U (d@s UddlZddlZddlZddlmZddlmZddlmZm Z m Z m Z m Z m Z mZmZddlZddlmZmZddlmZddlmZmZdd lmZmZdd lmZd d lmZmZd d l m!Z!d dl"m#Z#m$Z$m%Z%dddddddddddddddddd d!d"d#d$d%gZ&eGd&d'd'Z'Gd(d)d)e'Z(Gd*d+d+e'Z)Gd,d-d-ej*Z+Gd.d/d/ej*Z,Gd0ddej*Z-e ee(e)fe.e e/e ee0ee-d1d2d3Z1e2eee ee(e)fe e/fd4d5d6Z3d7e!iZ4e e2efe5d8<e4d9d:d;Z6e4dddeZ8Gd?ddeZ9Gd@ddeZ:GdAddeZ;GdBddeZGdEddeZ?GdFddeZ@GdGddeZAGdHddeZBe#dIe8jCfdJddKdLe e8e0ee-dMdNdZDe#dIe9jCfdJddKdLe e9e0ee-dMdOdZEe#dIe:jCfdJddKdLe e:e0ee-dMdPdZFe#dIe;jCfdJddKdLe e;e0ee-dMdQdZGe#dIejCfdJddKdLe e>e0ee-dMdTd!ZJe#dIe?jCfdJddKdLe e?e0ee-dMdUd"ZKe#dIe@jCfdJddKdLe e@e0ee-dMdVd#ZLe#dIeAjCfdJddKdLe eAe0ee-dMdWd$ZMe#dIeBjCfdJddKdLe eBe0ee-dMdXd%ZNd dYl"mOZOeOe8jCjPe9jCjPe:jCjPe;jCjPejCjPe?jCjPdZZQdS)[N) dataclass)partial)AnyCallableDictOptionalListSequenceTupleUnion)nnTensor)StochasticDepth)Conv2dNormActivationSqueezeExcitation)ImageClassificationInterpolationMode)_log_api_usage_once) WeightsEnumWeights)_IMAGENET_CATEGORIES)handle_legacy_interface_ovewrite_named_param_make_divisible EfficientNetEfficientNet_B0_WeightsEfficientNet_B1_WeightsEfficientNet_B2_WeightsEfficientNet_B3_WeightsEfficientNet_B4_WeightsEfficientNet_B5_WeightsEfficientNet_B6_WeightsEfficientNet_B7_WeightsEfficientNet_V2_S_WeightsEfficientNet_V2_M_WeightsEfficientNet_V2_L_Weightsefficientnet_b0efficientnet_b1efficientnet_b2efficientnet_b3efficientnet_b4efficientnet_b5efficientnet_b6efficientnet_b7efficientnet_v2_sefficientnet_v2_mefficientnet_v2_lc@sneZdZUeed<eed<eed<eed<eed<eed<edejfed<e d eee eed d d Z d S) _MBConvConfig expand_ratiokernelstrideinput_channels out_channels num_layers.blockN)channels width_mult min_valuereturncCst||d|S)N)r)r;r<r=r@C/tmp/pip-unpacked-wheel-vx7f76es/torchvision/models/efficientnet.pyadjust_channels9sz_MBConvConfig.adjust_channels)N) __name__ __module__ __qualname__float__annotations__intrr Module staticmethodrrBr@r@r@rAr3/s r3c sXeZdZd eeeeeeeeeedejfdd fdd Z e eeddd Z Z S) MBConvConfig?N.) r4r5r6r7r8r9r< depth_multr:r>c sL|||}|||}|||}| dkr0t} t||||||| dSN)rB adjust_depthMBConvsuper__init__) selfr4r5r6r7r8r9r<rMr: __class__r@rArR@s   zMBConvConfig.__init__r9rMcCstt||SrN)rHmathceilrVr@r@rArOSszMBConvConfig.adjust_depth)rLrLN) rCrDrErFrHrrr rIrRrJrO __classcell__r@r@rTrArK>s" rKc s@eZdZdeeeeeeeedejfddfdd Z Z S)FusedMBConvConfigN.)r4r5r6r7r8r9r:r>c s(|dkr t}t|||||||dSrN) FusedMBConvrQrR)rSr4r5r6r7r8r9r:rTr@rArRZs zFusedMBConvConfig.__init__)N) rCrDrErFrHrrr rIrRrYr@r@rTrArZXs rZcsReZdZefeeedejfedejfddfdd Z e e dddZ Z S) rP.N)cnfstochastic_depth_prob norm_layerse_layerr>c s td|jkr dks*ntd|jdko>|j|jk|_g}tj}| |j|j }||jkr| t |j|d||d| t |||j |j|||dtd|jd}| |||ttjddd | t ||jd|ddtj||_t|d |_|j|_dS) Nrrillegal stride value kernel_sizer^activation_layer)rbr6groupsr^rcT)inplace)Z activationrow)rQrRr6 ValueErrorr7r8use_res_connectr SiLUrBr4appendrr5maxr Sequentialr:rstochastic_depth) rSr\r]r^r_layersrcexpanded_channelsZsqueeze_channelsrTr@rArRjsT      zMBConv.