AVhdZddlZddlZddlZddlmZddlmZddlmZddlm Z ddlm Z ddlm Z ejd d ejfd Zejd d ejfd Zejdd ejfdZejdd ejfdZejdd ejfdZejdd ejfdZejdd ejfdZejdd ejfdZejdd ejfdZejdd ejfdZejdd ejfdZejd d ejfd!Zejd"d ejfd#Zejd$d ejfd%Zejd&d ejfd'Zejd(d ejfd)Zejd*d+ejfd,Zejd-d ejfd.Zejd/d ejfd0Z ejd1d ejfd2Z!ejd3d ejfd4Z"ejd5d ejfd6Z#ejd7d ejfd8Z$ejd9d ejfd:Z%ejd;d ejfd<Z&ejd=d ejfd>Z'ejd?d ejfd@Z(ejdAd ejfdBZ)ejdCd ejfdDZ*ejdEd ejfdFZ+dGZ,ejdHd ejfdIZ-ejdJd ejfdKZ.ejdLd ejfdMZ/ejdNd ejfdOZ0ejdPd ejfdQZ1ejdRd ejfdSZ2ejdTd ejfdUZ3ejdVd ejfdWZ4ejdXd ejfdYZ5ejdZd ejfd[Z6ejd\d ejfd]Z7ejd^d_Z8ejd`d ejfdaZ9ejdbd ejfdcZ:ejddd ejfdeZ;ejdfd ejfdgZ<ejdhd ejfdiZ=d ejfdjZ>ejdkd ejfdlZ?ejdmd ejfdnZ@ejdod ejfdpZAejdqd ejfdrZBejdsejdtd ejfduZCejdvd ejfdwZDejdxd ejfdyZEdzZFejd{d ejfd|ZGy)}z-Gradients for operators defined in nn_ops.py.N)dtypes)ops) array_ops)array_ops_stack) gen_nn_ops)math_opsConv2DBackpropInputopcdtj|tj|jd|jd|j d|j d|j d|j d|j d|j d j tj||jd|j d|j d|j d|j d|j d|j d j gS) The derivatives for deconvolution. Args: op: the Deconvolution op. grad: the tensor representing the gradient w.r.t. the output Returns: the gradients w.r.t. the input and the filter N dilationsstridespaddingexplicit_paddingsuse_cudnn_on_gpu data_formatrrrrrr)rconv2d_backprop_filterrshapeinputsget_attrdecodeconv2dr grads M/home/dcms/DCMS/lib/python3.12/site-packages/tensorflow/python/ops/nn_grad.py_Conv2DBackpropInputGradrs ''  //"))A, ' ))A,KK ,++i(++i(KK(;<;;'9:kk-0779 ;  ))A,KK ,++i(++i(KK(;<;;'9:kk-0779; Conv2DBackpropFilterctjtj|jd||jd|j d|j d|j d|j d|j d|j dj dtj|jd||j d|j d|j d|j d|j d|j dj gS) Nrrrrrrrrr)rconv2d_backprop_inputrrrrrrrs r_Conv2DBackpropFilterGradr$Bs && //"))A, '  ))A,KK ,++i(++i(KK(;<;;'9:kk-0779 ;=A ))A, KK ,++i(++i(KK(;<;;'9:kk-0779; r "DepthwiseConv2dNativeBackpropInputc dtj|tj|jd|jd|j d|j d|j d|j d|j d tj ||jd|j d|j d|j d|j d|j d gS) r Nr rrrrrrrrrrr)r'depthwise_conv2d_native_backprop_filterrrrrdepthwise_conv2d_nativers r'_DepthwiseConv2dNativeBackpropInputGradr*]s 88  //"))A, ' ))A,KK ,++i(++i(KK(;<kk-02((  ))A,KK ,++i(++i(KK(;<kk-02 r #DepthwiseConv2dNativeBackpropFilterc tjtj|jd||jd|j d|j d|j d|j d|j ddtj |jd||j d|j d|j d|j d|j dgS) Nrrrrrrrr')r&depthwise_conv2d_native_backprop_inputrrrrr)rs r(_DepthwiseConv2dNativeBackpropFilterGradr.~s77 //"))A, '  ))A,KK ,++i(++i(KK(;<kk-0248(( ))A, KK ,++i(++i(KK(;<kk-02 r Conv3Dc |jdj}tj|jd|jdg\}}t j ||jd||jd|jd|jd|t j|jd|||jd|jd|jd|gS)Nrrr rrrrrrr)rrrshape_nrrconv3d_backprop_input_v2conv3d_backprop_filter_v2)r rrshape_0shape_1s r _Conv3DGradr7s M*113+&& ! biil'CD'7))  ))A, KK ,++i(++i(!