\;jO&ddlZddlmZddlmZmZmZddlmZddlm Z gZ ddd d d d d dZ dZ dZ GddejZGddejZGddejZddZddZddZddZddZddZddZddZy) N)nn)AdaptiveAvgPool2DLinear MaxPool2D)get_weights_path_from_url)ConvNormActivation)zEhttps://paddle-hapi.bj.bcebos.com/models/shufflenet_v2_x0_25.pdparams 1e509b4c140eeb096bb16e214796d03b)zEhttps://paddle-hapi.bj.bcebos.com/models/shufflenet_v2_x0_33.pdparams 3d7b3ab0eaa5c0927ff1026d31b729bd)zDhttps://paddle-hapi.bj.bcebos.com/models/shufflenet_v2_x0_5.pdparams 5e5cee182a7793c4e4c73949b1a71bd4)zDhttps://paddle-hapi.bj.bcebos.com/models/shufflenet_v2_x1_0.pdparams 122d42478b9e81eb49f8a9ede327b1a4)zDhttps://paddle-hapi.bj.bcebos.com/models/shufflenet_v2_x1_5.pdparams faced5827380d73531d0ee027c67826d)zDhttps://paddle-hapi.bj.bcebos.com/models/shufflenet_v2_x2_0.pdparams cd3dddcd8305e7bcd8ad14d1c69a5784)zEhttps://paddle-hapi.bj.bcebos.com/models/shufflenet_v2_swish.pdparams adde0aa3b023e5b0c94a68be1c394b84)shufflenet_v2_x0_25shufflenet_v2_x0_33shufflenet_v2_x0_5shufflenet_v2_x1_0shufflenet_v2_x1_5shufflenet_v2_x2_0shufflenet_v2_swishcx|dk(rtjS|dk(rtjS|ytd|)Nswishreluz*The activation function is not supported: )rSwishReLU RuntimeError)acts jG:\00. PROJECTS\API\Inventory\templateJSON\kerjaOCR\Lib\site-packages\paddle/vision/models/shufflenetv2.pycreate_activation_layerr 8s= g~xx ww GuMNNc|jdd\}}}}||z}tj||||||g}tj|gd}tj|||||g}|S)Nr)shape)rrr#)perm)r$paddlereshape transpose)xgroups batch_size num_channelsheightwidthchannels_per_groups rchannel_shuffler2Csz./ggal+J fe%/  *f&8&%H A 1A q\65 IJA Hr!c<eZdZejffd ZdZxZS)InvertedResidualc t|t|dz|dzdddd||_t|dz|dzd|d|dzd|_t|dz|dzdddd||_y)Nrr%r in_channels out_channels kernel_sizestridepaddingr,activation_layerr&)super__init__r _conv_pw_conv_dw _conv_linearselfr7r8r:r< __class__s rr>zInvertedResidual.__init__Us *#q(%*-  +$)%*1$!  /$)%*- r!c,tj||jddz|jddzgd\}}|j|}|j |}|j |}tj ||gd}t|dS)Nr%r)num_or_sectionsaxisrG)r(splitr$r?r@rAconcatr2rCinputsx1x2outs rforwardzInvertedResidual.forwardus #\\!_16<<?a3GH B ]]2  ]]2    r "mmRH1-sA&&r!__name__ __module__ __qualname__rrr>rP __classcell__rDs@rr4r4TsBD'' @ 'r!r4c<eZdZejffd ZdZxZS)InvertedResidualDSc 2t|t||d|d|d|_t||dzdddd||_t||dzdddd||_t|dz|dzd|d|dzd|_t|dz|dzdddd||_y)Nr&r%r6rr)r=r>r _conv_dw_1_conv_linear_1 _conv_pw_2 _conv_dw_2_conv_linear_2rBs rr>zInvertedResidualDS.__init__s -#$! 1#%*- -#%*- -$)%*1$! 1$)%*- r!c|j|}|j|}|j|}|j|}|j |}t j ||gd}t|dS)Nr%rHr)rZr[r\r]r^r(rJr2rKs rrPzInvertedResidualDS.forwardsm __V $   $ __V $ __R    $mmRH1-sA&&r!rQrVs@rrXrXsBD''3 j'r!rXc*eZdZdZdfd ZdZxZS) ShuffleNetV2aShuffleNetV2 model from `"ShuffleNet V2: Practical Guidelines for Efficient CNN Architecture Design" `_. Args: scale (float, optional): Scale of output channels. Default: True. act (str, optional): Activation function of neural network. Default: "relu". num_classes (int, optional): Output dim of last fc layer. If num_classes <= 0, last fc layer will not be defined. Default: 1000. with_pool (bool, optional): Use pool before the last fc layer or not. Default: True. Returns: :ref:`api_paddle_nn_Layer`. An instance of ShuffleNetV2 model. Examples: .. code-block:: python >>> import paddle >>> from paddle.vision.models import ShuffleNetV2 >>> shufflenet_v2_swish = ShuffleNetV2(scale=1.0, act="swish") >>> x = paddle.rand([1, 3, 224, 224]) >>> out = shufflenet_v2_swish(x) >>> print(out.shape) [1, 1000] c t |||_||_||_gd}t |}|dk(rgd}nL|dk(rgd}nB|dk(rgd}n8|dk(rgd }n.|d k(rgd }n$|d k(rgd }nt dt|zdztd|dddd||_ tddd|_ g|_ t|D]\}} t| D]} | dk(rK|jt!