__init__inputr>cCs&||}|jr"||}||7}|SrNr:rirnrSrrresultr@r@rAforwards   zMBConv.forward) rCrDrErrKrFrr rIrRr rvrYr@r@rTrArPis  :rPcsBeZdZeeedejfddfdd Ze e dddZ Z S) r[.N)r\r]r^r>c std|jkr dks*ntd|jdko>|j|jk|_g}tj}| |j|j }||jkr| t |j||j |j||d| t ||jd|ddn"| t |j|j|j |j||dtj||_t|d|_|j|_dS)Nrrr`rbr6r^rcrarg)rQrRr6rhr7r8rir rjrBr4rkrr5rmr:rrn)rSr\r]r^rorcrprTr@rArRsP     zFusedMBConv.__init__rqcCs&||}|jr"||}||7}|SrNrsrtr@r@rArvs   zFusedMBConv.forward) rCrDrErZrFrr rIrRr rvrYr@r@rTrAr[s  4r[c speZdZd eeeefeeee e de j fe ee ddfdd Zeedd d Zeedd d ZZS)r皙?N.)inverted_residual_settingdropoutr] num_classesr^ last_channelkwargsr>c sztt||s tdn$t|trsz)EfficientNet.__init__..z:The inverted_residual_setting should be List[MBConvConfig]r:zThe parameter 'block' is deprecated since 0.13 and will be removed 0.15. Please pass this information on 'MBConvConfig.block' instead.Nrrrwcss|] }|jVqdSrN)r9)rr\r@r@rA "sz(EfficientNet.__init__..rreraT)prfZfan_out)moderL)-rQrRrrhrr all TypeErrorwarningswarnrKr:r BatchNorm2dr7rkrrjsumranger9copyr8r6rFrmfeaturesZAdaptiveAvgPool2davgpoolZDropoutZLinear classifiermodulesZConv2dinitZkaiming_normal_ZweightZbiasZzeros_Z GroupNormZones_rWsqrtZ out_featuresZuniform_)rSrzr{r]r|r^r}r~rroZfirstconv_output_channelsZtotal_stage_blocksZstage_block_idr\Zstage_Z block_cnfZsd_probZlastconv_input_channelsZlastconv_output_channelsmZ init_rangerTr@rArRs               zEfficientNet.__init__)xr>cCs.||}||}t|d}||}|S)Nr)rrtorchflattenrrSrr@r@rA _forward_implXs     zEfficientNet._forward_implcCs ||SrN)rrr@r@rArvbszEfficientNet.forward)rxryNN)rCrDrEr r rKrZrFrHrrr rIrrRr rrvrYr@r@rTrArs n )rzr{r}weightsprogressr~r>cKsR|dk rt|dt|jdt||fd|i|}|dk rN||j|d|S)Nr| categoriesr})r)rlenmetarZload_state_dictZget_state_dict)rzr{r}rrr~modelr@r@rA _efficientnetfs r)archr~r>c Ks:|drtt|d|dd}|dddddd|d dd dd d |d d d d d d |d dd d dd|d d dddd|d d d ddd|d dddddg}d}n|drtdddd d d tddd d ddtddd dddtddd ddd td dddddtd dd dddg}d}n|drtdddd d dtddd d dd tddd ddd tddd dddtd dddddtd dd dd d!td ddd d"d g}d}n|d#r$tddddddtddd dddtddd dd$dtddd d$dd%td dddd&d'td dd d&d(d)td ddd(d*dg}d}ntd+|||fS),NZefficientnet_br<rMr<rMrr r(Ppre@r00@ ir1i0ir2` izUnsupported model type ) startswithrrKpoprZrh)rr~Z bneck_confrzr}r@r@rA_efficientnet_confysT       rr _COMMON_META)rrzUhttps://github.com/pytorch/vision/tree/main/references/classification#efficientnet-v1)Zmin_sizerecipe)!rzUhttps://github.com/pytorch/vision/tree/main/references/classification#efficientnet-v2c @sBeZdZedeeddejdeddddd id d d Z e Z d S)rzJhttps://download.pytorch.org/models/efficientnet_b0_rwightman-3dd342df.pthrr crop_size resize_size interpolationidP ImageNet-1Kg?5^IlS@g5^I bW@zacc@1zacc@51These weights are ported from the original paper. num_params_metrics_docsurlZ transformsrN rCrDrErrrrBICUBIC_COMMON_META_V1 