#** ))A,  KK ,++i(++i(! r Conv3DBackpropInputV2c |jdj}dtj|t j |j d|j d|jd|jd|jd|tj||j d|jd|jd|jd|gS)Nrr rrrrr1)rrrr4rrrconv3dr rrs r_Conv3DBackpropInputGradr<s M*113+ **  //"))A, ' ))A,KK ,++i(++i(!#  ))A,KK ,++i(++i(! # r Conv3DBackpropFilterV2c |jdj}tjt j |j d||j d|jd|jd|jd|dtj|j d||jd|jd|jd|gS)Nrrrrrrr1)rrrr3rrrr:r;s r_Conv3DBackpropFilterGradr?s M*113+)) //"))A, '  ))A,KK ,++i(++i(!#%) ))A, KK ,++i(++i(! # r AvgPool3Dc  tjtj|jd||j d|j d|j d|j dj SNrksizerrr)rCrrr)ravg_pool3d_gradrrrrrrs r_AvgPool3DGradrEse  # #oobiil# KK kk)$kk)$++m,335  77r AvgPool3DGradc tj|jdtj||j d|j d|j d|j dj fSNrrCrrrr)r stop_gradientrr avg_pool3drrrs r_AvgPool3DGradGradrLsn  ! !"))A, /   kk'"kk)$kk)$++m4;;=  ? @@r MaxPool3Dc tj|jd|jd||j d|j d|j d|j dj SrB)rmax_pool3d_gradroutputsrrrs r_MaxPool3DGradrQsg  # #iiljjm KK kk)$kk)$++m,335 77r MaxPool3DGradc tj|jdtj|jdtj|jd|jd||j d|j d|j d|j dj fSNrr rCrrrrr)r zeros_likerrmax_pool3d_grad_gradrrrs r_MaxPool3DGradGradrXs   ryy| ,   ryy| ,  ) )iiliilkk'"kk)$kk),++m4;;= ? @ @r MaxPool3DGradGradc tj|jdtj|jdtj|jd|jd||j d|j d|j d|j dj fSrT)rrVrrrOrrrs r_MaxPool3DGradGradGradr[ s   ryy| ,   ryy| ,  $ $iiliilkk'"kk)$kk),++m4;;= ? @ @r Softmaxcf|jd}tj||zdd}||z |zS)a The derivative of the softmax nonlinearity. We assume that probs is of shape [batch_size * dim] The formula for dsoftmax / dx = (diag(softmax) - softmax * softmax'). This matrix is diagonal minus a rank one matrix, so it is easy to implement as follows: grad_x = grad_softmax * softmax - sum(grad_softmax * softmax) * softmax Args: op: the Softmax op. grad_softmax: the tensor representing the gradient w.r.t. the softmax output. Returns: gradient w.r.t the input to the softmax rTkeepdims)rPr reduce_sum)r grad_softmaxsoftmax sum_channelss r _SoftmaxGradres:( JJqM'$$\G%;R$O,  % 00r LogSoftmaxctj|jd}|tj|dd|zz S)aThe gradient for log_softmax. log_softmax = input - log(sum(exp(input)) dlog_softmax/dinput = diag - softmax(input) Args: op: The log softmax op. grad: The tensor representing the gradient w.r.t. the output. Returns: The gradients w.r.t. the input. rr^Tr_)rexprPra)r rrcs r_LogSoftmaxGradri2s9 LLA '' ##D"tB-  z 22% KsD++ D:9D: BiasAddV1unused_bias_opctjtj|dz }|tj||fS)aReturn the gradients for the 2 inputs of bias_op. The first input of unused_bias_op is the tensor t, and its gradient is just the gradient the unused_bias_op received. The second input of unused_bias_op is the bias vector which has one fewer dimension than "received_grad" (the batch dimension.) Its gradient