AVh'dZddlZddlmZddlmZddlmZddlmZddlm Z ddlm Z dd lm Z dd lm Z d Z ejd d ejfdZejdd ejfdZejdd ejfdZdZdZejdd ejfdZy)z1Gradients for operators defined in random_ops.py.N) constant_op)dtypes)ops) array_ops)clip_ops) gen_array_ops)gen_random_ops)math_opsctjtj|gtjtj |gd}tj ||S)N)dtyperaxis)rconcatonesrint32shapereshape)xnum_dimensions new_shapes Q/home/dcms/DCMS/lib/python3.12/site-packages/tensorflow/python/ops/random_grad.pyadd_leading_unit_dimensionsrsP~~~&fll;q!"$)   1i (( RandomGammaopc|jd}|jd}|jd}tj|g5t j |d}t ||}tj||}dtj||ztj|fcdddS#1swYyxYw)aReturns the gradient of a Gamma sample w.r.t. alpha. The gradient is computed using implicit differentiation (Figurnov et al., 2018). Args: op: A `RandomGamma` operation. We assume that the inputs to the operation are `shape` and `alpha` tensors, and the output is the `sample` tensor. grad: The incoming gradient `dloss / dsample` of the same shape as `op.outputs[0]`. Returns: A `Tensor` with derivatives `dloss / dalpha`. References: Implicit Reparameterization Gradients: [Figurnov et al., 2018] (http://papers.nips.cc/paper/7326-implicit-reparameterization-gradients) ([pdf] (http://papers.nips.cc/paper/7326-implicit-reparameterization-gradients.pdf)) rNr ) inputsoutputsrcontrol_dependenciesrrrr random_gamma_gradr reduce_sumrange)rgradralphasamplenum_sample_dimensionsalpha_broadcastable partial_as r_RandomGammaGradr*$s. ))A,% ))A,% ::a=& ' G&OOE2155 $&001DfMI (%% yx~~.CDF G G G Gs A)B77CStatelessRandomGammaV2c |jd}|jd}|jd}tj|g5ddt ||||fcdddS#1swYyxYw)aReturns the gradient of a Gamma sample w.r.t. alpha. The gradient is computed using implicit differentiation (Figurnov et al., 2018). Args: op: A `StatelessRandomGamma` operation. We assume that the inputs to the operation are `shape`, `seed` and `alpha` tensors, and the output is the `sample` tensor. grad: The incoming gradient `dloss / dsample` of the same shape as `op.outputs[0]`. Returns: A `Tensor` with derivatives `dloss / dalpha`. References: Implicit Reparameterization Gradients: [Figurnov et al., 2018] (http://papers.nips.cc/paper/7326-implicit-reparameterization-gradients) ([pdf] (http://papers.nips.cc/paper/7326-implicit-reparameterization-gradients.pdf)) rNrrrr _StatelessGammaGradAlpharr$rr%r&s r_StatelessRandomGammaV2Gradr1Lsj0 ))A,% ))A,% ::a=& 'N $0vtL MNNNs AA(StatelessRandomGammaV3c |jd}|jd}|jd}tj|g5ddddt ||||fcdddS#1swYyxYw)aReturns the gradient of a Gamma sample w.r.t. alpha. The gradient is computed using implicit differentiation (Figurnov et al., 2018). Args: op: A `StatelessRandomGamma` operation. We assume that the inputs to the operation are `shape`, `key`, `counter`, `alg`, and `alpha` tensors, and the output is the `sample` tensor. grad: The incoming gradient `dloss / dsample` of the same shape as `op.outputs[0]`. Returns: A `Tensor` with derivatives `dloss / dalpha`. References: Implicit Reparameterization Gradients: [Figurnov et al., 2018] (http://papers.nips.cc/paper/7326-implicit-reparameterization-gradients) ([pdf] (http://papers.nips.cc/paper/7326-implicit-reparameterization-gradients.pdf)) rNr.r0s r_StatelessRandomGammaV3Gradr5lsq0 ))A,% ))A,% ::a=& 'B $d $UE64 @ BBBBs A!!A*ctj|dtj|z }t||}t j ||}t j||zt j|S)z/Returns gradients of a gamma sampler wrt alpha.rr ) rrrankrr r!r r"r#)rr%r&r$r'r(r)s rr/r/sw$//%03innU6KK4E4IK../BFK)    YX^^,AB DDrc tjdtjdz|jd}||z}t j |}tj||kdt j|ztj|dkDdt j|z t j|}d|zS)zNormal distribution function.?g@ half_sqrt_2)r nameg?g) rconstantnpsqrtr r absrwhereerferfc)rr:wzys r_NdtrrFs$$ BGGBKqww]<++o!ll1o!oo+o8<<?oo b&"x}}Q''q)9;>:+678G    +EG &J??6*L??6*L??6*L77L"O77o'O77o'Ou  ??AJ##D5LzBI%%dWn:FK%%dWn:FK%%dWn:FK44%HAu66'JAw66'JAw66'JAw!!IEDA:OI##KgEK##KgEK##KgEK $ ; [ IGCJCJCJs 1N%P  P))__doc__numpyr=tensorflow.python.frameworkrrrtensorflow.python.opsrrrr r rRegisterGradient Operationr*r1r5r/rFrsrrr{s83.++*/0*)m$$G$G%$GN./NCMMN0N>./BCMMB0B@ D =>aJ3==aJ?aJr