||dz||dzd|t|dzdzt| dzz} nJ|jt#||dz||dzd|t|dzdzt| dzz} |jj%| t|d|dddd||_|rt)d|_|dkDr|d|_t/|d||_yy)N)r#r#?)rf0`Q?)rerf @ri?)rerfrgrh?)rerftirp?)rerfi`irp@)rerfiiizThis scale size:[z] is not implemented!r&r%r)r7r8r9r:r;r<)r9r:r;r)r7r8r:r<_)sublayernamere)r=r>scale num_classes with_poolr NotImplementedErrorstrr _conv1r _max_pool _block_list enumeraterange add_sublayerrXr4append _last_convr _pool2d_avg_out_cr_fc) rCr|rr}r~ stage_repeatsr<stage_out_channelsstage_id num_repeatiblockrDs rr>zShuffleNetV2.__init__sD  &"! 237 D=!:  d]!;  c\!<  c\!>  c\!>  c\!> %#c%j03JJ )+A.-   #qAF$-m$< Hj:&6 --!3(:8a<(H);HqL)I#$-= " !A.4s1q5zA.E!--!1(:8a<(H);HqL)I#$-= " !A.4s1q5zA.E  ''.+'%=0-*2.+B/-   03D  ?,R0DK04kBDH r!cH|j|}|j|}|jD] }||} |j|}|jr|j |}|j dkDr)tj|dd}|j|}|S)Nrr%re) start_axis stop_axis) rrrrr~rr}r(flattenr)rCrLr+invs rrPzShuffleNetV2.forward,s KK  NN1 ##CAA$ OOA  >>  #A   a qQ"=A Ar!)rqriT)rRrSrT__doc__r>rPrUrVs@rraras4LC\ r!rac tdi|}|rX|tvs J|dtt|dt|d}tj|}|j ||S)NzJ model do not have a pretrained model now, you should set pretrained=Falserr%)ra model_urlsrr(loadset_dict)arch pretrainedkwargsmodel weight_pathparams r_shufflenet_v2r<sx  "6 "E J  _V] ^ _ / t Q D!1!!4  K( u Lr!c t dd|d|S)aRShuffleNetV2 with 0.25x output channels, as described in `"ShuffleNet V2: Practical Guidelines for Ecient CNN Architecture Design" `_. Args: pretrained (bool, optional): Whether to load pre-trained weights. If True, returns a model pre-trained on ImageNet. Default: False. **kwargs (optional): Additional keyword arguments. For details, please refer to :ref:`ShuffleNetV2 `. Returns: :ref:`api_paddle_nn_Layer`. An instance of ShuffleNetV2 with 0.25x output channels. Examples: .. code-block:: python >>> import paddle >>> from paddle.vision.models import shufflenet_v2_x0_25 >>> # build model >>> model = shufflenet_v2_x0_25() >>> # build model and load imagenet pretrained weight >>> # model = shufflenet_v2_x0_25(pretrained=True) >>> x = paddle.rand([1, 3, 224, 224]) >>> out = model(x) >>> print(out.shape) [1, 1000] rdr|r)rrrrs rrrK&<  %)j DJ r!c t dd|d|S)aRShuffleNetV2 with 0.33x output channels, as described in `"ShuffleNet V2: Practical Guidelines for Ecient CNN Architecture Design" `_. Args: pretrained (bool, optional): Whether to load pre-trained weights. If True, returns a model pre-trained on ImageNet. Default: False. **kwargs (optional): Additional keyword arguments. For details, please refer to :ref:`ShuffleNetV2 `. Returns: :ref:`api_paddle_nn_Layer`. An instance of ShuffleNetV2 with 0.33x output channels. Examples: .. code-block:: python >>> import paddle >>> from paddle.vision.models import shufflenet_v2_x0_33 >>> # build model >>> model = shufflenet_v2_x0_33() >>> # build model and load imagenet pretrained weight >>> # model = shufflenet_v2_x0_33(pretrained=True) >>> x = paddle.rand([1, 3, 224, 224]) >>> out = model(x) >>> print(out.shape) [1, 1000] rjr)rrrs rrrnrr!c t dd|d|S)aMShuffleNetV2 with 0.5x output channels, as described in `"ShuffleNet V2: Practical Guidelines for Ecient CNN Architecture Design" `_. Args: pretrained (bool, optional): Whether to load pre-trained weights. If True, returns a model pre-trained on ImageNet. Default: False. **kwargs (optional): Additional keyword arguments. For details, please refer to :ref:`ShuffleNetV2 `. Returns: :ref:`api_paddle_nn_Layer`. An instance of ShuffleNetV2 with 0.5x output channels. Examples: .. code-block:: python >>> import paddle >>> from paddle.vision.models import shufflenet_v2_x0_5 >>> # build model >>> model = shufflenet_v2_x0_5() >>> # build model and load imagenet pretrained weight >>> # model = shufflenet_v2_x0_5(pretrained=True) >>> x = paddle.rand([1, 3, 224, 224]) >>> out = model(x) >>> print(out.shape) [1, 