IMAGENET1K_V1DEFAULTr@r@r@rArs&c @sveZdZedeeddejdeddddd id d d Z ed eeddej dedddddd iddd Z e Z dS)rzJhttps://download.pytorch.org/models/efficientnet_b1_rwightman-533bc792.pthrrivrg +S@gClW@rrrrz@https://download.pytorch.org/models/efficientnet_b1-c27df63c.pthzOhttps://github.com/pytorch/vision/issues/3995#new-recipe-with-lr-wd-crop-tuninggʡS@gƻW@$ These weights improve upon the results of the original paper by using a modified version of TorchVision's `new training recipe `_. )rrrrN) rCrDrErrrrrrrBILINEARZ IMAGENET1K_V2rr@r@r@rArsLc @sBeZdZedeeddejdedddddid d d Z e Z d S) rzJhttps://download.pytorch.org/models/efficientnet_b2_rwightman-bcdf34b7.pthi rirgx&T@gp= W@rrrrNrr@r@r@rArs&c @sBeZdZedeeddejdeddddd id d d Z e Z d S)r zJhttps://download.pytorch.org/models/efficientnet_b3_rwightman-cf984f9c.pthi,rrirgnT@g~jtX@rrrrNrr@r@r@rAr s&c @sBeZdZedeeddejdeddddd id d d Z e Z d S)r!zJhttps://download.pytorch.org/models/efficientnet_b4_rwightman-7eb33cd5.pthi|rri0!'rgjtT@gt&X@rrrrNrr@r@r@rAr!1s&c @sBeZdZedeeddejdedddddid d d Z e Z d S) r"zJhttps://download.pytorch.org/models/efficientnet_b5_lukemelas-b6417697.pthirirg#~jT@gx&1(X@rrrrNrr@r@r@rAr"Gs&c @sBeZdZedeeddejdedddddid d d Z e Z d S) r#zJhttps://download.pytorch.org/models/efficientnet_b6_lukemelas-c76e70fd.pthirirgnU@gv:X@rrrrNrr@r@r@rAr#]s&c @sBeZdZedeeddejdedddddid d d Z e Z d S) r$zJhttps://download.pytorch.org/models/efficientnet_b7_lukemelas-dcc49843.pthiXricrg+U@g'1:X@rrrrNrr@r@r@rAr$ss&c @sBeZdZedeeddejdedddddid d d Z e Z d S) r%zBhttps://download.pytorch.org/models/efficientnet_v2_s-dd5fe13b.pthrri8nGrg;OU@gx&18X@rrrrN rCrDrErrrrr_COMMON_META_V2rrr@r@r@rAr%s&c @sBeZdZedeeddejdedddddid d d Z e Z d S) r&zBhttps://download.pytorch.org/models/efficientnet_v2_m-dc08266a.pthri:rgI +GU@gDlIX@rrrrNrr@r@r@rAr&s&c @sFeZdZedeeddejdddeddddd id d d Z e Z d S)r'zBhttps://download.pytorch.org/models/efficientnet_v2_l-59c71312.pthr)?rr)rrrZmeanZstdiHfrgʡEsU@gOnrX@rrrrN) rCrDrErrrrrrrrr@r@r@rAr's* Z pretrained)rT)rr)rrr~r>cKs0t|}tdddd\}}t|d|||f|S)aEfficientNet B0 model architecture from the `EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks `_ paper. Args: weights (:class:`~torchvision.models.EfficientNet_B0_Weights`, optional): The pretrained weights to use. See :class:`~torchvision.models.EfficientNet_B0_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.efficientnet.EfficientNet`` base class. Please refer to the `source code `_ for more details about this class. .. autoclass:: torchvision.models.EfficientNet_B0_Weights :members: r(rLrrx)rverifyrrrrr~rzr}r@r@rAr(s cKs0t|}tdddd\}}t|d|||f|S)aEfficientNet B1 model architecture from the `EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks `_ paper. Args: weights (:class:`~torchvision.models.EfficientNet_B1_Weights`, optional): The pretrained weights to use. See :class:`~torchvision.models.EfficientNet_B1_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.efficientnet.EfficientNet`` base class. Please refer to the `source code `_ for more details about this class. .. autoclass:: torchvision.models.EfficientNet_B1_Weights :members: r)rL皙?rrx)rrrrrr@r@rAr)s cKs0t|}tdddd\}}t|d|||f|S)aEfficientNet B2 model architecture from the `EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks `_ paper. Args: weights (:class:`~torchvision.models.EfficientNet_B2_Weights`, optional): The pretrained weights to use. See :class:`~torchvision.models.EfficientNet_B2_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.efficientnet.EfficientNet`` base class. Please refer to the `source code `_ for more details about this class. .. autoclass:: torchvision.models.EfficientNet_B2_Weights :members: r*r333333?r333333?)rrrrrr@r@rAr*s cKs0t|}tdddd\}}t|d|||f|S)aEfficientNet B3 model architecture from the `EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks `_ paper. Args: weights (:class:`~torchvision.models.EfficientNet_B3_Weights`, optional): The pretrained weights to use. See :class:`~torchvision.models.EfficientNet_B3_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.efficientnet.EfficientNet`` base class. Please refer to the `source code `_ for more details about this class. .. autoclass:: torchvision.models.EfficientNet_B3_Weights :members: r+rffffff?rr)r rrrrr@r@rAr+0s cKs0t|}tdddd\}}t|d|||f|S)aEfficientNet B4 model architecture from the `EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks `_ paper. Args: weights (:class:`~torchvision.models.EfficientNet_B4_Weights`, optional): The pretrained weights to use. See :class:`~torchvision.models.EfficientNet_B4_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.efficientnet.EfficientNet`` base class. Please refer to the `source code `_ for more details about this class. .. autoclass:: torchvision.models.EfficientNet_B4_Weights :members: r,r?r皙?)r!rrrrr@r@rAr,Ls cKsDt|}tdddd\}}t|d|||fdttjddd i|S) aEfficientNet B5 model architecture from the `EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks `_ paper. Args: weights (:class:`~torchvision.models.EfficientNet_B5_Weights`, optional): The pretrained weights to use. See :class:`~torchvision.models.EfficientNet_B5_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.efficientnet.EfficientNet`` base class. Please refer to the `source code `_ for more details about this class. .. autoclass:: torchvision.models.EfficientNet_B5_Weights :members: r-g?g@rrr^MbP?