is the received gradient Summed on the batch dimension, which is the first dimension. Args: unused_bias_op: The BiasOp for which we need to generate gradients. received_grad: Tensor. The gradients passed to the BiasOp. Returns: Two tensors, the first one for the "tensor" input of the BiasOp, the second one for the "bias" input of the BiasOp. r )rrangerrankra)r~roreduction_dim_tensors r_BiasAddGradV1rsC&" }(E(IJ ,,]-AC DDr RelucHtj||jdSNr)r relu_gradrPrs r _ReluGradr   dBJJqM 22r EluGradc|jd}tj||tj|dk||jdztj |fS)Nr r)rrelu_gradrwhererV)r relu_xs r _EluGradGradrs[ ))A,%   dE * //ai ! ,i.B.B5.I K LLr SeluGradc|jd}tj||tj|dk||jdztj |fS)Nr gr)rr selu_gradrrrV)r rselu_xs r _SeluGradGradrs[ 99Q<&   tV , //rk4"))A,. 0D0DV0L N OOr Relu6cHtj||jdSr)r relu6_gradrPrs r _Relu6Gradrs   tRZZ] 33r Relu6Gradcv|jd}tj||tj|fSNr )rrrrrVr rxs r_Relu6GradGradrs2iil!   a ()*>*>q*A BBr LeakyRelucr|jd}|jd}tj|||S)Nralphar)rrrleaky_relu_gradr rrrs r_LeakyReluGradrs2iil! ++g %  # #D!5 99r LeakyReluGradc|jd}|jd}tj|||t j |fS)Nr rr)rrrrrrVrs r_LeakyReluGradGradrsLiil! ++g %  $ $T1+0 23<3G3G3J LLr ElucHtj||jdSr)rrrPrs r_EluGradrs   T2::a= 11r SelucHtj||jdSr)rrrPrs r _SeluGradrrr SoftpluscL|tj|jdzSr)rsigmoidrrs r _SoftplusGradrs   1. ..r SoftplusGradc|j\}}tj|g5tj||}||zt j | dzt j |zz }||fcdddS#1swYyxYw)Ng@)rrcontrol_dependenciesr softplus_gradrrh)r rdyrddyd2xs r_SoftplusGradGradrsz ))%"a '  " "4 +C )x||QB'#- Q? @C :s AA??BSoftsigncHtj||jdSr)r softsign_gradrrs r _SoftsignGradrs  ! !$ ! 55r ReluGradcv|jd}tj||tj|fSr)rrrrrVrs r _ReluGradGradrs2iil!   tQ ')=)=a)@ AAr c8tj|d}||zS)zMultiply after broadcasting vec to match dimensions of mat. Args: vec: A 1-D tensor of dimension [D0] mat: A 2-D tensor of dimension [D0, D1] Returns: A tensor of dimension [D0, D1], the result of vec * mat r^)r expand_dims)vecmats r _BroadcastMulrs  c2&# sr SoftmaxCrossEntropyWithLogitsc |jd}t||}|jd}|~t|ddsqt j |}||t jtjt j|dt j|ddz |zz }|t|t j| fS)z4Gradient function for SoftmaxCrossEntropyWithLogits.r r_is_zeros_tensorFraxis) rPrrgetattrrrcrsqueezermatmulr log_softmaxr grad_loss grad_grad softmax_gradrlogitsrcs r"_SoftmaxCrossEntropyWithLogitsGradrsA, y, /$ 99Q<& )/ 7  (Gi)++  ! !)Q /  ! !'1 - /  D }Y)?)