o + iD@sUddlmZddlZddlmZddlmZddlZddlmZddl m Z ddl m Z m Z mZmZmZmZmZddlmZdd lmZermdd lmZmZdd lmZdd lmZdd lmZedZGdddeZgZddddddZ de!d<Gdddej"Z#Gdddej"Z$Gdddej"Z%Gdddej"Z&Gd d!d!ej"Z'Gd"d#d#ej"Z(d:d-d.Z) /d;dr@rCyr#r#r$forwardn  zBNACConvLayer.forwardr(rr(r)r*r+r,r+r-rr.rr/rr0r+r1r2r3r4rCrr3rrr r!r<rH __classcell__r#r#rAr$r'Ur'cs2eZdZUded<dfd d Zdd dZZS) DenseLayerfloatrr*r+ growth_raterr3r4csZt||_t|||dddd|_t|||dddd|_|r+t|dd|_dSdS)Nr(rr*r,r-r/r.Zdownscale_in_infer)pmode)r;r<rr' bn_ac_func1 bn_ac_func2r dropout_func)r@r*rRrrrAr#r$r<ws& zDenseLayer.__init__rCrcCs:||}||}|jr||}tj||gdd}|S)Nr()Zaxis)rWrXrrYpaddleconcat)r@rCconvr#r#r$rHs   zDenseLayer.forward) r*r+rRr+rr+rrQr3r4rLrr r!r"r<rHrNr#r#rAr$rPts rPcs6eZdZUded< ddfd d ZdddZZS) DenseBlockrQrNr*r+ num_layersrrRname str | Noner3r4c sbt||_g|_|}t|D]}|j||d|dt||||d||}qdS)N_r()r*rRrr)r;r<rdense_layer_funcrangeappend add_sublayerrP) r@r*r_rrRrr`Z pre_channellayerrAr#r$r<s$   zDenseBlock.__init__rCrcCs|}|jD]}||}q|SrE)rc)r@rCr\funcr#r#r$rHs  zDenseBlock.forwardrE)r*r+r_r+rr+rRr+rrQr`rar3r4rLr]r#r#rAr$r^s  r^cs(eZdZd fdd Zd d d ZZS)TransitionLayerr*r+num_output_featuresr3r4cs2tt||dddd|_tdddd|_dS)Nr(rrSr7r.r8)r;r<r' conv_ac_funcr pool2d_avg)r@r*rjrAr#r$r<s zTransitionLayer.__init__rCrcCrDrE)rmrnrFr#r#r$rHrIzTransitionLayer.forward)r*r+rjr+r3r4rLrMr#r#rAr$ris ricr&) ConvBNLayerr(rr)r*r+r,r-rr.r/r0r1r2r3r4c s8tt||||||tdd|_t||d|_dS)NFr6r5)r;r<r rr>r r=r?rAr#r$r<s  zConvBNLayer.__init__rCrcCrDrE)r>r=rFr#r#r$rHrIzConvBNLayer.forwardrJrKrLrMr#r#rAr$rorOrocsJeZdZUdZded<ded<    ddfdd ZdddZZS)DenseNetaDenseNet model from `"Densely Connected Convolutional Networks" `_. Args: layers (int, optional): Layers of DenseNet. Default: 121. bn_size (int, optional): Expansion of growth rate in the middle layer. Default: 4. dropout (float, optional): Dropout rate. Default: :math:`0.0`. 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 DenseNet model. Examples: .. code-block:: python >>> import paddle >>> from paddle.vision.models import DenseNet >>> # Build model >>> densenet = DenseNet() >>> x = paddle.rand([1, 3, 224, 224]) >>> out = densenet(x) >>> print(out.shape) [1, 1000] r+rboolryTlayersrrrQr3r4cst||_||_gd}||vsJd|d|ddgdfddgd fddgd fddgd fddgd fd}||\}} } td |ddd dd|_td ddd|_| |_g|_ g|_ |} |} t | D]O\} }|j | d| dt| ||| |dt| dd| || } | } | t| dkr|j | d| ddt| | dd| d} | d} qit| dd|_|jrtd|_|jdkrdt| d}t| |tt| |dtd|_dSdS)N)rrzsupported layers are z but input layer is @ ) `0)r}r~$r)r}r~r|r|)r}r~rr|)r}r~r{rrTrkr))r*r,r-r.r/r1r(rlZdb_conv_r\)r*r_rrRrr`Ztr_convZ_blk)r*rjr5rg?)Z initializer)r9r:)r;r<rrro conv1_funcr pool2d_max block_configdense_block_func_listtransition_func_list enumeratererfr^r2lenrir batch_normrrnmathsqrtr rrout)r@rvrrrrZsupported_layersZ densenet_specZnum_init_featuresrRrZpre_num_channelsZ num_featuresir_ZstdvrAr#r$r<s             zDenseNet.__init__rCrcCs||}||}t|jD]\}}|j||}|t|jdkr*|j||}q||}|jr8| |}|j dkrJt j |ddd}| |}|S)Nr(r)Z start_axisZ stop_axis)rrrrrrrrrrnrrZflattenr)r@rCr\rr_rGr#r#r$rHfs      zDenseNet.forward)rrrsrtruT) rvr+rr+rrQrr+rrqr3r4rL)rr r!__doc__r"r<rHrNr#r#rAr$rps Rrparch _DenseNetArchrvr+ pretrainedrqkwargsUnpack[_DenseNetOptions]r3cKs^tdd|i|}|r-|tvsJ|dtt|dt|d}t|}|||S)NrvzJ model do not have a pretrained model now, you should set pretrained=Falserr(r#)rpr%rrZloadZset_dict)rrvrrmodelZ weight_pathparamr#r#r$ _densenet{s   rFcKtdd|fi|S)aDenseNet 121-layer model from `"Densely Connected Convolutional Networks" `_. 