o 2hG5@sdZddlmZddlmZddlmZddlmZddlmZddlmZddlm Z e d e d e d d ej fd dZ e dd ej fddZe dd ej fddZe dd ej fddZe dd ej fddZe dd ej fddZe ddd Zd ej fd!d"Ze d#d ej fd$d%Ze d&d ej fd'd(Ze d)d ej fd*d+Ze d,d-ej fd.d/Ze d0d-ej fd1d2Ze d3d ej fd4d5Ze d6d ej fd7d8Zd9S):z1Gradients for operators defined in sparse_ops.py.)dtypes)ops) sparse_tensor) array_ops)gen_sparse_ops)math_ops) sparse_ops SparseAddGrad SparseConcat SparseReorderopc Csb|jd}|jd}t|d}t|}t|||}t|}t |j } dt || dfS)aGradients for the SparseReorder op. Args: op: the SparseReorder op unused_output_indices_grad: the incoming gradients of the output indices output_values_grad: the incoming gradients of the output values Returns: Gradient for each of the 3 input tensors: (input_indices, input_values, input_shape) The gradients for input_indices and input_shape is None. rN) inputsrshaperranger SparseTensorrsparse_reorderinvert_permutationvaluesgather) r unused_output_indices_gradoutput_values_grad input_indices input_shape num_entries entry_indices sp_unordered sp_orderedinverted_permutationr\/var/www/html/chatgem/venv/lib/python3.10/site-packages/tensorflow/python/ops/sparse_grad.py_SparseReorderGrads     r! SparseAddcGst|d}|jd}|jd}|jd}t||||\}}||jd||jdd|dd|ddfS)amThe backward operator for the SparseAdd op. The SparseAdd op calculates A + B, where A, B, and the sum are all represented as `SparseTensor` objects. This op takes in the upstream gradient w.r.t. non-empty values of the sum, and outputs the gradients w.r.t. the non-empty values of A and B. Args: op: the SparseAdd op *grads: the incoming gradients, one element per output of `op` Returns: Gradient for each of the 6 input tensors of SparseAdd: (a_indices, a_values, a_shape, b_indices, b_values, b_shape, thresh) The gradients for the indices, shapes, and the threshold are None. rN)routputsrsparse_add_grad set_shape get_shape)r gradsval_grad a_indices b_indices sum_indices a_val_grad b_val_gradrrr _SparseAddGrad<s   r1SparseTensorDenseAddcCs|jd}dt||d|fS)Nr)rr gather_nd)r out_grad sp_indicesrrr _SparseTensorDenseAddGrad^s r6SparseReduceSumcCs\|jd}|jd}t||jd}t||}|t|tj}dt|||ddfS)z:Similar to gradient for the Sum Op (i.e. tf.reduce_sum()).rr r$N) rr reduced_shaperreshapecastrint64r3)r r4r5sp_shapeoutput_shape_kept_dimsout_grad_reshapedscalerrr _SparseReduceSumGrades   r@ SparseSlicecGsX|d}|jd}|jd}|jd}t||||}||jdd|dddfS)aThe backward operator for the SparseSlice op. This op takes in the upstream gradient w.r.t. non-empty values of the sliced `SparseTensor`, and outputs the gradients w.r.t. the non-empty values of input `SparseTensor`. Args: op: the SparseSlice op *grads: the incoming gradients, one element per output of `op` Returns: Gradient for each of the 5 input tensors of SparseSlice: (indices, values, shape, start, size) The gradients for the indices, shape, start and the size are None. r#rr$N)rr&rsparse_slice_gradr(r))r r*backprop_val_gradr input_startoutput_indicesr+rrr _SparseSliceGradrs   rFSparseTensorDenseMatMulcCs|jdd\}}}|jd}|d}|d}|jj}|jj} || kr0td|d| dtj||||| d} |rDtj| d d } |ddd f} |ddd f} t ||s[| n| } t |sd|nt ||sl| n| }|s|st