?)G(GH HHr #SparseSoftmaxCrossEntropyWithLogitsc ^|jd}t||}|jd}|~t|ddsqt j |}||t jtjt j|dt j|ddz |zz }|dfS)z:Gradient function for SparseSoftmaxCrossEntropyWithLogits.r rNrFrr) rPrrrrrcrrrrrrs r(_SparseSoftmaxCrossEntropyWithLogitsGradrsA, y, /$ 99Q<& )/ 7  (Gi)++  ! !)Q /  ! !'1 - /  D tr Conv2Dc |jd}|jd}|jd}|jd}|jd}|jd}tj|jd|jdg\}} t j ||jd||||||| t j |jd| ||||||| gS) zGradient function for Conv2D.rrrrrrrr r)rrr2rrr#r) r rrrrrrrr5r6s r _Conv2DGradr6skk+&) KK "' KK "'kk"56[[!34 M*+&& ! biil'CD'7&&  ))A, -+! #'' ))A,  -+! # r DepthwiseConv2dNativec Vtjtj|jd|jd||j d|j d|j d|j d|j dtj |jdtj|jd||j d|j d|j d|j d|j dgS) Nrr rrrrrr')rr-rrrrr(rs r_DepthwiseConv2dNativeGradr_s77 //"))A, ' ))A, KK ,++i(++i(KK(;<kk-0288 ))A, //"))A, ' KK ,++i(++i(KK(;<kk-02 r Dilation2Dc tj|jd|jd||jd|jd|jdtj|jd|jd||jd|jd|jdgS)Nrr rratesr)rdilation2d_backprop_inputrrdilation2d_backprop_filterrs r_Dilation2DGradrws**299Q<1t+-;;y+A+-;;w+?+-;;y+AC++BIIaL"))A,,.KK ,B,.KK,@,.KK ,BD  r LRNc |jd}|jd}|jd}|jd}tj||jd|jd||||gS)N depth_radiusbiasrbetar)rrlrn_gradrrP)r rrrrrs r_LRNGradrsq^,, V $ ++g % V $$ ! bjjm\4' r AvgPoolc tjtj|jdt j ||jd|jd|jd|jdS)Nr)out_typerCrrrrI)r avg_pool_gradrrrrint32rrs r _AvgPoolGradrsb  ! !oobiilV\\: kk'kk)kk)++m,  ..r AvgPoolGradc tj|jdtj||j d|j d|j d|j dfSrH)rrJrravg_poolrrs r_AvgPoolGradGradrsc  ! !"))A, /   kk'"kk)$kk)$++m4  6 77r MaxPoolc tj|jd|jd||j d|j d|j d|j d|j dS)NrrCrrrr)rrr)r max_pool_gradrrPrrs r _MaxPoolGradrsk  ! !iiljjm kk'kk)kk)$ $78++m, ..r MaxPoolV2c |jd}|jd}tj|jd|jd||||j d|j dddfS)Nr rrrrrU)rrmax_pool_grad_v2rPrr rrCrs r_MaxPoolGradV2rsr ))A,% IIaL'  $ $iiljjm   kk)$++m, .04T ::r MaxPoolWithArgmaxc ~tj|jd||jd|j d|j d|j d|j dS)Nrr rCrrinclude_batch_in_index)rr)rmax_pool_grad_with_argmaxrrPr)r runused_argmax_grads r_MaxPoolGradWithArgmaxrsb  - -iil jjmkk'kk)kk)$[[)AB r MaxPoolGradc ltj|jdtj|jdtj|jd|jd||j d|j d|j d|j dfSrT)rrVrrmax_pool_grad_gradrrs r_MaxPoolGradGradrs299Q<(299Q<(## ))A, ))A,  ++g  ++i ++i(kk-0  r MaxPoolGradV2c p|jd}|jd}tj|jdtj|jdtj|jd|jd||||j d|j dddfS)Nrr rrrU)rrrVrmax_pool_grad_grad_v2rrs r_MaxPoolGradGradV2rs ))A,% IIaL'299Q<(299Q<(&& ))A, ))A,   ++i(kk-0   r MaxPoolGradGradc ltj|jdtj|jdtj|jd|jd||j d|j d|j d|j dfSrT)rrVrrrrrs r_MaxPoolGradGradGradrs299Q<(299Q<( ))A, ))A,  ++g  ++i ++i(kk-0  r FractionalMaxPoolc tj|jd|jd||jd|jd|j dS)aReturns gradient for FractionalMaxPool. Since FractionalMaxPool has three outputs, there are three gradients passed in for each of the outputs. Only the first one is useful, the other two gradients are empty. Args: op: The FractionalMaxPoolOp. grad_0: Gradient with respect to op.outputs[0] unused_grad_1: Gradient with respect to op.outputs[1]/row_seq. It is empty. unused_grad_2: Gradient with respect to op.outputs[2]/col_seq. It is empty. Returns: Input backprop for FractionalMaxPool op. rr r overlapping)rfractional_max_pool_gradrrPrr