U (dA @sddlZddlmZddlmZddlmZmZmZm Z ddl Z ddl m Z ddl m mZddlmmZddl mZddlmZddlmZd d lmZmZd d lmZd d lmZmZd ddddddddg Z Gddde j!Z"Gddde j#Z$Gddde j%Z&Gdd d e j!Z'e j!ee(ddddZ)e*e e*e*e*e*fe*eee(ee'd d!d"Z+d#ed$d%d&Z,Gd'ddeZ-Gd(ddeZ.Gd)ddeZ/Gd*ddeZ0ed+e-j1fd,dd-d.ee-e(ee'd/d0dZ2ed+e.j1fd,dd-d.ee.e(ee'd/d1dZ3ed+e/j1fd,dd-d.ee/e(ee'd/d2dZ4ed+e0j1fd,dd-d.ee0e(ee'd/d3dZ5d d4lm6Z6e6e-j1j7e/j1j7e0j1j7e.j1j7d5Z8dS)6N) OrderedDict)partial)AnyListOptionalTuple)Tensor)ImageClassification)_log_api_usage_once) WeightsEnumWeights)_IMAGENET_CATEGORIES)handle_legacy_interface_ovewrite_named_paramDenseNetDenseNet121_WeightsDenseNet161_WeightsDenseNet169_WeightsDenseNet201_Weights densenet121 densenet161 densenet169 densenet201cseZdZdeeeeeddfdd ZeeedddZ eeed d d Z e j j eeed d d Ze j jeeed ddZe j jeed ddZeed ddZZS) _DenseLayerFN)num_input_features growth_ratebn_size drop_ratememory_efficientreturncstt||_tjdd|_tj|||dddd|_t|||_ tjdd|_ tj|||ddddd|_ t ||_ ||_dS)NTZinplacer F kernel_sizestridebiasr$r%paddingr&)super__init__nn BatchNorm2dnorm1ReLUrelu1Conv2dconv1norm2relu2conv2floatrr )selfrrrrr  __class__?/tmp/pip-unpacked-wheel-vx7f76es/torchvision/models/densenet.pyr+!s   z_DenseLayer.__init__)inputsr!cCs&t|d}||||}|SNr )torchcatr2r0r.)r7r<Zconcated_featuresbottleneck_outputr:r:r; bn_function0s z_DenseLayer.bn_function)inputr!cCs|D]}|jrdSqdS)NTF)Z requires_grad)r7rBZtensorr:r:r;any_requires_grad6sz_DenseLayer.any_requires_gradcsfdd}tj|f|S)Ncs |SN)rA)r<r7r:r;closure>sz7_DenseLayer.call_checkpoint_bottleneck..closure)cp checkpoint)r7rBrFr:rEr;call_checkpoint_bottleneck<s z&_DenseLayer.call_checkpoint_bottleneckcCsdSrDr:r7rBr:r:r;forwardCsz_DenseLayer.forwardcCsdSrDr:rJr:r:r;rKGscCst|tr|g}n|}|jrD||rDtjr8td||}n | |}| | | |}|j dkrtj||j |jd}|S)Nz%Memory Efficient not supported in JITr)ptraining) isinstancerr rCr>jitZ is_scripting ExceptionrIrAr5r4r3rFZdropoutrM)r7rBZ prev_featuresr@ new_featuresr:r:r;rKMs     )F)__name__ __module__ __qualname__intr6boolr+rrrArCr>rOZunusedrIZ_overload_methodrK __classcell__r:r:r8r;r s$rc sDeZdZdZd eeeeeeddfdd Zeeddd Z Z S) _DenseBlockr FN) num_layersrrrrr r!c sJtt|D]2}t|||||||d}|d|d|qdS)N)rrrr z denselayer%dr )r*r+ranger add_module) r7rZrrrrr ilayerr8r:r;r+ds   z_DenseBlock.__init__) init_featuresr!cCs6|g}|D]\}}||}||qt|dSr=)itemsappendr>r?)r7r_featuresnamer^rRr:r:r;rKxs  z_DenseBlock.forward)F) rSrTrU_versionrVr6rWr+rrKrXr:r:r8r;rYas rYcs&eZdZeeddfdd ZZS) _TransitionN)rnum_output_featuresr!csNtt||_tjdd|_tj||dddd|_tj ddd|_ dS)NTr"r Fr#r )r$r%) r*r+r,r-Znormr/relur1convZ AvgPool2dpool)r7rrfr8r:r;r+s   z_Transition.__init__)rSrTrUrVr+rXr:r:r8r;resrec sReZdZdZdeeeeeefeeeeed d fd d Ze e d ddZ Z S)raKDensenet-BC model class, based on `"Densely Connected Convolutional Networks" `_. Args: growth_rate (int) - how many filters to add each layer (`k` in paper) block_config (list of 4 ints) - how many layers in each pooling block num_init_features (int) - the number of filters to learn in the first convolution layer bn_size (int) - multiplicative factor for number of bottle neck layers (i.e. bn_size * k features in the bottleneck layer) drop_rate (float) - dropout rate after each dense layer num_classes (int) - number of classification classes memory_efficient (bool) - If True, uses checkpointing. Much more memory efficient, but slower. Default: *False*. See `"paper" `_.  @rFN)r block_confignum_init_featuresrr num_classesr r!c stt|ttdtjd|dddddfdt|fdtjd d fd tj ddd d fg|_ |}t |D]|\} } t | |||||d} |j d| d | || |}| t|d krrt||dd} |j d| d | |d}qr|j dt|t|||_|D]r} t| tjr||}tj|dd}t|d}t|d}||}|S)NTr")r r r )rbrQrgZadaptive_avg_pool2dr>flattenrz)r7rrboutr:r:r;rKs     zDenseNet.forward)rjrkrprqrrrF) rSrTrU__doc__rVrr6rWr+rrKrXr:r:r8r;rs&<)modelweightsprogressr!cCsjtd}|j|d}t|D]8}||}|r"|d|d}||||<||=q"||dS)Nz]^(.*denselayer\d+\.(?:norm|relu|conv))\.((?:[12])\.(?:weight|bias|running_mean|running_var))$)rr r )recompileZget_state_dictlistkeysmatchgroupZload_state_dict)rrrpatternZ state_dictkeyresZnew_keyr:r:r;_load_state_dicts   r)rrsrtrrkwargsr!cKsH|dk rt|dt|jdt|||f|}|dk rDt|||d|S)Nru categories)rrr)rrymetarr)rrsrtrrrrr:r:r; _densenets r)rz*https://github.com/pytorch/vision/pull/116z'These weights are ported from LuaTorch.)Zmin_sizerZrecipeZ_docsc @s:eZdZedeeddedddddid d ZeZd S) rz`_. Args: weights (:class:`~torchvision.models.DenseNet121_Weights`, optional): The pretrained weights to use. See :class:`~torchvision.models.DenseNet121_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.densenet.DenseNet`` base class. Please refer to the `source code `_ for more details about this class. .. autoclass:: torchvision.models.DenseNet121_Weights :members: rjrkrp)rverifyrrrrr:r:r;rTs cKst|}tddd||f|S)a{Densenet-161 model from `Densely Connected Convolutional Networks `_. Args: weights (:class:`~torchvision.models.DenseNet161_Weights`, optional): The pretrained weights to use. See :class:`~torchvision.models.DenseNet161_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.densenet.DenseNet`` base class. Please refer to the `source code `_ for more details about this class. .. autoclass:: torchvision.models.DenseNet161_Weights :members: 0)rlrm$rn`)rrrrr:r:r;rms cKst|}tddd||f|S)a{Densenet-169 model from `Densely Connected Convolutional Networks `_. Args: weights (:class:`~torchvision.models.DenseNet169_Weights`, optional): The pretrained weights to use. See :class:`~torchvision.models.DenseNet169_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.densenet.DenseNet`` base class. Please refer to the `source code `_ for more details about this class. .. autoclass:: torchvision.models.DenseNet169_Weights :members: rj)rlrmrjrjrp)rrrrr:r:r;rs cKst|}tddd||f|S)a{Densenet-201 model from `Densely Connected Convolutional Networks `_. Args: weights (:class:`~torchvision.models.DenseNet201_Weights`, optional): The pretrained weights to use. See :class:`~torchvision.models.DenseNet201_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.densenet.DenseNet`` base class. Please refer to the `source code `_ for more details about this class. .. autoclass:: torchvision.models.DenseNet201_Weights :members: rj)rlrmrrjrp)rrrrr:r:r;rs ) _ModelURLs)rrrr)9r collectionsr functoolsrtypingrrrrr>Ztorch.nnr,Ztorch.nn.functionalZ functionalrQZtorch.utils.checkpointutilsrHrGrZtransforms._presetsr r Z_apir rZ_metar_utilsrr__all__ModulerZ ModuleDictrYrwrerrWrrVrrrrrrrrrrrrrZ model_urlsr:r:r:r;sz        A U """"