a Sic@sdZddlZddlmZddlmZddlmZddlmZddlmm Z ddl m Z ddl m Z dd lmZdd lmZdd lmZdd lmZdd lmZdZdddddddddddddddddd d!d"d#d$d%Zd&gd&d'ggd(d)gd)gd)d*gd*gd+Zd,d-d.d/d0d.d)d1d.d2d-d.gd,d-d.d/d0d.d)d3d.d2d-d.gd,d-d.d/d4d.d)d5d.d2d-d.gd,d-d.d/d6d.d)d5d.d2d-d.gd,d0d.d/d7d.d)d8d.d2d0d.gd,d0d.d/d5d.d)d9d.d2d0d.gd,d0d.d/d:d.d)d;d.d2d0d.gd+Zdd?d@dAZdBZddCdDZdeeedHdIdJZ de!ee!dMdNdOZ"de!e!e#eeeeedQdRdSZ$deeeeedTdUdVZ%dWdXZ&e!dYdZd[Z'd\d]Z(de!eeeee!feeefdddedfZ)edgdhddidjZ*edkdlddmdnZ+edodpddqdrZ,edsdtddudvZ-edwdxddydzZ.ed{d|dd}d~Z/edddddZ0eddddZ1eddddZ2ej2je2_ej3djde*_ej3dnde+_ej3drde,_ej3dvde-_ej3dzde._ej3d~de/_ej3dde0_dS)zResNet-RS models for Keras. Reference: - [Revisiting ResNets: Improved Training and Scaling Strategies]( https://arxiv.org/pdf/2103.07579.pdf) N)Callable)Dict)List)Union)backend)layers)imagenet_utils)training) data_utils) layer_utils) keras_exportzGhttps://storage.googleapis.com/tensorflow/keras-applications/resnet_rs/Z 544b3434d00efc199d66e9058c7f3379Z 82d5b90c5ce9d710da639d6216d0f979Z eb285be29ab42cf4835ff20a5e3b5d23Z f9a0f6b85faa9c3db2b6e233c4eebb5bZ 8d72a301ed8a6f11a47c4ced4396e338Z 5fbf7ac2155cb4d5a6180ee9e3aa8704Z 31a46a92ab21b84193d0d71dd8c3d03bZ dc8b2cba2005552eafa3167f00dc2133Z ba6271b99bdeb4e7a9b15c05964ef4adZ fa79794252dbe47c89130f65349d654aZ a76930b741884e09ce90fa7450747d5fZ bbdb3994718dfc0d1cd45d7eff3f3d9cZ 20d575825ba26176b03cb51012a367a8Z 2c42ecb22e35f3e23d2f70babce0a2aaZ f4a039dc3c421321b7fc240494574a68Z 6e44b55025bbdff8f51692a023143d66Z 7ccb858cc738305e8ceb3c0140bee393Z ab0c1f9079d2f85a9facbd2c88aa6079Z ae0eb9bed39e64fc8d7e0db4018dc7e8Z fe6217c32be8305b1889657172b98884Z 69d9d925319f00a8bdd4af23c04e4102Z 90daa68cd26c95aa6c5d25451e095529)zresnet-rs-101-i160.h5zresnet-rs-101-i160_notop.h5zresnet-rs-101-i192.h5zresnet-rs-101-i192_notop.h5zresnet-rs-152-i192.h5zresnet-rs-152-i192_notop.h5zresnet-rs-152-i224.h5zresnet-rs-152-i224_notop.h5zresnet-rs-152-i256.h5zresnet-rs-152-i256_notop.h5zresnet-rs-200-i256.h5zresnet-rs-200-i256_notop.h5zresnet-rs-270-i256.h5zresnet-rs-270-i256_notop.h5zresnet-rs-350-i256.h5zresnet-rs-350-i256_notop.h5zresnet-rs-350-i320.h5zresnet-rs-350-i320_notop.h5zresnet-rs-420-i320.h5zresnet-rs-420-i320_notop.h5zresnet-rs-50-i160.h5zresnet-rs-50-i160_notop.h5)rri@)2e^@) input_filters num_repeatsi$5H,WVarianceScalingg@fan_outtruncated_normal)scalemode distribution) class_nameconfiga(Instantiates the {name} architecture. Reference: [Revisiting ResNets: Improved Training and Scaling Strategies]( https://arxiv.org/pdf/2103.07579.pdf) For image classification use cases, see [this page for detailed examples]( https://keras.io/api/applications/#usage-examples-for-image-classification-models). For transfer learning use cases, make sure to read the [guide to transfer learning & fine-tuning]( https://keras.io/guides/transfer_learning/). Note: each Keras Application expects a specific kind of input preprocessing. For ResNetRs, by default input preprocessing is included as a part of the model (as a `Rescaling` layer), and thus `tf.keras.applications.resnet_rs.preprocess_input` is