|jHddlmZddlmZddlmZmZmZddlm Z m Z ddl Z ddl m Z ddl mZdd lmZd d lmZerdd l mZGd deZgZdddZGdde jZGddZGdde jZGdde jZGddeZGddeZ d,d-d)Z d,d.d*Z d,d/d+Z dS)0) annotations)partial) TYPE_CHECKINGCallable TypedDict) NotRequiredUnpackN)nn)get_weights_path_from_url)ConvNormActivation_make_divisible)Tensorc$eZdZUded<ded<dS)_MobileNetV3OptionszNotRequired[int] num_classeszNotRequired[bool] with_poolN)__name__ __module__ __qualname____annotations__p/lsinfo/ai/hellotax_ai/data_center/backend/venv/lib/python3.11/site-packages/paddle/vision/models/mobilenetv3.pyrr$s*%%%%$$$$$$rr)zIhttps://paddle-hapi.bj.bcebos.com/models/mobilenet_v3_small_x1.0.pdparams 34fe0e7c1f8b00b2b056ad6788d0590c)zIhttps://paddle-hapi.bj.bcebos.com/models/mobilenet_v3_large_x1.0.pdparams 118db5792b4e183b925d8e8e334db3df)zmobilenet_v3_small_x1.0zmobilenet_v3_large_x1.0cNeZdZdZejejfdfd Zdd ZddZ xZ S)SqueezeExcitationa This block implements the Squeeze-and-Excitation block from https://arxiv.org/abs/1709.01507 (see Fig. 1). Parameters ``activation``, and ``scale_activation`` correspond to ``delta`` and ``sigma`` in eq. 3. This code is based on the torchvision code with modifications. You can also see at https://github.com/pytorch/vision/blob/main/torchvision/ops/misc.py#L127 Args: input_channels (int): Number of channels in the input image. squeeze_channels (int): Number of squeeze channels. activation (Callable[..., paddle.nn.Layer], optional): ``delta`` activation. Default: ``paddle.nn.ReLU``. scale_activation (Callable[..., paddle.nn.Layer]): ``sigma`` activation. Default: ``paddle.nn.Sigmoid``. input_channelsintsqueeze_channels activationCallable[..., nn.Layer]scale_activationreturnNonec"ttjd|_tj||d|_tj||d|_||_||_ dS)Nr) super__init__r AdaptiveAvgPool2DavgpoolConv2Dfc1fc2r$r&)selfr!r#r$r& __class__s rr+zSqueezeExcitation.__init__Es} +A.. 9^-=qAA9-~qAA$*,, 0 0 2 2rinputrc||}||}||}||}||SN)r-r/r$r0r&r1r3scales r_scalezSqueezeExcitation._scaleSsV U##&&$$U+++rc6||}||zSr5)r8r6s rforwardzSqueezeExcitation.forwardZs E""u}r) r!r"r#r"r$r%r&r%r'r()r3rr'r) rrr__doc__r ReLUSigmoidr+r8r: __classcell__r2s@rr r 7s  "/1g46J 3 3 3 3 3 3 3,,,,rr c2eZdZ dddZeddZdS)InvertedResidualConfig? in_channelsr"kernelexpanded_channels out_channelsuse_seboolr$strstrider7floatc p||||_||_||||_||||_||_|d|_nB|dkrtj|_n*|dkrtj |_ntd|||_ dS)N)r7relu hardswishz*The activation function is not supported: ) adjust_channelsrCrDrErFrGactivation_layerr r< Hardswish RuntimeErrorrJ) r1rCrDrErFrGr$rJr7s rr+zInvertedResidualConfig.__init__`s // 5/II !%!5!5 U"6" " !00U0KK  $(D ! ! 6 ! !$&GD ! ! ; & &$&LD ! !IZII  rc(t||zdS)Nr)channelsr7s rrOz&InvertedResidualConfig.adjust_channels~sx%/333rN)rB)rCr"rDr"rEr"rFr"rGrHr$rIrJr"r7rK)rrrr+ staticmethodrOrrrrArA_sR<444\444rrAc(eZdZdfd ZddZxZS)InvertedResidualrCr"rErF filter_sizerJrGrHrPr% norm_layerr'r(c t|dko||k|_||_||k|_|jrt ||ddd|||_t ||||t|dz dz||||_|jr1t|t|dztj |_ t ||ddd|d|_dS)Nrr)rCrF kernel_sizerJpaddingrZrPr rCrFr\rJr]groupsrZrP)r&)r*r+use_res_connectrGexpandr expand_convr"bottleneck_convr rr Hardsigmoidmid_se linear_conv) r1rCrErFrYrJrGrPrZr2s rr+zInvertedResidual.__init__s, %{J{l/J !%66 ; 1'.%!1   D  2)*#qQ.//$!-  ; +! 1Q 677!#DK .)%!!   