grad_0 unused_grad_1 unused_grad_2s r_FractionalMaxPoolGradrsR&  , ,iiljjm jjmjjmkk-  r FractionalAvgPoolc tj|jdj||jd|jd|j dS)aReturns gradient for FractionalAvgPool. Since FractionalAvgPool has three outputs, there are three gradients passed in for each of the outputs. Only the first one is useful, the other two gradients are empty. Args: op: The FractionalAvgPoolOp. grad_0: Gradient with respect to op.outputs[0] unused_grad_1: Gradient with respect to op.outputs[1]/row_seq. It is empty. unused_grad_2: Gradient with respect to op.outputs[2]/col_seq. It is empty. Returns: Input backprop for FractionalAvgPool op. rr rr)rfractional_avg_pool_gradr get_shaperPrrs r_FractionalAvgPoolGradr$sO&  , ,RYYq\-C-C-Ev-/ZZ]BJJqM-/[[-G IIr BatchNormWithGlobalNormalizationc tj|jd|jd|jd|jd||jd|jd\}}}}}|||||fS)a$Return the gradients for the 5 inputs of BatchNormWithGlobalNormalization. We do not backprop anything for the mean and var intentionally as they are not being trained with backprop in the operation. Args: op: The BatchNormOp for which we need to generate gradients. grad: Tensor. The gradients passed to the BatchNormOp. Returns: dx: Backprop for input, which is (grad * (g * rsqrt(v + epsilon))) dm: Backprop for mean, which is sum_over_rest(grad * g) * (-1 / rsqrt(v + epsilon)) dv: Backprop for variance, which is sum_over_rest(grad * g * (x - m)) * (-1/2) * (v + epsilon) ^ (-3/2) db: Backprop for beta, which is grad reduced in all except the last dimension. dg: Backprop for gamma, which is (grad * ((x - m) * rsqrt(v + epsilon))) rr rr variance_epsilonscale_after_normalization)r)batch_norm_with_global_normalization_gradrr)r rdxdmdvdbdgs r%_BatchNormWithGlobalNormalizationGradr*<sw*"KKiilBIIaL"))A, ! dkk$%r{{3N'OQ"b"b" RR r c |jd}|d}|jd}|jd}|jd}|jd}|dk(rtj} n&|dk(rtj} ntj } |rK||||j d|j d|||d } |dk(r|j d | d <| di| \} } } }}n|jd}|jd}|d k(r1tj|gd }tj|gd }n5|dk(r0tj|gd}tj|gd}|dvrdnd}||||||||d } |dk(r|j d | d <| di| \} } } }}|d k(rtj| gd} n|dk(rtj| gd} | | | ddfS)aReturn the gradients for the 3 inputs of BatchNorm. Args: op: The BatchNormOp for which we need to compute gradients. version: Integer indicating which version to use of the fused batch norm gradient. *grad: An argument list for tensors of gradients wrt the outputs with grad[0] as grad_y. Returns: grad_x: gradient for x, which is scale * rsqrt(variance + epsilon) * [grad_y - mean(grad_y) - (x - mean(x)) * mean(grad_y * (x - mean(x))) / (variance + epsilon)] in training mode; grad_y * scale * rsqrt(pop_variance + epsilon) in