actually a pass-through function. In this use case, ResNetRS models expect their inputs to be float tensors of pixels with values in the [0-255] range. At the same time, preprocessing as a part of the model (i.e. `Rescaling` layer) can be disabled by setting `include_preprocessing` argument to False. With preprocessing disabled ResNetRS models expect their inputs to be float tensors of pixels with values in the [-1, 1] range. Args: depth: Depth of ResNet network. input_shape: optional shape tuple. It should have exactly 3 inputs channels, and width and height should be no smaller than 32. E.g. (200, 200, 3) would be one valid value. bn_momentum: Momentum parameter for Batch Normalization layers. bn_epsilon: Epsilon parameter for Batch Normalization layers. activation: activation function. se_ratio: Squeeze and Excitation layer ratio. dropout_rate: dropout rate before final classifier layer. drop_connect_rate: dropout rate at skip connections. include_top: whether to include the fully-connected layer at the top of the network. block_args: list of dicts, parameters to construct block modules. model_name: name of the model. pooling: optional pooling mode for feature extraction when `include_top` is `False`. - `None` means that the output of the model will be the 4D tensor output of the last convolutional layer. - `avg` means that global average pooling will be applied to the output of the last convolutional layer, and thus the output of the model will be a 2D tensor. - `max` means that global max pooling will be applied. weights: one of `None` (random initialization), `'imagenet'` (pre-training on ImageNet), or the path to the weights file to be loaded. Note: one model can have multiple imagenet variants depending on input shape it was trained with. For input_shape 224x224 pass `imagenet-i224` as argument. By default, highest input shape weights are downloaded. input_tensor: optional Keras tensor (i.e. output of `layers.Input()`) to use as image input for the model. classes: optional number of classes to classify images into, only to be specified if `include_top` is True, and if no `weights` argument is specified. classifier_activation: A `str` or callable. The activation function to use on the "top" layer. Ignored unless `include_top=True`. Set `classifier_activation=None` to return the logits of the "top" layer. include_preprocessing: Boolean, whether to include the preprocessing layer (`Rescaling`) at the bottom of the network. Defaults to `True`. Note: Input image is normalized by ImageNet mean and standard deviation. Returns: A `keras.Model` instance. cs2durtd}d|fdd}|S)z Conv2D block with fixed padding.NZconv_c s:dkrt|}tjdkr(dnddtd|S)NsamevalidF)filters kernel_sizestridespaddinguse_biaskernel_initializername) fixed_paddingrConv2DCONV_KERNEL_INITIALIZER)inputsr3r4r9r5X/var/www/html/django/DPS/env/lib/python3.9/site-packages/keras/applications/resnet_rs.pyapplys z!Conv2DFixedPadding..applyrget_uid)r3r4r5r9counterrAr?r>r@Conv2DFixedPaddings    rEh㈵>relu