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 // See docs in ../ops/image_ops.cc. 17 18 #define USE_EIGEN_TENSOR 19 #define EIGEN_USE_THREADS 20 21 #include "tensorflow/core/kernels/extract_image_patches_op.h" 22 #include <vector> 23 #include "tensorflow/core/framework/numeric_op.h" 24 #include "tensorflow/core/framework/op_kernel.h" 25 #include "tensorflow/core/framework/register_types.h" 26 #include "tensorflow/core/framework/tensor.h" 27 #include "tensorflow/core/kernels/bounds_check.h" 28 #include "tensorflow/core/kernels/ops_util.h" 29 #include "tensorflow/core/lib/core/errors.h" 30 #include "tensorflow/core/platform/logging.h" 31 #include "tensorflow/core/platform/macros.h" 32 #include "tensorflow/core/util/tensor_format.h" 33 34 namespace tensorflow { 35 36 typedef Eigen::ThreadPoolDevice CPUDevice; 37 typedef Eigen::GpuDevice GPUDevice; 38 39 static inline void ParseAttributeVec4(OpKernelConstruction* context, 40 const string& attr_name, 41 std::vector<int32>* attr) { 42 OP_REQUIRES_OK(context, context->GetAttr(attr_name, attr)); 43 OP_REQUIRES( 44 context, (*attr)[0] == 1 && (*attr)[3] == 1, 45 errors::Unimplemented("Only support ", attr_name, " across space.")); 46 OP_REQUIRES(context, (*attr)[1] >= 1 && (*attr)[2] >= 1, 47 errors::OutOfRange(attr_name, " is out of range.")); 48 } 49 50 template <typename Device, typename T> 51 class ExtractImagePatchesOp : public UnaryOp<T> { 52 public: 53 explicit ExtractImagePatchesOp(OpKernelConstruction* context) 54 : UnaryOp<T>(context) { 55 ParseAttributeVec4(context, "ksizes", &ksizes_); 56 ParseAttributeVec4(context, "strides", &strides_); 57 ParseAttributeVec4(context, "rates", &rates_); 58 OP_REQUIRES_OK(context, context->GetAttr("padding", &padding_)); 59 } 60 61 void Compute(OpKernelContext* context) override { 62 // Input tensor is of the following dimensions: 63 // [ batch, in_rows, in_cols, channels ] 64 const Tensor& input = context->input(0); 65 OP_REQUIRES(context, input.dims() == 4, 66 errors::InvalidArgument("input must be 4-dimensional", 67 input.shape().DebugString())); 68 69 const int batch = input.dim_size(0); 70 const int in_rows = input.dim_size(1); 71 const int in_cols = input.dim_size(2); 72 const int depth = input.dim_size(3); 73 74 const int ksize_rows = ksizes_[1]; 75 const int ksize_cols = ksizes_[2]; 76 77 const int stride_rows = strides_[1]; 78 const int stride_cols = strides_[2]; 79 80 const int rate_rows = rates_[1]; 81 const int rate_cols = rates_[2]; 82 83 const int ksize_rows_eff = ksize_rows + (ksize_rows - 1) * (rate_rows - 1); 84 const int ksize_cols_eff = ksize_cols + (ksize_cols - 1) * (rate_cols - 1); 85 86 int64 out_rows = 0, out_cols = 0; 87 int64 pad_rows = 0, pad_cols = 0; 88 OP_REQUIRES_OK(context, 89 GetWindowedOutputSize(in_rows, ksize_rows_eff, stride_rows, 90 padding_, &out_rows, &pad_rows)); 91 OP_REQUIRES_OK(context, 92 GetWindowedOutputSize(in_cols, ksize_cols_eff, stride_cols, 93 padding_, &out_cols, &pad_cols)); 94 95 const std::vector<int64> out_sizes = {batch, out_rows, out_cols, 96 ksize_rows * ksize_cols * depth}; 97 TensorShape out_shape(out_sizes); 98 99 Tensor* output = nullptr; 100 OP_REQUIRES_OK(context, context->allocate_output(0, out_shape, &output)); 101 102 // If there is nothing to compute, return. 103 if (out_shape.num_elements() == 0) { 104 return; 105 } 106 107 functor::ExtractImagePatchesForward<Device, T>()( 108 context->eigen_device<Device>(), input.tensor<T, 4>(), ksize_rows, 109 ksize_cols, stride_rows, stride_cols, rate_rows, rate_cols, 110 BrainPadding2EigenPadding(padding_), output->tensor<T, 4>()); 111 } 112 113 private: 114 std::vector<int32> ksizes_; 115 std::vector<int32> strides_; 116 std::vector<int32> rates_; 117 118 Padding padding_; 119 120 TF_DISALLOW_COPY_AND_ASSIGN(ExtractImagePatchesOp); 121 }; 122 123 // Registration of the CPU implementations. 124 #define REGISTER(T) \ 125 REGISTER_KERNEL_BUILDER( \ 126 Name("ExtractImagePatches").Device(DEVICE_CPU).TypeConstraint<T>("T"), \ 127 ExtractImagePatchesOp<CPUDevice, T>); 128 129 TF_CALL_REAL_NUMBER_TYPES(REGISTER); 130 131 #undef REGISTER 132 133 #if GOOGLE_CUDA 134 135 // Forward declarations of the functor specializations for GPU. 136 namespace functor { 137 138 #define DECLARE_GPU_SPEC(T) \ 139 template <> \ 140 void ExtractImagePatchesForward<GPUDevice, T>::operator()( \ 141 const GPUDevice& d, typename TTypes<T, 4>::ConstTensor input, \ 142 int patch_rows, int patch_cols, int stride_rows, int stride_cols, \ 143 int rate_rows, int rate_cols, const Eigen::PaddingType& padding, \ 144 typename TTypes<T, 4>::Tensor output); \ 145 extern template struct ExtractImagePatchesForward<GPUDevice, T>; 146 147 TF_CALL_GPU_NUMBER_TYPES(DECLARE_GPU_SPEC); 148 149 #undef DECLARE_GPU_SPEC 150 151 } // namespace functor 152 153 // Registration of the GPU implementations. 154 #define REGISTER(T) \ 155 REGISTER_KERNEL_BUILDER( \ 156 Name("ExtractImagePatches").Device(DEVICE_GPU).TypeConstraint<T>("T"), \ 157 ExtractImagePatchesOp<GPUDevice, T>); 158 159 TF_CALL_GPU_NUMBER_TYPES(REGISTER); 160 161 #undef REGISTER 162 163 #endif // GOOGLE_CUDA 164 165 } // namespace tensorflow 166