freeze mode. grad_scale: gradient for scale, which is sum(grad_y * (x - mean(x)) * rsqrt(variance + epsilon)) in training mode; sum(grad_y * (x - pop_mean) * rsqrt(pop_variance + epsilon)) in freeze mode. grad_offset: gradient for offset, which is sum(grad_y) in training mode; sum(grad_y) in freeze mode. rr epsilonr is_trainingrr r ) y_backproprscalereserve_space_1reserve_space_2r,rr-reserve_space_3rs)rrr r sNCDHW)rrr r r )rssNHWCNHWCNDHWC)rr r r)rr r rr N) rrrfused_batch_norm_grad_v3fused_batch_norm_grad_v2fused_batch_norm_gradrPr transpose)r versionrrgrad_yr/r,rr-grad_funargsr%dscaledoffset_pop_meanpop_vartarget_data_formats r_BaseFusedBatchNormGradrEWs 2 iil! 7& ))A,% KK "' M*+ M*+ \22H!|22H//H ::a=::a="" D!| " 1 d  ( 04 0BAyy|HiilGg   a .a""6<8f    a 1a""6?;f$/4=%=&BI #")" D!| " 1 d  ( 04 0BAg   r< 0b    r? 3b VWdD ((r FusedBatchNormct|dg|SrrErs r_FusedBatchNormGradrI Q . ..r FusedBatchNormV2ct|dg|SrrHrs r_FusedBatchNormV2GradrMrJr FusedBatchNormV3ct|dg|S)NrrHrs r_FusedBatchNormV3GradrPrJr L2Lossc&|jd|zS)zReturn the gradients for L2Loss. Args: op: The L2LossOp for which we need to generate gradients. grad: Tensor containing a single number. Returns: The gradient, which is (x * grad). r)rrs r _L2LossGradrSs 1 r TopKTopKV2ctj|jd}tj|jd}tjt j |tjtj|dz }tj|jdtjd|g}tjt j |tjtj|dz }tj|d}tj|t j tjt jdt j |tj|z|dtjzdg} tjtj tj| dtj|dgt j"|g|tj$gtjgS)aEReturn the gradients for TopK. Args: op: The TopKOp for which we need to generate gradients. grad: Tensor. The gradients passed to the TopKOp. Returns: A list of two tensors, the first being the gradient w.r.t to the input and TopK, and the second being the gradient w.r.t. to the indices (all zero). rr r^)dtype)rrrrPgatherrcastrint64sizervrstackrrr scatter_nd reduce_prodzeros) r rrAin_shape ind_shape ind_lastdimind_2d in_lastdimouterdiminds r _TopKGradrgs__RYYq\ *(oobjjm,)  mmIv||,nnY!#%+   jjm_**B +<= ?&mmHfll+nnX"$*__V $Q '( x}}   nnQ%]]8V\\BZO')*, .06|| ==@Bd  D#   ##C,i.?.?rd.K##H-. 019;oob -  r ApproxTopKc jdjdgz tjtj }j jd dkr z  fdtjtfdt D }tj|| g}tj||g}tj||jdjS)zReturn the gradients for ApproxTopK. Args: op: The ApproxTopK for which we need to generate gradients. grad: The gradients for backprop. Returns: Scattered gradient based on the top-k indices. r reduction_dimensionrc*|k(r#tjjdS|}tt j ddz}|||<tjt j||}tj|Sr) rrvrPlist itertoolsrepeatrr broadcast_to) diota_len iota_shapeiota idx_shapelifted_idx_shaper r reduction_dims r GetLiftedIdxz)_ApproxTopKGradient..GetLiftedIdxsM   rzz!}.> ??|Hi&&q$(34JJqM   X^^H5z BD  ! !