bn_momentum bn_epsilon activationcs2durtd}d|fdd}|S)zResNet-D type STEM block.NZstem_cs8tdkrdnd}tddddd|}tj|dd |}tjd d |}tdddd d|}tj|d d |}tjdd |}tddddd|}tj|dd |}tjdd |}tddddd|}tj|dd |}tjdd |}|S)N channels_lastrr0 Z _stem_conv_1r3r4r5r9Z_stem_batch_norm_1axismomentumepsilonr9Z _stem_act_1r9Z _stem_conv_2Z_stem_batch_norm_2Z _stem_act_2rZ _stem_conv_3Z_stem_batch_norm_3Z _stem_act_3Z _stem_conv_4Z_stem_batch_norm_4Z _stem_act_4)rimage_data_formatrErBatchNormalization Activation)r=bn_axisxrLrKrJr9r?r@rAst    zSTEM..applyrB)rJrKrLr9rDrAr?r[r@STEMs   4r\?r0) in_filtersse_ratio expand_ratiocsHtdkrdnddur0td}d|fdd}|S)zSqueeze and Excitation block.rMrr0NZse_c stjdd|}dkr.|jdddf}ndd|jdf}tj|dd|}tdtd}tj|ddgddgtddd d d |}tjdddgddgtddd d d |}tj||gddS)NZ _se_squeezerUr0Z _se_reshaperr1TrHZ _se_reduce)r3r4r5r8r6r7rLr9sigmoidZ _se_expandZ _se_excite) rGlobalAveragePooling2DshapeReshapemaxintr;r<multiply)r=rZse_shapeZnum_reduced_filtersrYr`r^r9r_r?r@rA1sF    zSE..apply)rrVrC)r^r_r`r9rDrAr?rjr@SE(s   $rk皙?)r3r5use_projectionrJrKrLr_survival_probabilityc s<durtd} d| f dd} | S)z7Bottleneck block variant for residual networks with BN.NZblock_0_cstdkrdnd}|}rd}dkr`tjddddd |}t|ddd d |}nt|dd d |}tj|d d |}tdddd |}tj|dd |}tjdd|}tddd |}tj|dd |}tjdd|}tddddd |}tj|dd |}dkrpdkrnntdd|}rtjddd|}t ||g}tjdd|S)NrMrr0rrO)rOrOr1Z_projection_pooling) pool_sizer5r6r9Z_projection_convrPZ_projection_batch_normrQZ_conv_1Z batch_norm_1Z_act_1rUZ_conv_2Z _batch_norm_2Z_act_2Z_conv_3Z _batch_norm_3r_se)r_r9)Nr0r0r0Z_drop) noise_shaper9Z _output_act) rrVrAveragePooling2DrErWrXrkDropoutAdd)r=rYshortcut filters_outrZ rLrKrJr3r9r_r5rnrmr?r@rAhs zBottleneckBlock..applyrB) r3r5rmrJrKrLr_rnr9rDrAr?rwr@BottleneckBlockXs   Xrx)r_rKrJrLrnc s<durtd} d| f dd} | S)z0Create one group of blocks for the ResNet model.NZ block_group_cs`tddd |}tdD].}tddd|dd |}q,|S) NTZ _block_0_) r3r5rmr_rKrJrLrnr9r0F_block__) r3r5rmr_rLrKrJrnr9)rxrange)r=rZi rLrKrJr3r9rr_r5rnr?r@rAs8    zBlockGroup..applyrB) r3r5rr_rKrJrLrnr9rDrAr?r}r@ BlockGroups   r~cCs|t||S)z@Get survival probability based on block number and initial rate.)floatZ init_rate block_numZ total_blocksr?r?r@get_survival_probabilitysr) target_limitcCst}||krt|dS)z9Increase default recursion limit to create larger models.N)sysgetrecursionlimitsetrecursionlimit)rZ current_limitr?r?r@allow_bigger_recursionsrcCs8|d}|d}||}tj||f||ffd|}|S)zKPad the input along the spatial dimensions independently of input size.r0rO)r6)r ZeroPadding2D)r=r4Z pad_totalZpad_begpad_end padded_inputsr?r?r@r:sr:皙?T resnet-rsimagenetsoftmax)depthrL block_argsclassifier_activationc" Cst|}| dkr&t|}| d|} dd|D}| h|dvsltjj| sltd|d|d| | |vr|r|d krtd |d |tj|d d t || d}| durt j |d}nt | st j | |d}n| }t dkrdnd}|}|rB||d}t jdd|}|dkrBt jgdgd|d|}t|||d|}| durft|} t| D]d\}}t||dt| dd}t|d||dkrdnd|d||||d|dd d! |}qn|r(t jd"d#|}|dkrt j|d$d#|}t|| t j||d%d&|}n6| d'krDt jd"d#|}n| d(kr^t jd)d#|}| durtt| }n|}tj||| d#}| |vr| d*d+}| d*|}|s|d,7}|d-}t |} t!j"|| d.t#|d/}!