U ʹj: @sXdZddlmZddlZddlmZddlmZmZddl m Z m Z m Z m Z mZddlmZddlmZdd lmZmZGd d d ejZGd d d ejZGdddejZGdddejZGdddejZGdddejZd+ddZeeddeddedddeddddd d!Zd,d#d$Zed-d%d&Z ed.d'd(Z!ed/d)d*Z"dS)0z InceptionNeXt paper: https://arxiv.org/abs/2303.16900 Original implementation & weights from: https://github.com/sail-sg/inceptionnext )partialNIMAGENET_DEFAULT_MEANIMAGENET_DEFAULT_STD) trunc_normal_DropPath to_2tuple get_paddingSelectAdaptivePool2d)build_model_with_cfg)checkpoint_seq)register_modelgenerate_default_cfgscs*eZdZdZd fdd Zdd ZZS) InceptionDWConv2dz% Inception depthwise convolution  ?r c stt||}t||d}t||d}tj||||||d|_tj||d|fd|fd|f|d|_tj|||df|df|df|d|_|d||||f|_ dS)Ndilation)paddingrgroupsr rr) super__init__intr nnConv2d dwconv_hwdwconv_wdwconv_h split_indexes) selfin_chsZsquare_kernel_sizeZband_kernel_sizeZ branch_ratiorgcZsquare_paddingZ band_padding __class__Y/var/www/html/bwave/python/venv/lib/python3.8/site-packages/timm/models/inception_next.pyrs:    zInceptionDWConv2d.__init__cCsBtj||jdd\}}}}tj|||||||fddS)Nr )dim)torchsplitr catrrr)r!xZx_idZx_hwZx_wZx_hr&r&r'forward.szInceptionDWConv2d.forward)rrrr )__name__ __module__ __qualname____doc__rr- __classcell__r&r&r$r'rsrcs8eZdZdZddejdddffdd ZddZZS) ConvMlpz MLP using 1x1 convs that keeps spatial dims copied from timm: https://github.com/huggingface/pytorch-image-models/blob/v0.6.11/timm/models/layers/mlp.py NTcst|p|}|p|}t|}tj||d|dd|_|rF||nt|_||_t ||_ tj||d|dd|_ dS)Nr r) kernel_sizebias) rrrrrfc1IdentitynormactDropoutdropfc2)r! in_featureshidden_featuresZ out_features act_layer norm_layerr6r<r$r&r'r>s  zConvMlp.__init__cCs6||}||}||}||}||}|SN)r7r9r:r<r=r!r,r&r&r'r-Ss      zConvMlp.forward) r.r/r0r1rZReLUrr-r2r&r&r$r'r39sr3csDeZdZdZdddejeejddddffd d Zd d Z Z S) MlpClassifierHeadz MLP classification head avgrư>)epsr4Tc sxtt|dd|_||j} t|| } tj| | |d|_||_ || |_ tj| ||d|_ t ||_ dS)NT pool_typeflatten)r6)rrr global_poolZ feat_multrrLinearr7r:r9r=r;r<) r!r( num_classesrJ mlp_ratior@rAr<r6r>r?r$r&r'r`s   zMlpClassifierHead.__init__cCs@||}||}||}||}||}||}|SrB)rLr7r:r9r<r=rCr&r&r'r-us      zMlpClassifierHead.forward) r.r/r0r1rGELUrZ LayerNormrr-r2r&r&r$r'rD\s rDcs>eZdZdZdeejedejddffdd Z dd Z Z S) MetaNeXtBlockz MetaNeXtBlock Block Args: dim (int): Number of input channels. drop_path (float): Stochastic depth rate. Default: 0.0 ls_init_value (float): Init value for Layer Scale. Default: 1e-6. r rGr4c svt|||d|_|||_||t|||d|_|rRt|t |nd|_ | dkrht | nt |_ dS)Nr)r@r4)rr token_mixerr9rmlpr Parameterr)Zonesgammarr8 drop_path) r!r(rrSrAZ mlp_layerrOr@ ls_init_valuerWr$r&r'rs  zMetaNeXtBlock.__init__cCsV|}||}||}||}|jdk rD||jdddd}|||}|S)Nr )rSr9rTrVmulZreshaperW)r!r,Zshortcutr&r&r'r-s    zMetaNeXtBlock.forward) r.r/r0r1rr BatchNorm2dr3rPrr-r2r&r&r$r'rQs rQc s:eZdZdddddeejddf fdd Zdd ZZS) MetaNeXtStage)r r N?rRc  std|_|dks(|d|dkrPt| |tj||d||dd|_n t|_|pfdg|}g} t|D]*} | t ||d|| ||| | | dqttj| |_ dS)NFr rr])r5striderr4)r(rrWrXrSr@rArO) rrgrad_checkpointingr Sequentialr downsampler8rangeappendrQblocks)r!r"out_chsr_depthrdrop_path_ratesrXrSr@rArOZ stage_blocksir$r&r'rs8    zMetaNeXtStage.