$(8 99r c3.K|] }|yw)Nr6).0rprws r z&_ApproxTopKGradient..s 0q<? 0sr)rPr functoolsreduceoperatormulrrrrtrlrrvr]r) r rrAflat_shape_len lifted_idxflat_idx flat_gradrwrtrurrvs ` @@@@@r_ApproxTopKGradientrs"jjm!!)!_##HLL)<. $++34-Q=(M:: 0E$K 00t=*   zND+A B(~&67)   h 299Q<3E3E FFr NthElementc|jd}|jd}tjtjt j |d||j}t j |d}t j tj|dd}tj|||zdgS)a:Return the gradients for NthElement. Args: op: The NthElementOp for which we need to generate gradients. grad: Tensor. The gradients passed to the NthElementOp Returns: A list of two tensors, the first being the gradient w.r.t. the input, the second being the gradient w.r.t. the N (None). rr^N) rrPrrYequalrrrWradivide)r rinputoutput indicators num_selecteds r_NthElementGradrs ))A,% ::a=& }}nnY**626> L*   tR ($&&x':'::r'JBO, //*l 3d :D AAr c g}ttj||jdtj||jdD]m\}}t j ||t j |dz}|jtjtj||dgotj|dS)a}Replaces each segment with its mean along the last axis. Specifically, each value in the `inputs` tensor gets replaced by the mean value computed from the values that belong to the same segment. Args: inputs: A 2-tensor. Aggregation is done over dimension 1. segments: A 2-tensor, same shape as `input`. Returns: The result, same shape and type as `inputs`. rr ) num_segmentsr^r) ziprsplitrrunsorted_segment_mean reduce_maxappendrvrXrr\)rsegmentsresultinputs_i segments_imeans_is r_MeanAggregatorr9s &!ooffll1o.oohq 124Hh ,,*8+>+>z+JQ+NPG MM)**7J?"FH H   vA ..r IsotonicRegressionc:~|jd}t||S)afGradient for the isotonic regression function. Args: op: The IsotonicRegression tensorflow op. grad_output: Tensor of incoming gradients with respect to the output. grad_segments: Tensor of incoming gradients with respect to the segments. Returns: A tensor, same size as `grad_output` with the gradient with respect to the input. r )rPr)r grad_output grad_segmentsrs r_IsotonicRegressionGradrTs! ZZ]( h //r )H__doc__r{rmr}tensorflow.python.frameworkrrtensorflow.python.opsrrrrRegisterGradient Operationrr$r*r.r7r<r?rErLrQrXr[rerirpr|rrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrr*rErIrMrPrSrgrrrrr6r rrs 4.++1,*+,!!-!H,-#--.4:; <@;<=,hCMM 0-./,./#--0*k"7s}}7#7o&@3==@'@k"7s}}7#7o& @3== @' @)* @s}} @+ @i 1S]]1!10l#G G$G"i DS]]D!D6m$ 3 3% 3Fk"D3==D#D.f3#--33i LS]]L!Lj!OcmmO"Og43==44k"Cs}}C#C k":s}}:#: o&L3==L'Le222f3#--33j!/cmm/"/n% #-- & j!6cmm6"6j!BcmmB"B 56I3==I7I0;<=6h%CMM% %P-.3==/.l#  $ ei .S]].!.m$77%7i  .S]] .! .k" :s}} :# :)* s}} + m$ %  o&3=='('( S]] )  )* +8)*I I+I.89cmm:4O) O)d&'/CMM/(/()/cmm/*/()/cmm/*/h CMM   fh*#--* *Zl#%GCMM%G$%GPl#B B$B4/6*+0 0,0r