U (d{: @s8ddlmZddlmZmZmZmZmZddlZddlm Z m Z ddl m Z ddlmZmZddlmZdd lmZdd lmZd d lmZmZd d lmZd dlmZmZdddddddddg ZGddde j Z!Gddde j"Z#GdddZ$Gddde j"Z%ee$e&eee'ee%dd d!Z(d"ed#d$d%Z)Gd&ddeZ*Gd'ddeZ+Gd(ddeZ,Gd)ddeZ-ed*e*j.fd+dd,d-ee*e'ee%d.d/dZ/ed*e+j.fd+dd,d-ee+e'ee%d.d0dZ0ed*e,j.fd+dd,d-ee,e'ee%d.d1dZ1ed*e-j.fd+dd,d-ee-e'ee%d.d2dZ2dS)3)partial)AnyCallableListOptionalSequenceN)nnTensor) functional)Conv2dNormActivationPermute)StochasticDepth)ImageClassification)_log_api_usage_once) WeightsEnumWeights)_IMAGENET_CATEGORIES)handle_legacy_interface_ovewrite_named_paramConvNeXtConvNeXt_Tiny_WeightsConvNeXt_Small_WeightsConvNeXt_Base_WeightsConvNeXt_Large_Weights convnext_tinyconvnext_small convnext_baseconvnext_largec@seZdZeedddZdS) LayerNorm2dxreturncCs>|dddd}t||j|j|j|j}|dddd}|S)Nrr r)ZpermuteFZ layer_normZnormalized_shapeweightbiasepsselfr"r+?/tmp/pip-unpacked-wheel-vx7f76es/torchvision/models/convnext.pyforwardszLayerNorm2d.forwardN)__name__ __module__ __qualname__r r-r+r+r+r,r sr csHeZdZd eeeedejfddfdd Ze e dddZ Z S) CNBlockN.) layer_scalestochastic_depth_prob norm_layerr#c st|dkr ttjdd}ttj||dd|ddtdddd g||tj|d |dd t tjd ||dd tddd dg|_ t t |d d ||_t|d |_dS) Nư>r(r$T) kernel_sizepaddinggroupsr'rr r)Z in_featuresZ out_featuresr'row)super__init__rr LayerNorm SequentialConv2dr LinearZGELUblock ParametertorchZonesr2rstochastic_depth)r*Zdimr2r3r4 __class__r+r,r>'s  zCNBlock.__init__)inputr#cCs&|j||}||}||7}|SN)r2rCrF)r*rIresultr+r+r,r->s zCNBlock.forward)N) r.r/r0floatrrrModuler>r r- __classcell__r+r+rGr,r1&sr1c@s2eZdZeeeeddddZedddZdS) CNBlockConfigN)input_channels out_channels num_layersr#cCs||_||_||_dSrJ)rPrQrR)r*rPrQrRr+r+r,r>GszCNBlockConfig.__init__)r#cCs:|jjd}|d7}|d7}|d7}|d7}|jf|jS)N(zinput_channels={input_channels}z, out_channels={out_channels}z, num_layers={num_layers}))rHr.format__dict__)r*sr+r+r,__repr__Qs  zCNBlockConfig.__repr__)r.r/r0intrr>strrXr+r+r+r,rOEs  rOc sreZdZdeeeeeeede j feede j fe ddfdd Z e e d d d Ze e d d d ZZS)rr5N.) block_settingr3r2 num_classesrCr4kwargsr#c stt||s tdn$t|trjsz%ConvNeXt.__init__..z/The block_setting should be List[CNBlockConfig]r5r6rr$r;T)r8strider9r4Zactivation_layerr'css|] }|jVqdSrJ)rR)racnfr+r+r, sz$ConvNeXt.__init__..g?rr )r8rcg{Gz?)Zstd)!r=r>r ValueErrorr`rall TypeErrorr1rr rPappendr sumrangerRrr@rQrAfeaturesZAdaptiveAvgPool2davgpoolZFlattenrB classifiermodulesinitZ trunc_normal_r&r'Zzeros_)r*r]r3r2r^rCr4r_ZlayersZfirstconv_output_channelsZtotal_stage_blocksZstage_block_idrdZstage_Zsd_probZ lastblockZlastconv_output_channelsmrGr+r,r>[sn            zConvNeXt.__init__r!cCs"||}||}||}|SrJ)rmrnror)r+r+r, _forward_impls   zConvNeXt._forward_implcCs ||SrJ)rtr)r+r+r,r-szConvNeXt.forward)r[r5r\NN)r.r/r0rrOrLrYrrrrMrr>r rtr-rNr+r+rGr,rZs"N)r]r3weightsprogressr_r#cKsP|dk rt|dt|jdt|fd|i|}|dk rL||j|d|S)Nr^ categoriesr3)rv)rlenmetarZload_state_dictZget_state_dict)r]r3rurvr_modelr+r+r, _convnexts r{) r|zNhttps://github.com/pytorch/vision/tree/main/references/classification#convnexta These weights improve upon the results of the original paper by using a modified version of TorchVision's `new training recipe `_. )Zmin_sizerwZrecipeZ_docsc @s<eZdZedeedddeddddd id d ZeZd S) rz>https://download.pytorch.org/models/convnext_tiny-983f1562.pthZ crop_sizeZ resize_sizeiH< ImageNet-1KgzGT@gMbX X@zacc@1zacc@5Z num_paramsZ_metricsurlZ transformsryN r.r/r0rrr _COMMON_META IMAGENET1K_V1DEFAULTr+r+r+r,rs c @s<eZdZedeedddeddddd id d ZeZd S) rz?https://download.pytorch.org/models/convnext_small-0c510722.pthr}riHZrgClT@g)X@rrrNrr+r+r+r,rs c @s<eZdZedeedddeddddd id d ZeZd S) rz>https://download.pytorch.org/models/convnext_base-6075fbad.pthr}rihGrgU@gHz7X@rrrNrr+r+r+r,rs c @s<eZdZedeedddeddddd id d ZeZd S) rz?https://download.pytorch.org/models/convnext_large-ea097f82.pthr}rri rg"~U@gX9v>X@rrrNrr+r+r+r,rs Z pretrained)ruT)rurv)rurvr_r#cKsTt|}tdddtdddtdddtdddg}|dd }t||||f|S) aConvNeXt Tiny model architecture from the `A ConvNet for the 2020s `_ paper. Args: weights (:class:`~torchvision.models.convnext.ConvNeXt_Tiny_Weights`, optional): The pretrained weights to use. See :class:`~torchvision.models.convnext.ConvNeXt_Tiny_Weights` below for more details and possible values. By default, no pre-trained weights are used. progress (bool, optional): If True, displays a progress bar of the download to stderr. Default is True. **kwargs: parameters passed to the ``torchvision.models.convnext.ConvNext`` base class. Please refer to the `source code `_ for more details about this class. .. autoclass:: torchvision.models.ConvNeXt_Tiny_Weights :members: `r$ Nr3g?)rverifyrOpopr{rurvr_r]r3r+r+r,rs      cKsTt|}tdddtdddtdddtdddg}|dd }t||||f|S) aConvNeXt Small model architecture from the `A ConvNet for the 2020s `_ paper. Args: weights (:class:`~torchvision.models.convnext.ConvNeXt_Small_Weights`, optional): The pretrained weights to use. See :class:`~torchvision.models.convnext.ConvNeXt_Small_Weights` below for more details and possible values. By default, no pre-trained weights are used. progress (bool, optional): If True, displays a progress bar of the download to stderr. Default is True. **kwargs: parameters passed to the ``torchvision.models.convnext.ConvNext`` base class. Please refer to the `source code `_ for more details about this class. .. autoclass:: torchvision.models.ConvNeXt_Small_Weights :members: rrr$rrNr3g?)rrrOrr{rr+r+r,r7s      cKsTt|}tdddtdddtdddtdddg}|dd }t||||f|S) aConvNeXt Base model architecture from the `A ConvNet for the 2020s `_ paper. Args: weights (:class:`~torchvision.models.convnext.ConvNeXt_Base_Weights`, optional): The pretrained weights to use. See :class:`~torchvision.models.convnext.ConvNeXt_Base_Weights` below for more details and possible values. By default, no pre-trained weights are used. progress (bool, optional): If True, displays a progress bar of the download to stderr. Default is True. **kwargs: parameters passed to the ``torchvision.models.convnext.ConvNext`` base class. Please refer to the `source code `_ for more details about this class. .. autoclass:: torchvision.models.ConvNeXt_Base_Weights :members: r$iirNr3?)rrrOrr{rr+r+r,rWs      cKsTt|}tdddtdddtdddtdddg}|dd }t||||f|S) aConvNeXt Large model architecture from the `A ConvNet for the 2020s `_ paper. Args: weights (:class:`~torchvision.models.convnext.ConvNeXt_Large_Weights`, optional): The pretrained weights to use. See :class:`~torchvision.models.convnext.ConvNeXt_Large_Weights` below for more details and possible values. By default, no pre-trained weights are used. progress (bool, optional): If True, displays a progress bar of the download to stderr. Default is True. **kwargs: parameters passed to the ``torchvision.models.convnext.ConvNext`` base class. Please refer to the `source code `_ for more details about this class. .. autoclass:: torchvision.models.ConvNeXt_Large_Weights :members: rrr$rirNr3r)rrrOrr{rr+r+r,rus      )3 functoolsrtypingrrrrrrErr Ztorch.nnr r%Zops.miscr r Zops.stochastic_depthrZtransforms._presetsrutilsrZ_apirrZ_metar_utilsrr__all__r?r rMr1rOrrLboolr{rrrrrrrrrrr+r+r+r,s~       Z  ""