__init__cCs6||}|jr(tjs(t|j|}n ||}|SrB)rbr`r)jitZ is_scriptingr rerCr&r&r'r-s   zMetaNeXtStage.forward) r.r/r0rrrPrr-r2r&r&r$r'r\s-r\cseZdZdZddddddeejejded d d ffd d Z d dZ e j j d$ddZe j j ddZddefddZe j j d%ddZe j j ddZddZd&edd d!Zd"d#ZZS)'MetaNeXta MetaNeXt A PyTorch impl of : `InceptionNeXt: When Inception Meets ConvNeXt` - https://arxiv.org/abs/2303.16900 Args: in_chans (int): Number of input image channels. Default: 3 num_classes (int): Number of classes for classification head. Default: 1000 depths (tuple(int)): Number of blocks at each stage. Default: (3, 3, 9, 3) dims (tuple(int)): Feature dimension at each stage. Default: (96, 192, 384, 768) token_mixers: Token mixer function. Default: nn.Identity norm_layer: Normalization layer. Default: nn.BatchNorm2d act_layer: Activation function for MLP. Default: nn.GELU mlp_ratios (int or tuple(int)): MLP ratios. Default: (4, 4, 4, 3) head_fn: classifier head drop_rate (float): Head dropout rate drop_path_rate (float): Stochastic depth rate. Default: 0. ls_init_value (float): Init value for Layer Scale. Default: 1e-6. rrErF rr r`i)rRrRrRrr4rGcstt|}t|ttfs*|g|}t| ttfsB| g|} ||_||_| |_g|_ t t j ||dddd||d|_ ddtd| t||D}|d}d}d}t |_t|D]}|dks|dkrdnd}||kr|dkr||9}d}||9}|dkrdnd}||}|jt|||dkr:|nd||f|||||| |||| |d |}|j t||d |d g7_ q||_|jdkrt| tr|jstd | |j||j| d |_n$|jrt|jdd|_n t |_||jdS)NrrR)r5r_cSsg|] }|qSr&)tolist).0r,r&r&r' sz%MetaNeXt.__init__..r r])r r]) r_rrgrhrXr@rSrArOzstages.)Znum_chsZ reductionmodule4Cannot disable global pooling with MLP head present.rJr<TrI) rrlen isinstancelisttuplerNrL drop_rateZ feature_inforrarstemr)Zlinspacesumr*stagesrcrdr\dict num_features issubclassrDAssertionErrorheadr r8apply _init_weights)r!Zin_chansrNrLZ output_stridedepthsdims token_mixersrAr@Z mlp_ratioshead_fnr}Zdrop_path_raterXZ num_stageZdp_ratesZprev_chsZ curr_striderrir_Zfirst_dilationrfr$r&r'rsf    "   "   zMetaNeXt.__init__cCs>t|tjtjfr:t|jdd|jdk r:tj|jddS)Ng{Gz?)stdr) rzrrrMrweightr6initZ constant_)r!mr&r&r'r;s zMetaNeXt._init_weightsFcCstd|r dnddgdS)Nz^stemz^stages\.(\d+))z^stages\.(\d+)\.downsample)r)z^stages\.(\d+)\.blocks\.(\d+)N)r~re)r)r!Zcoarser&r&r' group_matcherAszMetaNeXt.group_matchercCs|jjSrB)rr=r!r&r&r'get_classifierKszMetaNeXt.get_classifierrNcCsn|dk r||_|dkrHt|tr.|js.td||j||j|jd|_n"|jr`t|jdd|_n t |_dS)NrrwrxTrI) rLrrDrrr}rr rr8)r!rNrLrr&r&r'reset_classifierOs zMetaNeXt.reset_classifierTcCs|jD] }||_qdSrB)rr`)r!enablesr&r&r'set_grad_checkpointing\s zMetaNeXt.set_grad_checkpointingcCstSrB)setrr&r&r'no_weight_decayaszMetaNeXt.no_weight_decaycCs||}||}|SrB)r~rrCr&r&r'forward_featureses  zMetaNeXt.forward_features) pre_logitscCs*|r t|jdr|j|}|S||S)NrL)hasattrrrL)r!r,rr&r&r' forward_headjs   zMetaNeXt.forward_headcCs||}||}|SrB)rrrCr&r&r'r-qs  zMetaNeXt.forward)F)T)F)r.r/r0r1rrr[rPrDrrr)rjignorerrrrrrboolrr-r2r&r&r$r'rks:K     rkc Ks|dddddttddd |S) NrE)rr)rg?Zbicubiczstem.0zhead.fc2) urlrN input_size pool_sizecrop_pct interpolationZmeanrZ first_conv classifierr)rkwargsr&r&r'_cfgwsrztimm/) hf_hub_idgffffff?)rr)rrrrr) rr^)rrrr)zinception_next_tiny.sail_in1kzinception_next_small.sail_in1kzinception_next_base.sail_in1kz!inception_next_base.sail_in1k_384FcKs$tt||fdtdddi|}|S)NZ feature_cfg)rr r]rT)Z out_indicesZflatten_sequential)r rkr)variant pretrainedrmodelr&r&r'_create_inception_nexts rcKs(tddtd}tdd|it|f|S)Nrmrorrrinception_next_tinyr)rrrrrrZ model_argsr&r&r'rs rcKs(tddtd}tdd|it|f|S)Nrrrrorinception_next_smallr)rrrr&r&r'rs rcKs(tddtd}tdd|it|f|S)Nr)iirinception_next_baser)rrrr&r&r'rs r)r)F)F)F)F)#r1 functoolsrr)Ztorch.nnrZ timm.datarrZ timm.layersrrrr r Z_builderr Z _manipulater _registryrrModulerr3rDrQr\rkrZ default_cfgsrrrrr&r&r&r'sR    '##'7