• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /* Copyright 2016 The TensorFlow Authors. All Rights Reserved.
2 
3 Licensed under the Apache License, Version 2.0 (the "License");
4 you may not use this file except in compliance with the License.
5 You may obtain a copy of the License at
6 
7     http://www.apache.org/licenses/LICENSE-2.0
8 
9 Unless required by applicable law or agreed to in writing, software
10 distributed under the License is distributed on an "AS IS" BASIS,
11 WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12 See the License for the specific language governing permissions and
13 limitations under the License.
14 ==============================================================================*/
15 
16 #include "tensorflow/core/framework/op_kernel.h"
17 #include "tensorflow/core/framework/register_types.h"
18 #include "tensorflow/core/framework/tensor.h"
19 #include "tensorflow/core/framework/tensor_util.h"
20 #include "tensorflow/core/framework/types.h"
21 #include "tensorflow/core/util/sparse/sparse_tensor.h"
22 
23 namespace tensorflow {
24 
25 template <typename T>
26 class SparseAddGradOp : public OpKernel {
27  public:
SparseAddGradOp(OpKernelConstruction * ctx)28   explicit SparseAddGradOp(OpKernelConstruction *ctx) : OpKernel(ctx) {}
29 
Compute(OpKernelContext * ctx)30   void Compute(OpKernelContext *ctx) override {
31     // Gradient for op: SparseAdd(a, b) == sum.
32     const Tensor *backprop_val_grad, *a_indices, *b_indices, *sum_indices;
33     OP_REQUIRES_OK(ctx, ctx->input("backprop_val_grad", &backprop_val_grad));
34     OP_REQUIRES_OK(ctx, ctx->input("a_indices", &a_indices));
35     OP_REQUIRES_OK(ctx, ctx->input("b_indices", &b_indices));
36     OP_REQUIRES_OK(ctx, ctx->input("sum_indices", &sum_indices));
37 
38     OP_REQUIRES(ctx,
39                 TensorShapeUtils::IsMatrix(a_indices->shape()) &&
40                     TensorShapeUtils::IsMatrix(b_indices->shape()) &&
41                     TensorShapeUtils::IsMatrix(sum_indices->shape()),
42                 errors::InvalidArgument(
43                     "Input indices should be matrices but received shapes: ",
44                     a_indices->shape().DebugString(), " and ",
45                     b_indices->shape().DebugString(), " and ",
46                     sum_indices->shape().DebugString()));
47     OP_REQUIRES(
48         ctx, TensorShapeUtils::IsVector(backprop_val_grad->shape()),
49         errors::InvalidArgument(
50             "Input backprop_val_grad should be a vector but received shape: ",
51             backprop_val_grad->shape().DebugString()));
52     OP_REQUIRES(
53         ctx,
54         a_indices->dim_size(1) == b_indices->dim_size(1) &&
55             b_indices->dim_size(1) == sum_indices->dim_size(1),
56         errors::InvalidArgument("The densified operands should have the same "
57                                 "ndims; for A, B, sum got: ",
58                                 a_indices->dim_size(1), b_indices->dim_size(1),
59                                 sum_indices->dim_size(1)));
60     OP_REQUIRES(
61         ctx, backprop_val_grad->NumElements() == sum_indices->dim_size(0),
62         errors::InvalidArgument("# elements of backprop_val_grad and # rows of "
63                                 "sum_indices should match (#nnz of sum): got ",
64                                 backprop_val_grad->NumElements(), " and ",
65                                 sum_indices->dim_size(0)));
66 
67     const int num_dims = a_indices->dim_size(1);
68     const int64 a_nnz = a_indices->dim_size(0);
69     const int64 b_nnz = b_indices->dim_size(0);
70     const int64 sum_nnz = backprop_val_grad->NumElements();
71 
72     const auto a_indices_mat = a_indices->matrix<int64>();
73     const auto b_indices_mat = b_indices->matrix<int64>();
74     const auto sum_indices_mat = sum_indices->matrix<int64>();
75 
76     Tensor *a_val_grad, *b_val_grad;
77     OP_REQUIRES_OK(ctx,
78                    ctx->allocate_output(0, TensorShape({a_nnz}), &a_val_grad));
79     OP_REQUIRES_OK(ctx,
80                    ctx->allocate_output(1, TensorShape({b_nnz}), &b_val_grad));
81 
82     T *a_val_grad_flat = a_val_grad->flat<T>().data();
83     T *b_val_grad_flat = b_val_grad->flat<T>().data();
84     const T *backprop_val_grad_flat = backprop_val_grad->flat<T>().data();
85     memset(a_val_grad_flat, 0, sizeof(T) * a_nnz);
86     memset(b_val_grad_flat, 0, sizeof(T) * b_nnz);
87 
88 #define COMPARE(a_or_b, idx)                                                \
89   switch (sparse::DimComparator::cmp(a_or_b##_indices_mat, sum_indices_mat, \
90                                      idx, k, num_dims)) {                   \
91     case 0:                                                                 \
92       a_or_b##_val_grad_flat[idx] = backprop_val_grad_flat[k];              \
93       ++idx;                                                                \
94       break;                                                                \
95     case -1:                                                                \
96       ++idx;                                                                \
97       a_or_b##_idx_geq = false;                                             \
98       break;                                                                \
99     case 1:                                                                 \
100       break;                                                                \
101   }
102 
103     // Set-intersect the indices; fill in grads for positions in the
104     // intersection.
105     int64 i = 0, j = 0, k = 0;
106     bool a_idx_geq, b_idx_geq;
107     while (i < a_nnz && j < b_nnz && k < sum_nnz) {
108       a_idx_geq = b_idx_geq = true;
109       COMPARE(a, i);
110       COMPARE(b, j);
111       // increment pointer into sum_indices iff both the current A, B indices >=
112       // the current sum index.
113       if (a_idx_geq && b_idx_geq) ++k;
114     }
115 
116     // at most one loop below will run
117     while (i < a_nnz && k < sum_nnz) {
118       a_idx_geq = true;
119       COMPARE(a, i);
120       if (a_idx_geq) ++k;
121     }
122     while (j < b_nnz && k < sum_nnz) {
123       b_idx_geq = true;
124       COMPARE(b, j);
125       if (b_idx_geq) ++k;
126     }
127 #undef COMPARE
128   }
129 };
130 
131 #define REGISTER_KERNELS(type)                                            \
132   REGISTER_KERNEL_BUILDER(                                                \
133       Name("SparseAddGrad").Device(DEVICE_CPU).TypeConstraint<type>("T"), \
134       SparseAddGradOp<type>)
135 
136 // This op should work for any T that SparseAdd is registered with.
137 REGISTER_KERNELS(float);
138 REGISTER_KERNELS(double);
139 REGISTER_KERNELS(int64);
140 REGISTER_KERNELS(int32);
141 REGISTER_KERNELS(int16);
142 REGISTER_KERNELS(int8);
143 REGISTER_KERNELS(complex64);
144 REGISTER_KERNELS(complex128);
145 #undef REGISTER_KERNELS
146 }  // namespace tensorflow
147