|$|!n| dur|$| |S)0a" Build Resnet-RS model, given provided parameters. Args: depth: Depth of ResNet network. input_shape: optional shape tuple. It should have exactly 3 inputs channels, and width and height should be no smaller than 32. E.g. (200, 200, 3) would be one valid value. bn_momentum: Momentum parameter for Batch Normalization layers. bn_epsilon: Epsilon parameter for Batch Normalization layers. activation: activation function. se_ratio: Squeeze and Excitation layer ratio. dropout_rate: dropout rate before final classifier layer. drop_connect_rate: dropout rate at skip connections. include_top: whether to include the fully-connected layer at the top of the network. block_args: list of dicts, parameters to construct block modules. model_name: name of the model. pooling: optional pooling mode for feature extraction when `include_top` is `False`. - `None` means that the output of the model will be the 4D tensor output of the last convolutional layer. - `avg` means that global average pooling will be applied to the output of the last convolutional layer, and thus the output of the model will be a 2D tensor. - `max` means that global max pooling will be applied. weights: one of `None` (random initialization), `'imagenet'` (pre-training on ImageNet), or the path to the weights file to be loaded. Note- one model can have multiple imagenet variants depending on input shape it was trained with. For input_shape 224x224 pass `imagenet-i224` as argument. By default, highest input shape weights are downloaded. input_tensor: optional Keras tensor (i.e. output of `layers.Input()`) to use as image input for the model. classes: optional number of classes to classify images into, only to be specified if `include_top` is True, and if no `weights` argument is specified. classifier_activation: A `str` or callable. The activation function to use on the "top" layer. Ignored unless `include_top=True`. Set `classifier_activation=None` to return the logits of the "top" layer. include_preprocessing: Boolean, whether to include the preprocessing layer (`Rescaling`) at the bottom of the network. Defaults to `True`. Note- Input image is normalized by ImageNet mean and standard deviation. Returns: A `tf.keras.Model` instance. Raises: ValueError: in case of invalid argument for `weights`, or invalid input shape. ValueError: if `classifier_activation` is not `softmax` or `None` when using a pretrained top layer. rz-icSsg|]}d|qS)z imagenet-ir?).0rZr?r?r@ ^zResNetRS..NzThe `weights` argument should be either `None` (random initialization), `'imagenet'` (pre-training on ImageNet, with highest available input shape), or the path to the weights file to be loaded. For ResNetRSz- the following weight variants are available z% (default=highest). Received weights=rz-If using `weights` as `'imagenet'` or any of zH with `include_top` as true, `classes` should be 1000. Received classes=rrN) default_sizemin_size data_formatrequire_flattenweights)rd)tensorrdrMrr0gp?)r+)g ףp= ?gv/?gCl?)gg٪?g.?gQ?)meanvariancerRrIrOrrrrr~rz) r3rLr5rr_rJrKrnr9avg_poolrU top_dropout predictions)rLr9avgrfmax_pool-raZ_notopz.h5models)fnameorigin