rxrc|}|jr||}||}|jr||}||}|jrtj||}|Sr5) rbrcrdrGrfrgrapaddleadd)r1rhidentitys rr:zInvertedResidual.forwards ; $  ##A   # # ;  AA   Q     ( 8Q''Ar)rCr"rEr"rFr"rYr"rJr"rGrHrPr%rZr%r'r(rhrr'r)rrrr+r:r>r?s@rrXrXsR5 5 5 5 5 5 n        rrXcTeZdZUdZded<ded<ded< ddfd ZddZxZS) MobileNetV3aOMobileNetV3 model from `"Searching for MobileNetV3" `_. Args: config (list[InvertedResidualConfig]): MobileNetV3 depthwise blocks config. last_channel (int): The number of channels on the penultimate layer. scale (float, optional): Scale of channels in each layer. Default: 1.0. num_classes (int, optional): Output dim of last fc layer. If num_classes <=0, last fc layer will not be defined. Default: 1000. with_pool (bool, optional): Use pool before the last fc layer or not. Default: True. rKr7r"rrHrrBTconfiglist[InvertedResidualConfig] last_channelr'r(c \t||_||_||_||_||_|dj|_|dj|_ |j dz|_ ttj ddtd|jddd d tj |_tjfd |jD|_t|j |j d d dd tj |_|rtjd |_|dkrwtjtj|j |jtjtjd tj|j||_dSdS)NrgMbP?gGz?)epsilonmomentumr r)rCrFr\rJr]r_rPrZc g|]<}t|j|j|j|j|j|j|j=S))rCrErFrYrJrGrPrZ)rXrCrErFrDrJrGrP).0cfgrZs r z(MobileNetV3.__init__..sc   ! #&)&;!$!1 # ::%(%9)      rr^g?)p)r*r+rqr7rsrrrCfirstconv_in_channelslastconv_in_channelslastconv_out_channelsrr BatchNorm2Dr rQconv Sequentialblockslastconvr,r-LinearDropout classifier)r1rqrsr7rrrZr2s @rr+zMobileNetV3.__init__s   (&"%+AY%:"$*2J$:!%)%>%B"R^UTJJJ &3\!     m     ;     +13!\       3/22DL ?? m $4d6GHH  S!!! $+[99 DOOO ?rrhrc&||}||}||}|jr||}|jdkr*t j|d}||}|S)Nrr) rrrrr-rrjflattenr)r1rhs rr:zMobileNetV3.forwards IIaLL KKNN MM!   >  QA  a  q!$$A""ArrBrpT) rqrrrsr"r7rKrr"rrHr'r(rm)rrrr;rr+r:r>r?s@rroros  LLLOOO  CCCCCCCJ        rroc,eZdZdZ ddfd ZxZS)MobileNetV3SmallaMobileNetV3 Small architecture model from `"Searching for MobileNetV3" `_. Args: scale (float, optional): Scale of channels in each layer. Default: 1.0. num_classes (int, optional): Output dim of last fc layer. If num_classes <= 0, last fc layer will not be defined. Default: 1000. with_pool (bool, optional): Use pool before the last fc layer or not. Default: True. Returns: :ref:`api_paddle_nn_Layer`. An instance of MobileNetV3 Small architecture model. Examples: .. code-block:: pycon >>> import paddle >>> from paddle.vision.models import MobileNetV3Small >>> # Build model >>> model = MobileNetV3Small(scale=1.0) >>> x = paddle.rand([1, 3, 224, 224]) >>> out = model(x) >>> print(out.shape) paddle.Size([1, 1000]) rBrpTr7rKrr"rrHr'r(cLtddddddd|tddddddd|tddd dddd |tdd d d ddd|td d dd ddd |td d dd ddd |td d ddddd |tdd ddddd |tdd dd ddd|td d dd ddd |td d dd ddd |g }td|zd}t|||||dS)NryTrMr HFXr`(rNx0i i@irTrsr7rrrArr*r+r1r7rrrqrsr2s rr+zMobileNetV3Small.__init__Jsn #2q"b$5 I I "2q"b%E J J "2q"b%E J J "2q"b$ Q N N "2q#r4a O O "2q#r4a O O "2q#r4a O O "2q#r4a O O "2q#r4a O O "2q#r4a O O "2q#r4a O O  'te|Q77   %#      rrr7rKrr"rrHr'r(rrrr;r+r>r?s@rrr-sW<            rrc,eZdZdZ ddfd ZxZS)MobileNetV3LargeaMobileNetV3 Large architecture model from `"Searching for MobileNetV3" `_. Args: scale (float, optional): Scale of channels in each layer. Default: 1.0. num_classes (int, optional): Output dim of last fc layer. If num_classes <= 0, last fc layer will not be defined. Default: 1000. with_pool (bool, optional): Use pool before the last fc layer or not. Default: True. Returns: :ref:`api_paddle_nn_Layer`. An instance of MobileNetV3 Large architecture model. Examples: .. code-block:: pycon >>> import paddle >>> from paddle.vision.models import MobileNetV3Large >>> # Build model >>> model = MobileNetV3Large(scale=1.0) >>> x = paddle.rand([1, 3, 224, 224]) >>> out = model(x) >>> print(out.shape) paddle.Size([1, 1000]) rBrpTr7rKrr"rrHr'r(ctddddddd|tddddddd|tddd dddd|tdd d d d dd|td d d d d dd|td d d d d dd|td dddddd|tddddddd|tddddddd|tddddddd|tddddd dd|tddddd dd|tdd ddd dd|tdd ddd dd|tdd ddd dd|g}td|zd}t|||||dS)NrryFrMr@rr rrrTrrPrNipiiirTrrrs rr+zMobileNetV3Large.