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:`DenseNet `. Returns: :ref:`api_paddle_nn_Layer`. An instance of DenseNet 121-layer model. Examples: .. code-block:: python >>> import paddle >>> from paddle.vision.models import densenet121 >>> # Build model >>> model = densenet121() >>> # Build model and load imagenet pretrained weight >>> # model = densenet121(pretrained=True) >>> x = paddle.rand([1, 3, 224, 224]) >>> out = model(x) >>> print(out.shape) [1, 1000] rrrrrrr#r#r$r rcKr)aDenseNet 161-layer model from `"Densely Connected Convolutional Networks" `_. Args: pretrained (bool, optional): Whether to load pre-trained weights. If True, returns a model pre-trained on ImageNet. Default: False. **kwargs (optional): Additional keyword arguments. For details, please refer to :ref:`DenseNet `. Returns: :ref:`api_paddle_nn_Layer`. An instance of DenseNet 161-layer model. Examples: .. code-block:: python >>> import paddle >>> from paddle.vision.models import densenet161 >>> # Build model >>> model = densenet161() >>> # Build model and load imagenet pretrained weight >>> # model = densenet161(pretrained=True) >>> x = paddle.rand([1, 3, 224, 224]) >>> out = model(x) >>> print(out.shape) [1, 1000] rrwrrr#r#r$rrrcKr)aDenseNet 169-layer model from `"Densely Connected Convolutional Networks" `_. 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:`DenseNet `. Returns: :ref:`api_paddle_nn_Layer`. An instance of DenseNet 169-layer model. Examples: .. code-block:: python >>> import paddle >>> from paddle.vision.models import densenet169 >>> # Build model >>> model = densenet169() >>> # Build model and load imagenet pretrained weight >>> # model = densenet169(pretrained=True) >>> x = paddle.rand([1, 3, 224, 224]) >>> out = model(x) >>> print(out.shape) [1, 1000] rrxrrr#r#r$rrrcKr)aDenseNet 201-layer model from `"Densely Connected Convolutional Networks" `_. Args: pretrained (bool, optional): Whether to load pre-trained weights. If True, returns a model pre-trained on ImageNet. Default: False. **kwargs (optional): Additional keyword arguments. For details, please refer to :ref:`DenseNet `. Returns: :ref:`api_paddle_nn_Layer`. An instance of DenseNet 201-layer model. Examples: .. code-block:: python >>> import paddle >>> from paddle.vision.models import densenet201 >>> # Build model >>> model = densenet201() >>> # Build model and load imagenet pretrained weight >>> # model = densenet201(pretrained=True) >>> x = paddle.rand([1, 3, 224, 224]) >>> out = model(x) >>> print(out.shape) [1, 1000] rryrrr#r#r$rsrcKr)aDenseNet 264-layer model from `"Densely Connected Convolutional Networks" `_. 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:`DenseNet `. Returns: :ref:`api_paddle_nn_Layer`. An instance of DenseNet 264-layer model. Examples: .. code-block:: python >>> import paddle >>> from paddle.vision.models import densenet264 >>> # Build model >>> model = densenet264() >>> # Build model and load imagenet pretrained weight >>> # model = densenet264(pretrained=True) >>> x = paddle.rand([1, 3, 224, 224]) >>> out = model(x) >>> print(out.shape) [1, 1000] rrzrrr#r#r$rrr) rrrvr+rrqrrr3rp)F)rrqrrr3rp)/ __future__rrtypingrZtyping_extensionsrrZrZpaddle.base.param_attrrZ paddle.nnrrr r r r r Zpaddle.nn.initializerrZpaddle.utils.downloadrrrrrZpaddle._typingrrr__all__r%r"ZLayerr'rPr^rirorprrrrrrr#r#r#r$sX     $    %&  $ $ $ #