1000] rnr)rrrs rrr&<  $'J BH r!c t dd|d|S)aMShuffleNetV2 with 1.0x output channels, as described in `"ShuffleNet V2: Practical Guidelines for Ecient CNN Architecture Design" `_. Args: pretrained (bool, optional): Whether to load pre-trained weights. If True, returns a model pre-trained on ImageNet. Default: False. **kwargs (optional): Additional keyword arguments. For details, please refer to :ref:`ShuffleNetV2 `. Returns: :ref:`api_paddle_nn_Layer`. An instance of ShuffleNetV2 with 1.0x output channels. Examples: .. code-block:: python >>> import paddle >>> from paddle.vision.models import shufflenet_v2_x1_0 >>> # build model >>> model = shufflenet_v2_x1_0() >>> # build model and load imagenet pretrained weight >>> # model = shufflenet_v2_x1_0(pretrained=True) >>> x = paddle.rand([1, 3, 224, 224]) >>> out = model(x) >>> print(out.shape) [1, 1000] rqr)rrrs rrrrr!c t dd|d|S)aMShuffleNetV2 with 1.5x output channels, as described in `"ShuffleNet V2: Practical Guidelines for Ecient CNN Architecture Design" `_. Args: pretrained (bool, optional): Whether to load pre-trained weights. If True, returns a model pre-trained on ImageNet. Default: False. **kwargs (optional): Additional keyword arguments. For details, please refer to :ref:`ShuffleNetV2 `. Returns: :ref:`api_paddle_nn_Layer`. An instance of ShuffleNetV2 with 1.5x output channels. Examples: .. code-block:: python >>> import paddle >>> from paddle.vision.models import shufflenet_v2_x1_5 >>> # build model >>> model = shufflenet_v2_x1_5() >>> # build model and load imagenet pretrained weight >>> # model = shufflenet_v2_x1_5(pretrained=True) >>> x = paddle.rand([1, 3, 224, 224]) >>> out = model(x) >>> print(out.shape) [1, 1000] rtr)rrrs rrrrr!c t dd|d|S)aMShuffleNetV2 with 2.0x output channels, as described in `"ShuffleNet V2: Practical Guidelines for Ecient CNN Architecture Design" `_. Args: pretrained (bool, optional): Whether to load pre-trained weights. If True, returns a model pre-trained on ImageNet. Default: False. **kwargs (optional): Additional keyword arguments. For details, please refer to :ref:`ShuffleNetV2 `. Returns: :ref:`api_paddle_nn_Layer`. An instance of ShuffleNetV2 with 2.0x output channels. Examples: .. code-block:: python >>> import paddle >>> from paddle.vision.models import shufflenet_v2_x2_0 >>> # build model >>> model = shufflenet_v2_x2_0() >>> # build model and load imagenet pretrained weight >>> # model = shufflenet_v2_x2_0(pretrained=True) >>> x = paddle.rand([1, 3, 224, 224]) >>> out = model(x) >>> print(out.shape) [1, 1000] rvr)rrrs rrrrr!c "t ddd|d|S)aZShuffleNetV2 with swish activation function, as described in `"ShuffleNet V2: Practical Guidelines for Ecient CNN Architecture Design" `_. Args: pretrained (bool, optional): Whether to load pre-trained weights. If True, returns a model pre-trained on ImageNet. Default: False. **kwargs (optional): Additional keyword arguments. For details, please refer to :ref:`ShuffleNetV2 `. Returns: :ref:`api_paddle_nn_Layer`. An instance of ShuffleNetV2 with swish activation function. Examples: .. code-block:: python >>> import paddle >>> from paddle.vision.models import shufflenet_v2_swish >>> # build model >>> model = shufflenet_v2_swish() >>> # build model and load imagenet pretrained weight >>> # model = shufflenet_v2_swish(pretrained=True) >>> x = paddle.rand([1, 3, 224, 224]) >>> out = model(x) >>> print(out.shape) [1, 1000] rqr)r|rr)rrrs rrrs,<        r!)F)r(r paddle.nnrrrpaddle.utils.downloadropsr __all__rr r2Layerr4rXrarrrrrrrrrr!rrs::;$ 3 @O "+'rxx+'\>'>'Bv288vr  F F F F F F$r!