r?r5r6r@ conv_kwargs next_in_chs mid_convsr*convrrr r.s*       zOsaBlock.__init__cCsn|g}|jdur ||}|||}||}|jdur"||}|jdur,||}|jr5||d}|S)Nr)r;r=r>r?r@r8)rr!outputrrr r,Zs         zOsaBlock.forward r-r.r/r nnReLUrr,r1rrrr r2,s,r2cs6eZdZddddeejdffdd ZddZZS) OsaStageTFeseNc stt|d|_|rtjdddd|_nd|_g} t|D]6}||dk}| dur8| |dkr8t| |}nd}| t |||||oE|dk||rJ| nd | | |d g7} |}qtj | |_ dS) NFr7T) kernel_sizestride ceil_moderrr3)r8r9r?r5r6r@) rrMrgrad_checkpointingrK MaxPool2dpoolr<r r2 Sequentialblocks)rrArBrCblock_per_stagerD downsampler8r9r?r5r6drop_path_ratesrXr* last_blockr@rrr rks2      zOsaStage.__init__cCsB|jdur ||}|jrtjst|j|}|S||}|Sr)rVrTr(jit is_scriptingrrXrr!rrr r,s    zOsaStage.forwardrJrrrr rMis -rMcs,eZdZddddeejddffdd Zejj d*d d Z ejj d+d d Z ejj dej fddZ d,deefddZ     d-dejdeeeeefdededededeeejeejeejfffddZ   d.deeeefded efd!d"Zd#d$Zd*d%efd&d'Zd(d)ZZS)/rr7avg rSc  stt|||_||_|dksJt|fi| }|dd} |d} |d} |d}|d}|d}t||d }| d }|d rEtnt}t j t|| d d fdd i||| d | dd fddi||| d| d d fd|i|g|_ t| dd d| dkrdnd dg|_ | }t t d | t||}| dd|dd}td|d|d |dd|}g}tdD]D}| d kp|d k}|t||| ||||||f|||d|g7}|||_||rd nd9}|j t|j|d|dg7_ qt j ||_|j|_t|j|||d|_|D]&\}}t|t jr2t jj|jdddqt|t jr@t j|j qdS)a Args: cfg (dict): Model architecture configuration in_chans (int): Number of input channels (default: 3) num_classes (int): Number of classifier classes (default: 1000) global_pool (str): Global pooling type (default: 'avg') output_stride (int): Output stride of network, one of (8, 16, 32) (default: 32) norm_layer (Union[str, nn.Module]): normalization layer act_layer (Union[str, nn.Module]): activation layer drop_rate (float): Dropout rate (default: 0.) drop_path_rate (float): Stochastic depth drop-path rate (default: 0.) kwargs (dict): Extra kwargs overlayed onto cfg rb stem_stridestem_chsstage_conv_chs stage_out_chsrYrDr4rOr9rr7rQrzstem.)num_chs reductionr+r$Nr8r?)r8r9r?)rZr[zstages.) pool_type drop_ratefan_outrelu)mode nonlinearityr)!rrr num_classesrkr:getr r rKrWstem feature_infor(splitlinspacesumr<rM num_featuresstageshead_hidden_sizer head named_modules isinstanceConv2dinitkaiming_normal_weightLinearzeros_bias)rcfgin_chansrp global_pool output_strider5r6rkdrop_path_ratekwargsrcrerfrgrYrDrElast_stem_stride conv_typecurrent_stride stage_dpr in_ch_list stage_argsrxr*rZnmrrr rsl    "  $ zVovNet.__init__FcCstd|rddSddS)Nz^stemz^stages\.(\d+)z^stages\.(\d+).blocks\.