cache_subdir file_hash)%DEPTH_TO_WEIGHT_VARIANTSrftfiogfileexists ValueErrorrobtain_input_shaperrVrInputis_keras_tensor Rescaling Normalizationr\ BLOCK_ARGS enumeraterlenr~rcrsvalidate_activationDenseGlobalMaxPooling2Dr get_source_inputsr ModelsplitBASE_WEIGHTS_URLr get_file WEIGHT_HASHES load_weights)"r input_shaperJrKrLr_ dropout_ratedrop_connect_rate include_topr model_namepoolingr input_tensorclassesrinclude_preprocessingZavailable_weight_variantsZmax_input_shapeZweights_allow_list img_inputrYrZ num_channelsr|argsrnr=modelZweights_input_shapeZ weights_namefilename download_url weights_pathr?r?r@ResNetRSsI                     rz'keras.applications.resnet_rs.ResNetRS50zkeras.applications.ResNetRS50cCs td|dd||||||d|d S)zBuild ResNet-RS50 model.rrFr]z resnet-rs-50 rrrrrrrrrrrrrrrrrrrrrr?r?r@ ResNetRS50srz(keras.applications.resnet_rs.ResNetRS101zkeras.applications.ResNetRS101cCs td|dd||||||d|d S)zBuild ResNet-RS101 model.rrFr]z resnet-rs-101rrrr?r?r@ ResNetRS101srz(keras.applications.resnet_rs.ResNetRS152zkeras.applications.ResNetRS152cCs td|dd||||||d|d S)zBuild ResNet-RS152 model.rrFr]z resnet-rs-152rrrr?r?r@ ResNetRS152srz(keras.applications.resnet_rs.ResNetRS200zkeras.applications.ResNetRS200cCs td|dd||||||d|d S)zBuild ResNet-RS200 model.r皙?r]z resnet-rs-200rrrr?r?r@ ResNetRS200:srz(keras.applications.resnet_rs.ResNetRS270zkeras.applications.ResNetRS270cCs(tdtd|dd||||||d|d S)zBuild ResNet-RS-270 model.irrr]z resnet-rs-270rrrrr?r?r@ ResNetRS270Xsrz(keras.applications.resnet_rs.ResNetRS350zkeras.applications.ResNetRS350cCs(tdtd|dd||||||d|d S)zBuild ResNet-RS350 model.irr皙?z resnet-rs-350rrrr?r?r@ ResNetRS350wsrz(keras.applications.resnet_rs.ResNetRS420zkeras.applications.ResNetRS420cCs(tdtd|dd||||||d|d S)zBuild ResNet-RS420 model.irrrz resnet-rs-420) rrrrrrrrrrrrrrr?r?r@ ResNetRS420srz-keras.applications.resnet_rs.preprocess_inputcCs|S)aA placeholder method for backward compatibility. The preprocessing logic has been included in the ResnetRS model implementation. Users are no longer required to call this method to normalize the input data. This method does nothing and only kept as a placeholder to align the API surface between old and new version of model. Args: x: A floating point `numpy.array` or a `tf.Tensor`. data_format: Optional data format of the image tensor/array. Defaults to None, in which case the global setting `tf.keras.backend.image_data_format()` is used (unless you changed it, it defaults to "channels_last").{mode} Returns: Unchanged `numpy.array` or `tf.Tensor`. r?)rZrr?r?r@preprocess_inputsrz/keras.applications.resnet_rs.decode_predictionscCstj||dS)N)top)rdecode_predictions)predsrr?r?r@rsrrU)N)rFrGrHN)r]r0N)rFrGrHr]rlN)r]rGrFrHrlN)NrFrGrHr]r]rTNrNrNrrT)TrrNNNrT)TrrNNNrT)TrrNNNrT)TrrNNNrT)TrrNNNrT)TrrNNNrT)TrrNNNrT)N)r)4__doc__rtypingrrrrZtensorflow.compat.v2compatv2rkerasrrZkeras.applicationsr keras.enginer keras.utilsr r tensorflow.python.util.tf_exportr rrrrr<BASE_DOCSTRINGrErstrr\rgrkboolrxr~rrr:rrrrrrrrrrformatr?r?r?r@s            - L  C 4 o 0 S