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/register_types.h" 25 #include "tensorflow/core/framework/tensor.h" 26 #include "tensorflow/core/framework/tensor_util.h" 27 #include "tensorflow/core/framework/types.h" 28 #include "tensorflow/core/util/sparse/sparse_tensor.h" 29 30 using tensorflow::gtl::ArraySlice; 31 using tensorflow::sparse::SparseTensor; 32 33 namespace tensorflow { 34 35 using CPUDevice = Eigen::ThreadPoolDevice; 36 37 template <typename Device, typename T> 38 class SparseSoftmaxOp : public OpKernel { 39 public: SparseSoftmaxOp(OpKernelConstruction * context)40 explicit SparseSoftmaxOp(OpKernelConstruction *context) : OpKernel(context) {} 41 Compute(OpKernelContext * context)42 void Compute(OpKernelContext *context) override { 43 const Tensor *indices_t, *values_t, *shape_t; 44 OP_REQUIRES_OK(context, context->input("sp_indices", &indices_t)); 45 OP_REQUIRES_OK(context, context->input("sp_values", &values_t)); 46 OP_REQUIRES_OK(context, context->input("sp_shape", &shape_t)); 47 48 // Validations. 49 OP_REQUIRES(context, TensorShapeUtils::IsMatrix(indices_t->shape()), 50 errors::InvalidArgument( 51 "Input sp_indices should be a matrix but received shape: ", 52 indices_t->shape().DebugString())); 53 OP_REQUIRES(context, 54 TensorShapeUtils::IsVector(values_t->shape()) && 55 TensorShapeUtils::IsVector(shape_t->shape()), 56 errors::InvalidArgument( 57 "Inputs sp_values and sp_shape should be vectors " 58 "but received shapes: ", 59 values_t->shape().DebugString(), " and ", 60 shape_t->shape().DebugString())); 61 OP_REQUIRES(context, shape_t->NumElements() >= 2, 62 errors::InvalidArgument( 63 "Input should have rank >= 2, but received shape: ", 64 shape_t->SummarizeValue(3))); 65 OP_REQUIRES(context, 66 indices_t->dim_size(0) < std::numeric_limits<int>::max(), 67 errors::InvalidArgument( 68 "Number of non-zero elements exceeds int32 range")); 69 70 const int nnz = static_cast<int>(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( 75 tensor::DeepCopy(*indices_t), tensor::DeepCopy(*values_t), 76 TensorShape(shape_t->flat<int64>()), &st)); 77 78 Tensor *output_values = nullptr; 79 OP_REQUIRES_OK(context, context->allocate_output(0, TensorShape({nnz}), 80 &output_values)); 81 typename TTypes<T>::Flat output_flat = output_values->flat<T>(); 82 83 Tensor tmp_t; 84 OP_REQUIRES_OK(context, context->allocate_temp(DataTypeToEnum<T>::value, 85 TensorShape({}), &tmp_t)); 86 typename TTypes<T>::Scalar tmp_scalar = tmp_t.scalar<T>(); 87 88 gtl::InlinedVector<int64, 4> dims(rank); 89 std::iota(dims.begin(), dims.end(), 0); 90 // { 0, ..., rank-1 }. 91 const ArraySlice<int64> kReorderDims(dims); 92 // All but the last dim -- the class dimension to be max-reduced along. 93 const ArraySlice<int64> kGroupByDims = kReorderDims.subspan(0, rank - 1); 94 st.Reorder<T>(kReorderDims); 95 int count = 0; 96 97 // The SparseTensor has logical shape [..., b, c], where the 98 // innermost size-"c" dimension is the class dimension to be max-reduced. 99 // Therefore we group by the first (rank - 1) dimensions. 100 const Device &device = context->eigen_device<Device>(); 101 for (const auto &g : st.group(kGroupByDims)) { 102 const auto group_vals = g.values<T>(); 103 const int group_size = group_vals.size(); 104 105 // Shifts by max, exponentiates, then renormalizes. 106 tmp_scalar.device(context->eigen_device<Device>()) = group_vals.maximum(); 107 const T group_max = tmp_scalar(); 108 109 Eigen::Tensor<T, 1, Eigen::RowMajor> tmp(group_size); 110 tmp.device(device) = (group_vals - tmp.constant(group_max)).exp(); 111 112 tmp_scalar.device(device) = tmp.sum().inverse(); 113 tmp.device(device) = tmp * tmp.constant(tmp_scalar()); 114 115 // Assigns back to output[count, count + group_size). 116 Eigen::TensorMap<Eigen::Tensor<T, 1, Eigen::RowMajor>> output_part( 117 output_flat.data() + count, group_size); 118 output_part.device(device) = tmp; 119 120 count += group_size; 121 } 122 } 123 }; 124 125 #define REGISTER_KERNEL(T) \ 126 REGISTER_KERNEL_BUILDER( \ 127 Name("SparseSoftmax").Device(DEVICE_CPU).TypeConstraint<T>("T"), \ 128 SparseSoftmaxOp<CPUDevice, T>) 129 130 REGISTER_KERNEL(float); 131 REGISTER_KERNEL(double); 132 #undef REGISTER_KERNEL 133 134 } // namespace tensorflow 135