(\d+))rrrX)r:)rcoarserrr group_matchers zVovNet.group_matcherTcCs|jD]}||_qdSr)rxrT)renablesrrr set_grad_checkpointings zVovNet.set_grad_checkpointingr#cCs|jjSr)rzfc)rrrr get_classifierszVovNet.get_classifierNrcCs||_|j||dSr)rprzreset)rrprrrr reset_classifierszVovNet.reset_classifierNCHWr!indicesnorm stop_early output_fmtintermediates_onlyc Cs|dvsJdg}td|\}} d} |jdd|}| |vr%|||jd|}tjs3|s7|j} n|jd| } t| ddD]\} } | |}| |vrU||qD|rZ|S||fS) a Forward features that returns intermediates. Args: x: Input image tensor indices: Take last n blocks if int, all if None, select matching indices if sequence norm: Apply norm layer to compatible intermediates stop_early: Stop iterating over blocks when last desired intermediate hit output_fmt: Shape of intermediate feature outputs intermediates_only: Only return intermediate features Returns: )rzOutput shape must be NCHW.rNr$r)start)rrrr'r(r]r^rxr&) rr!rrrrr intermediates take_indices max_indexfeat_idxrxstagerrr forward_intermediates s&  zVovNet.forward_intermediatesr prune_norm prune_headcCs2td|\}}|jd||_|r|dd|S)z@ Prune layers not required for specified intermediates. rNrr3)rrxr)rrrrrrrrr prune_intermediate_layers;s  z VovNet.prune_intermediate_layerscCs||}||Sr)rrrxr_rrr forward_featuresIs  zVovNet.forward_features pre_logitscCs|r |j||dS||S)N)r)rz)rr!rrrr forward_headMszVovNet.forward_headcCs||}||}|Sr)rrr_rrr r,Ps  zVovNet.forwardF)Tr)NFFrF)rFT)r-r.r/r rKrLrr(r]ignorerrModulerrstrrr0rintrboolrrrrrr,r1rrrr rsdT    1 )@r)r)iir)rrrOrOFr3)rerfrgrDrYr8r9r?)rrrdr7)rrr)rP`p)rrirr7)rrrrTrN)rr7 r7eca) vovnet39a vovnet57aese_vovnet19b_slim_dwese_vovnet19b_dwese_vovnet19b_slim ese_vovnet19b ese_vovnet39b ese_vovnet57b ese_vovnet99b eca_vovnet39brese_vovnet39b_evoscKs$tt||ft|tddd|S)NT)flatten_sequential) model_cfg feature_cfg)rr model_cfgsr:)variant pretrainedrrrr _create_vovnetsrc Ks|dddddttddd |S) Nr`)r7rr)rg?bicubicz stem.0.convzhead.fc) urlrp input_size pool_sizecrop_pct interpolationmeanstd first_conv classifierr)rrrrr _cfgsr)rztimm/)r7 rgffffff?) hf_hub_idtest_input_size test_crop_pct)?rr)r7rr)r)r7@rg?)rrrrrrrr) zvovnet39a.untrainedzvovnet57a.untrainedzese_vovnet19b_slim_dw.untrainedzese_vovnet19b_dw.ra_in1kzese_vovnet19b_slim.untrainedzese_vovnet39b.ra_in1kz!ese_vovnet57b.ra4_e3600_r256_in1kzese_vovnet99b.untrainedzeca_vovnet39b.untrainedzese_vovnet39b_evos.untrainedr#cKtdd|i|S)Nrr)rrrrrrr rrcKr)Nrr)rrrrrr rrrcKr)Nrr)rrrrrr rrrcKr)Nrr)rrrrrr rrrcKr)Nrr)rrrrrr rrrcKr)Nrr)rrrrrr r rcKr)Nrr)rrrrrr rrrcKr)Nrr)rrrrrr rrrcKr)Nrr)rrrrrr rrrcKsdd}td||d|S)Nc[std|fddi|S)N evonorms0r]F)r)rwnkwargsrrr norm_act_fn!sz'ese_vovnet39b_evos..norm_act_fnr)rr5)rr)rrrrrr rsr)r3)3__doc__typingrrrrr(torch.nnrK timm.datarr timm.layersr r r r r rr_builderr _featuresr _manipulater _registryrr__all__rWrrr2rMrr:rrr default_cfgsrrrrrrrrrrrrrr sH  $   =97          j