o + i'@sddlmZddlmZmZddlmZmZddlZddl m m Z ddlm Z ddl mZddlmZmZmZmZddlmZerWdd lmZdd lmZGd d d eZgZd ddZGddde jZGddde jZGddde jZd%ddZ d&d'd!d"Z! d&d'd#d$Z"dS)() annotations) TYPE_CHECKING TypedDict) NotRequiredUnpackN)nn) ParamAttr)AdaptiveAvgPool2DConv2DDropout MaxPool2D)get_weights_path_from_url)Tensor)Size2c@seZdZUded<ded<dS)_SqueezeNetOptionszNotRequired[int] num_classeszNotRequired[bool] with_poolN)__name__ __module__ __qualname____annotations__rrk/home/app/PaddleOCR-VL-test/.venv_paddleocr/lib/python3.10/site-packages/paddle/vision/models/squeezenet.pyr#s  r)z[https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/SqueezeNet1_0_pretrained.pdparamsZ 30b95af60a2178f03cf9b66cd77e1db1)z[https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/SqueezeNet1_1_pretrained.pdparamsZ a11250d3a1f91d7131fd095ebbf09eee) squeezenet1_0 squeezenet1_1cs,eZdZ ddfd d ZdddZZS) MakeFireConvrinput_channelsintoutput_channels filter_sizerpaddingreturnNonecs(tt||||ttd|_dS)N)r weight_attr bias_attr)super__init__r r_conv)selfrrrr  __class__rrr&7s  zMakeFireConv.__init__xrcCs||}t|}|S)N)r'Frelu)r(r+rrrforwardHs  zMakeFireConv.forward)r) rrrrrrr rr!r")r+rr!rrrrr&r. __classcell__rrr)rr6srcs(eZdZdfdd Zdd d ZZS)MakeFirerrsqueeze_channelsexpand1x1_channelsexpand3x3_channelsr!r"cs<tt||d|_t||d|_t||ddd|_dS)N)r )r%r&rr' _conv_path1 _conv_path2)r(rr2r3r4r)rrr&Os  zMakeFire.__init__inputsrcCs0||}||}||}tj||gddS)Nr5Zaxis)r'r7r8paddleconcat)r(r9r+x1Zx2rrrr.]s   zMakeFire.forward) rrr2rr3rr4rr!r"r9rr!rr/rrr)rr1Nsr1csJeZdZUdZded<ded<ded< ddfd d ZdddZZS) SqueezeNetaSqueezeNet model from `"SqueezeNet: AlexNet-level accuracy with 50x fewer parameters and <0.5MB model size" `_. Args: version (str): Version of SqueezeNet, which can be "1.0" or "1.1". 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 SqueezeNet model. Examples: .. code-block:: python >>> import paddle >>> from paddle.vision.models import SqueezeNet >>> # build v1.0 model >>> model = SqueezeNet(version='1.0') >>> # build v1.1 model >>> # model = SqueezeNet(version='1.1') >>> x = paddle.rand([1, 3, 224, 224]) >>> out = model(x) >>> print(out.shape) [1, 1000] strversionrrboolrTr!r"c st||_||_||_ddg}||vs Jd|d||jdkr{tddddttd |_tddd d |_ t dd d d |_ t dd d d |_ t dddd|_ t dddd|_t dddd|_t dddd|_t dd dd|_t dd dd|_nVtdd dddttd|_tddd d |_ t d d d d |_ t dd d d |_ t dddd|_ t dddd|_t dddd|_t dddd|_t dd dd|_t dd dd|_tddd|_td|dttd|_td|_dS)N1.01.1zsupported versions are z but input version is r6`)strider#r$r)Z kernel_sizerIr @ 0iir5)rIr r#r$g?Zdownscale_in_infer)pmode)r#r$)r%r&rArrr rr'r _poolr1_conv1_conv2_conv3_conv4_conv5_conv6_conv7_conv8r _drop_conv9r _avg_pool)r(rArrZsupported_versionsr)rrr&sb    zSqueezeNet.__init__r9rcCs@||}t|}||}|jdkrG||}||}||}||}||}| |}| |}| |}||}| |}n2||}||}||}||}||}||}| |}| |}| |}| |}|j dkr||}||}|jrt|}||}tj|ddgd}|S)NrDrrHr6r:)r'r,r-rSrArTrUrVrWrXrYrZr[rr\r]rr^r;Zsqueeze)r(r9r+rrrr.s@                             zSqueezeNet.forward)rCT)rAr@rrrrBr!r"r>)rrr__doc__rr&r.r0rrr)rr?ds  9r?archr@rA pretrainedrBkwargsUnpack[_SqueezeNetOptions]r!cKs\t|fi|}|r,|tvsJ|dtt|dt|d}t|}|||S)NzJ model do not have a pretrained model now, you should set pretrained=Falserr5)r? model_urlsr r;loadZset_dict)r`rArarbmodelZ weight_pathparamrrr _squeezenets   rhFcKtdd|fi|S)aSqueezeNet v1.0 model from `"SqueezeNet: AlexNet-level accuracy with 50x fewer parameters and <0.5MB model size" `_. 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:`SqueezeNet `. Returns: :ref:`api_paddle_nn_Layer`. An instance of SqueezeNet v1.0 model. Examples: .. code-block:: python >>> import paddle >>> from paddle.vision.models import squeezenet1_0 >>> # build model >>> model = squeezenet1_0() >>> # build model and load imagenet pretrained weight >>> # model = squeezenet1_0(pretrained=True) >>> x = paddle.rand([1, 3, 224, 224]) >>> out = model(x) >>> print(out.shape) [1, 1000] rrDrhrarbrrrr!rcKri)aSqueezeNet v1.1 model from `"SqueezeNet: AlexNet-level accuracy with 50x fewer parameters and <0.5MB model size" `_. 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:`SqueezeNet `. Returns: :ref:`api_paddle_nn_Layer`. An instance of SqueezeNet v1.1 model. Examples: .. code-block:: python >>> import paddle >>> from paddle.vision.models import squeezenet1_1 >>> # build model >>> model = squeezenet1_1() >>> # build model and load imagenet pretrained weight >>> # model = squeezenet1_1(pretrained=True) >>> x = paddle.rand([1, 3, 224, 224]) >>> out = model(x) >>> print(out.shape) [1, 1000] rrErjrkrrrrrlr) r`r@rAr@rarBrbrcr!r?)F)rarBrbrcr!r?)# __future__rtypingrrZtyping_extensionsrrr;Zpaddle.nn.functionalrZ functionalr,Zpaddle.base.param_attrrZ paddle.nnr r r r Zpaddle.utils.downloadr rZpaddle._typingrr__all__rdZLayerrr1r?rhrrrrrrs4         %