__init__s #2q"b%E J J "2q"b%E J J "2q"b%E J J "2q"b$5 I I "2q#r4E J J "2q#r4E J J "AsB{Au   #AsB{Au   #AsB{Au   #AsB{Au   #AsC{Au   #QS$ Q   #QS$ Q   #QS$ Q   #QS$ Q  ?" F'te|Q77   %#      rrrrr?s@rrrgsW< 0 0 0 0 0 0 0 0 0 0 0 rrFrBarchrI pretrainedrHr7rKkwargsUnpack[_MobileNetV3Options]r'c @|dkrtdd|i|}ntdd|i|}|rx|d|}|tvs J|dtt|dt|d}t j|}|||S)Nmobilenet_v3_larger7_xzJ model do not have a pretrained model now, you should set pretrained=Falserrr)rr model_urlsr rjloadset_dict)rrr7rmodel weight_pathparams r _mobilenet_v3rs  ### 77u777 77u777 !!%!!z!!! _ _ _"!!0 t Q D!1!!4   K(( u Lrc &t d||d|}|S)aMobileNetV3 Small architecture model from `"Searching for MobileNetV3" `_. Args: pretrained (bool, optional): Whether to load pre-trained weights. If True, returns a model pre-trained on ImageNet. Default: False. scale (float, optional): Scale of channels in each layer. Default: 1.0. **kwargs (optional): Additional keyword arguments. For details, please refer to :ref:`MobileNetV3Small `. Returns: :ref:`api_paddle_nn_Layer`. An instance of MobileNetV3 Small architecture model. Examples: .. code-block:: pycon >>> import paddle >>> from paddle.vision.models import mobilenet_v3_small >>> # Build model >>> model = mobilenet_v3_small() >>> # Build model and load imagenet pretrained weight >>> # model = mobilenet_v3_small(pretrained=True) >>> # Build mobilenet v3 small model with scale=0.5 >>> model = mobilenet_v3_small(scale=0.5) >>> x = paddle.rand([1, 3, 224, 224]) >>> out = model(x) >>> print(out.shape) paddle.Size([1, 1000]) mobilenet_v3_smallr7r)rrrr7rrs rrr5J  $)j  DJ  E Lrc &t d||d|}|S)aMobileNetV3 Large architecture model from `"Searching for MobileNetV3" `_. Args: pretrained (bool, optional): Whether to load pre-trained weights. If True, returns a model pre-trained on ImageNet. Default: False. scale (float, optional): Scale of channels in each layer. Default: 1.0. **kwargs (optional): Additional keyword arguments. For details, please refer to :ref:`MobileNetV3Large `. Returns: :ref:`api_paddle_nn_Layer`. An instance of MobileNetV3 Large architecture model. Examples: .. code-block:: pycon >>> import paddle >>> from paddle.vision.models import mobilenet_v3_large >>> # Build model >>> model = mobilenet_v3_large() >>> # Build model and load imagenet pretrained weight >>> # model = mobilenet_v3_large(pretrained=True) >>> # Build mobilenet v3 large model with scale=0.5 >>> model = mobilenet_v3_large(scale=0.5) >>> x = paddle.rand([1, 3, 224, 224]) >>> out = model(x) >>> print(out.shape) paddle.Size([1, 1000]) rr)rrrs rrrrr)FrB) rrIrrHr7rKrrr'ro)rrHr7rKrrr'r)rrHr7rKrrr'r)! __future__r functoolsrtypingrrrtyping_extensionsrr rjr paddle.utils.downloadr opsr _utilsrrr__all__rLayerr rArXrorrrrrrrrrs#""""" 21111111 ;;;;;;$$$$$$######%%%%%%i%%%      %%%%%%%%P!4!4!4!4!4!4!4!4HBBBBBrxBBBJbbbbb"(bbbJ7 7 7 7 7 {7 7 7 tM M M M M {M M M d2(((((X(((((((r