{Gz?epsZmomentum)r"rrrrr rrr@r@rAr-hs cKsDt|}tdddd\}}t|d|||fdttjddd i|S) aEfficientNet B6 model architecture from the `EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks `_ paper. Args: weights (:class:`~torchvision.models.EfficientNet_B6_Weights`, optional): The pretrained weights to use. See :class:`~torchvision.models.EfficientNet_B6_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.efficientnet.EfficientNet`` base class. Please refer to the `source code `_ for more details about this class. .. autoclass:: torchvision.models.EfficientNet_B6_Weights :members: r.rg@rrr^rrr)r#rrrrr rrr@r@rAr.s cKsDt|}tdddd\}}t|d|||fdttjddd i|S) aEfficientNet B7 model architecture from the `EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks `_ paper. Args: weights (:class:`~torchvision.models.EfficientNet_B7_Weights`, optional): The pretrained weights to use. See :class:`~torchvision.models.EfficientNet_B7_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.efficientnet.EfficientNet`` base class. Please refer to the `source code `_ for more details about this class. .. autoclass:: torchvision.models.EfficientNet_B7_Weights :members: r/g@g@rrr^rrr)r$rrrrr rrr@r@rAr/s cKs<t|}td\}}t|d|||fdttjddi|S)a Constructs an EfficientNetV2-S architecture from `EfficientNetV2: Smaller Models and Faster Training `_. Args: weights (:class:`~torchvision.models.EfficientNet_V2_S_Weights`, optional): The pretrained weights to use. See :class:`~torchvision.models.EfficientNet_V2_S_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.efficientnet.EfficientNet`` base class. Please refer to the `source code `_ for more details about this class. .. autoclass:: torchvision.models.EfficientNet_V2_S_Weights :members: r0rxr^rr)r%rrrrr rrr@r@rAr0s   cKs<t|}td\}}t|d|||fdttjddi|S)a Constructs an EfficientNetV2-M architecture from `EfficientNetV2: Smaller Models and Faster Training `_. Args: weights (:class:`~torchvision.models.EfficientNet_V2_M_Weights`, optional): The pretrained weights to use. See :class:`~torchvision.models.EfficientNet_V2_M_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.efficientnet.EfficientNet`` base class. Please refer to the `source code `_ for more details about this class. .. autoclass:: torchvision.models.EfficientNet_V2_M_Weights :members: r1rr^rr)r&rrrrr rrr@r@rAr1s   cKs<t|}td\}}t|d|||fdttjddi|S)a Constructs an EfficientNetV2-L architecture from `EfficientNetV2: Smaller Models and Faster Training `_. Args: weights (:class:`~torchvision.models.EfficientNet_V2_L_Weights`, optional): The pretrained weights to use. See :class:`~torchvision.models.EfficientNet_V2_L_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.efficientnet.EfficientNet`` base class. Please refer to the `source code `_ for more details about this class. .. autoclass:: torchvision.models.EfficientNet_V2_L_Weights :members: r2rr^rr)r'rrrrr rrr@r@rAr2s   ) _ModelURLs)r(r)r*r+r,r-r.r/)RrrWrZ dataclassesr functoolsrtypingrrrrrr r r rr r Ztorchvision.opsrZops.miscrrZtransforms._presetsrrutilsrZ_apirrZ_metar_utilsrrr__all__r3rKrZrIrPr[rrFrHboolrstrrrrGrrrrrr r!r"r#r$r%r&r'rr(r)r*r+r,r-r.r/r0r1r2rrZ model_urlsr@r@r@rAs~  (   C=~  8,###$$%