j t | d t |d d d}n=|r|st j t | dt |dd d}n(|s|rt t | d t |d}n|r|rt j t | dt |d d d d}dtj|d dgdd| fS)aGradients for the dense tensor in the SparseTensorDenseMatMul op. Args: op: the SparseTensorDenseMatMul op grad: the incoming gradient Returns: Gradient for each of the 4 input tensors: (sparse_indices, sparse_values, sparse_shape, dense_tensor) The gradients for indices and shape are None. Raises: TypeError: When the two operands don't have the same type. Nr$ adjoint_a adjoint_bzDSparseTensorDenseMatMul op received operands with different types: `z` and `z`.)rHT) conjugaterr#)rI)rHrI)axis)rget_attrdtype base_dtype TypeErrorrsparse_tensor_dense_mat_mulrmatrix_transposer transposermatmul expand_dimssqueeze)r gradr,a_valuesa_shapebadj_aadj_ba_typeb_typeb_gradrowscolsparts_aparts_b a_values_gradrrr _SparseTensorDenseMatMulGradsb          rfSparseDenseCwiseAddcCtd)Nz4Gradient for SparseDenseCwiseAdd is not implemented.NotImplementedError unused_op unused_gradrrr _SparseDenseCwiseAddGradrncCs |jd}|jd}|jd}tt|tj}tt|t|d}t t |t jj|gd}||} || } t | t dg|gdddg} t || } |ra|| } ||jd} n|| } ||jd t| } tt|t| | |}d| d|fS)z4Common code for SparseDenseCwise{Mul,Div} gradients.rr r$rLr#N)rrr:rrrr;rVsizeconcatonesrslicer3squarer sparse_add zeros_likerr)r rXis_mul x_indicesx_shapeyy_shapenum_added_dimsaugmented_y_shapescalingscaled_indices dense_valsdxdy_valdyrrr _SparseDenseCwiseMulOrDivGrads6      rSparseDenseCwiseMulcC t||dS)z"Gradients for SparseDenseCwiseMul.Trr rXrrr _SparseDenseCwiseMulGrad rSparseDenseCwiseDivcCr)z"Gradients for SparseDenseCwiseDiv.Frrrrr _SparseDenseCwiseDivGradrr SparseSoftmaxc Cs|jd|jd}}|jd}t|||}t|||}t||j|j|}tj|dgdd }t||} | j|j} d| dgS)aGradients for SparseSoftmax. The calculation is the same as SoftmaxGrad: grad_x = grad_softmax * softmax - sum(grad_softmax * softmax) * softmax where we now only operate on the non-zero values present in the SparseTensors. Args: op: the SparseSoftmax op. grad: the upstream gradient w.r.t. the non-zero SparseSoftmax output values. Returns: Gradients w.r.t. the input (sp_indices, sp_values, sp_shape). rr rLT)keepdimsN)rr&rrrrsparse_reduce_sumsparse_dense_cwise_add) r rXindicesrout_vals sp_outputsp_grad sp_product sum_reducedsp_sumgrad_xrrr _SparseSoftmaxGrads     rSparseSparseMaximumrlcCrh)Nz4Gradient for SparseSparseMaximum is not implemented.rirkrrr _SparseSparseMaximumGrad?rorSparseSparseMinimumcCrh)Nz4Gradient for SparseSparseMinimum is not implemented.rirkrrr _SparseSparseMinimumGradFrorSparseFillEmptyRowscCs(|jd}tj||d\}}d|d|gS)z"Gradients for SparseFillEmptyRows.r$)reverse_index_map grad_valuesN)r&rsparse_fill_empty_rows_grad)r unused_grad_output_indicesoutput_grad_valuesunused_grad_empty_row_indicatorunused_grad_reverse_index_maprd_valuesd_default_valuerrr _SparseFillEmptyRowsGradMs  r SparseToDensecCsF|j\}}}}t||}t|t|}t|t|||gS)N)rrr3r reduce_sumrv)r rXsparse_indices output_shape_sparse_values_graddefault_value_gradrrr _SparseToDenseGrad`s   rN)__doc__tensorflow.python.frameworkrrrtensorflow.python.opsrrrrNotDifferentiableRegisterGradient Operationr!r1r6r@rFrfrnrrrrrrrrrrrr sV          ! Y "!