1 /* Copyright 2015 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 #ifndef TENSORFLOW_KERNELS_ADJUST_CONTRAST_OP_H_ 17 #define TENSORFLOW_KERNELS_ADJUST_CONTRAST_OP_H_ 18 #include "third_party/eigen3/unsupported/Eigen/CXX11/Tensor" 19 #include "tensorflow/core/framework/tensor_types.h" 20 21 namespace tensorflow { 22 namespace functor { 23 24 // Functor used by AdjustContrastOp to do the computations. 25 template <typename Device, typename T> 26 struct AdjustContrast { 27 void operator()(const Device& d, typename TTypes<T, 4>::ConstTensor input, 28 typename TTypes<float>::ConstScalar contrast_factor, 29 typename TTypes<float>::ConstScalar min_value, 30 typename TTypes<float>::ConstScalar max_value, 31 typename TTypes<float, 4>::Tensor mean_values, 32 typename TTypes<float, 4>::Tensor output) { 33 const int batch = input.dimension(0); 34 const int height = input.dimension(1); 35 const int width = input.dimension(2); 36 const int channels = input.dimension(3); 37 38 Eigen::array<int, 4> scalar_broadcast; 39 scalar_broadcast[0] = batch; 40 scalar_broadcast[1] = height; 41 scalar_broadcast[2] = width; 42 scalar_broadcast[3] = channels; 43 #if !defined(EIGEN_HAS_INDEX_LIST) 44 Eigen::array<int, 2> reduction_axis; 45 reduction_axis[0] = 1; 46 reduction_axis[1] = 2; 47 Eigen::array<int, 4> broadcast_dims; 48 broadcast_dims[0] = 1; 49 broadcast_dims[1] = height; 50 broadcast_dims[2] = width; 51 broadcast_dims[3] = 1; 52 Eigen::Tensor<int, 4>::Dimensions reshape_dims; 53 reshape_dims[0] = batch; 54 reshape_dims[1] = 1; 55 reshape_dims[2] = 1; 56 reshape_dims[3] = channels; 57 #else 58 Eigen::IndexList<Eigen::type2index<1>, Eigen::type2index<2> > 59 reduction_axis; 60 Eigen::IndexList<Eigen::type2index<1>, int, int, Eigen::type2index<1> > 61 broadcast_dims; 62 broadcast_dims.set(1, height); 63 broadcast_dims.set(2, width); 64 Eigen::IndexList<int, Eigen::type2index<1>, Eigen::type2index<1>, int> 65 reshape_dims; 66 reshape_dims.set(0, batch); 67 reshape_dims.set(3, channels); 68 #endif 69 Eigen::Sizes<1, 1, 1, 1> scalar; 70 float num_reduced_coeffs = height * width; 71 mean_values.device(d) = 72 (input.template cast<float>().sum(reduction_axis).eval() / 73 num_reduced_coeffs) 74 .reshape(reshape_dims) 75 .broadcast(broadcast_dims); 76 77 auto contrast_factor_tensor = 78 contrast_factor.reshape(scalar).broadcast(scalar_broadcast); 79 auto adjusted = 80 (input.template cast<float>() - mean_values) * contrast_factor_tensor + 81 mean_values; 82 auto min_bcast = min_value.reshape(scalar).broadcast(scalar_broadcast); 83 auto max_bcast = max_value.reshape(scalar).broadcast(scalar_broadcast); 84 // TODO(wicke): This is rather slow and should be re-written as pure cuda. 85 output.device(d) = adjusted.cwiseMin(max_bcast).cwiseMax(min_bcast); 86 } 87 }; 88 89 // Functor used by AdjustContrastOpv2 to do the computations. 90 template <typename Device> 91 struct AdjustContrastv2 { 92 void operator()(const Device& d, typename TTypes<float, 4>::ConstTensor input, 93 typename TTypes<float>::ConstScalar contrast_factor, 94 typename TTypes<float, 4>::Tensor output) { 95 const int batch = input.dimension(0); 96 const int height = input.dimension(1); 97 const int width = input.dimension(2); 98 const int channels = input.dimension(3); 99 100 Eigen::array<int, 4> scalar_broadcast; 101 scalar_broadcast[0] = batch; 102 scalar_broadcast[1] = height; 103 scalar_broadcast[2] = width; 104 scalar_broadcast[3] = channels; 105 #if !defined(EIGEN_HAS_INDEX_LIST) 106 Eigen::array<int, 2> reduction_axis; 107 reduction_axis[0] = 0; 108 reduction_axis[1] = 1; 109 Eigen::array<int, 4> broadcast_dims; 110 broadcast_dims[0] = 1; 111 broadcast_dims[1] = height; 112 broadcast_dims[2] = width; 113 broadcast_dims[3] = 1; 114 Eigen::Tensor<int, 4>::Dimensions reshape_dims; 115 reshape_dims[0] = batch; 116 reshape_dims[1] = 1; 117 reshape_dims[2] = 1; 118 reshape_dims[3] = channels; 119 Eigen::array<int, 4> reduced_dims_first; 120 reduced_dims_first[0] = 1; 121 reduced_dims_first[1] = 2; 122 reduced_dims_first[2] = 0; 123 reduced_dims_first[3] = 3; 124 #else 125 Eigen::IndexList<Eigen::type2index<0>, Eigen::type2index<1> > 126 reduction_axis; 127 Eigen::IndexList<Eigen::type2index<1>, int, int, Eigen::type2index<1> > 128 broadcast_dims; 129 broadcast_dims.set(1, height); 130 broadcast_dims.set(2, width); 131 Eigen::IndexList<int, Eigen::type2index<1>, Eigen::type2index<1>, int> 132 reshape_dims; 133 reshape_dims.set(0, batch); 134 reshape_dims.set(3, channels); 135 Eigen::IndexList<Eigen::type2index<1>, Eigen::type2index<2>, 136 Eigen::type2index<0>, Eigen::type2index<3> > 137 reduced_dims_first; 138 #endif 139 Eigen::Sizes<1, 1, 1, 1> scalar; 140 float num_reduced_coeffs = height * width; 141 output.device(d) = 142 (input.shuffle(reduced_dims_first).sum(reduction_axis).eval() / 143 num_reduced_coeffs) 144 .reshape(reshape_dims) 145 .broadcast(broadcast_dims); 146 auto contrast_factor_tensor = 147 contrast_factor.reshape(scalar).broadcast(scalar_broadcast); 148 auto adjusted = (input - output) * contrast_factor_tensor; 149 output.device(d) += adjusted; 150 } 151 }; 152 153 } // namespace functor 154 } // namespace tensorflow 155 156 #endif // TENSORFLOW_KERNELS_ADJUST_CONTRAST_OP_H_ 157
