xref: /aosp_15_r20/external/tensorflow/tensorflow/core/kernels/sparse_softmax_op.cc (revision b6fb3261f9314811a0f4371741dbb8839866f948)
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 // See docs in ../ops/sparse_ops.cc.
17 
18 #define EIGEN_USE_THREADS
19 
20 #include <numeric>
21 
22 #include "third_party/eigen3/unsupported/Eigen/CXX11/Tensor"
23 #include "tensorflow/core/framework/op_kernel.h"
24 #include "tensorflow/core/framework/op_requires.h"
25 #include "tensorflow/core/framework/register_types.h"
26 #include "tensorflow/core/framework/tensor.h"
27 #include "tensorflow/core/framework/tensor_util.h"
28 #include "tensorflow/core/framework/types.h"
29 #include "tensorflow/core/util/sparse/sparse_tensor.h"
30 
31 using tensorflow::gtl::ArraySlice;
32 using tensorflow::sparse::SparseTensor;
33 
34 namespace tensorflow {
35 
36 using CPUDevice = Eigen::ThreadPoolDevice;
37 
38 template <typename Device, typename T>
39 class SparseSoftmaxOp : public OpKernel {
40  public:
SparseSoftmaxOp(OpKernelConstruction * context)41   explicit SparseSoftmaxOp(OpKernelConstruction *context) : OpKernel(context) {}
42 
Compute(OpKernelContext * context)43   void Compute(OpKernelContext *context) override {
44     const Tensor *indices_t, *values_t, *shape_t;
45     OP_REQUIRES_OK(context, context->input("sp_indices", &indices_t));
46     OP_REQUIRES_OK(context, context->input("sp_values", &values_t));
47     OP_REQUIRES_OK(context, context->input("sp_shape", &shape_t));
48 
49     // Validations.
50     OP_REQUIRES(context, TensorShapeUtils::IsMatrix(indices_t->shape()),
51                 errors::InvalidArgument(
52                     "Input sp_indices should be a matrix but received shape: ",
53                     indices_t->shape().DebugString()));
54     OP_REQUIRES(context,
55                 TensorShapeUtils::IsVector(values_t->shape()) &&
56                     TensorShapeUtils::IsVector(shape_t->shape()),
57                 errors::InvalidArgument(
58                     "Inputs sp_values and sp_shape should be vectors "
59                     "but received shapes: ",
60                     values_t->shape().DebugString(), " and ",
61                     shape_t->shape().DebugString()));
62     OP_REQUIRES(context, shape_t->NumElements() >= 2,
63                 errors::InvalidArgument(
64                     "Input should have rank >= 2, but received shape: ",
65                     shape_t->SummarizeValue(3)));
66     TensorShape shape;
67     OP_REQUIRES_OK(context, TensorShape::BuildTensorShape(
68                                 shape_t->flat<int64_t>(), &shape));
69 
70     const int64_t nnz = indices_t->dim_size(0);
71     const int rank = static_cast<int>(indices_t->dim_size(1));
72     SparseTensor st;
73     OP_REQUIRES_OK(
74         context, SparseTensor::Create(tensor::DeepCopy(*indices_t),
75                                       tensor::DeepCopy(*values_t), shape, &st));
76 
77     Tensor *output_values = nullptr;
78     OP_REQUIRES_OK(context, context->allocate_output(0, TensorShape({nnz}),
79                                                      &output_values));
80     typename TTypes<T>::Flat output_flat = output_values->flat<T>();
81 
82     Tensor tmp_t;
83     OP_REQUIRES_OK(context, context->allocate_temp(DataTypeToEnum<T>::value,
84                                                    TensorShape({}), &tmp_t));
85     typename TTypes<T>::Scalar tmp_scalar = tmp_t.scalar<T>();
86 
87     gtl::InlinedVector<int64_t, 4> dims(rank);
88     std::iota(dims.begin(), dims.end(), 0);
89     // { 0, ..., rank-1 }.
90     const ArraySlice<int64_t> kReorderDims(dims);
91     // All but the last dim -- the class dimension to be max-reduced along.
92     const ArraySlice<int64_t> kGroupByDims = kReorderDims.subspan(0, rank - 1);
93     st.Reorder<T>(kReorderDims);
94     int count = 0;
95 
96     // The SparseTensor has logical shape [..., b, c], where the
97     // innermost size-"c" dimension is the class dimension to be max-reduced.
98     // Therefore we group by the first (rank - 1) dimensions.
99     const Device &device = context->eigen_device<Device>();
100     for (const auto &g : st.group(kGroupByDims)) {
101       const auto group_vals = g.values<T>();
102       const int group_size = group_vals.size();
103 
104       // Shifts by max, exponentiates, then renormalizes.
105       tmp_scalar.device(context->eigen_device<Device>()) = group_vals.maximum();
106       const T group_max = tmp_scalar();
107 
108       Eigen::Tensor<T, 1, Eigen::RowMajor> tmp(group_size);
109       tmp.device(device) = (group_vals - tmp.constant(group_max)).exp();
110 
111       tmp_scalar.device(device) = tmp.sum().inverse();
112       tmp.device(device) = tmp * tmp.constant(tmp_scalar());
113 
114       // Assigns back to output[count, count + group_size).
115       Eigen::TensorMap<Eigen::Tensor<T, 1, Eigen::RowMajor>> output_part(
116           output_flat.data() + count, group_size);
117       output_part.device(device) = tmp;
118 
119       count += group_size;
120     }
121   }
122 };
123 
124 #define REGISTER_KERNEL(T)                                             \
125   REGISTER_KERNEL_BUILDER(                                             \
126       Name("SparseSoftmax").Device(DEVICE_CPU).TypeConstraint<T>("T"), \
127       SparseSoftmaxOp<CPUDevice, T>)
128 
129 REGISTER_KERNEL(float);
130 REGISTER_KERNEL(double);
131 #undef REGISTER_KERNEL
132 
133 }  // namespace tensorflow
134