1 /* Copyright 2019 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_CORE_KERNELS_CONV_GRAD_SHAPE_UTILS_H_ 17 #define TENSORFLOW_CORE_KERNELS_CONV_GRAD_SHAPE_UTILS_H_ 18 19 #include <vector> 20 21 #include "tensorflow/core/framework/tensor_shape.h" 22 #include "tensorflow/core/lib/core/stringpiece.h" 23 #include "tensorflow/core/util/padding.h" 24 #include "tensorflow/core/util/tensor_format.h" 25 26 namespace tensorflow { 27 // Information about a single spatial dimension for a convolution 28 // backpropagation. 29 struct ConvBackpropSpatialDimension { 30 int64_t input_size; 31 int64_t filter_size; 32 int64_t output_size; 33 int64_t stride; 34 int64_t dilation; 35 36 // Output size after scaling by the stride. 37 int64_t expanded_output_size; 38 39 // Number of padding elements to be added before/after this dimension of 40 // the input when computing Conv?DBackpropInput. 41 int64_t pad_before, pad_after; 42 }; 43 44 // Computed dimensions for a backwards convolution. 45 struct ConvBackpropDimensions { 46 // Information about each spatial dimension. 47 gtl::InlinedVector<ConvBackpropSpatialDimension, 3> spatial_dims; 48 49 // Batch size. 50 int64_t batch_size; 51 52 // Input and output feature depth. 53 int64_t in_depth, out_depth; 54 55 // Convenience access methods for spatial dimensions properties. input_sizeConvBackpropDimensions56 int64_t input_size(int dim) const { return spatial_dims[dim].input_size; } filter_sizeConvBackpropDimensions57 int64_t filter_size(int dim) const { return spatial_dims[dim].filter_size; } output_sizeConvBackpropDimensions58 int64_t output_size(int dim) const { return spatial_dims[dim].output_size; } strideConvBackpropDimensions59 int64_t stride(int dim) const { return spatial_dims[dim].stride; } dilationConvBackpropDimensions60 int64_t dilation(int dim) const { return spatial_dims[dim].dilation; } 61 62 // Compute padding for the given spatial dimension. 63 int SpatialPadding(const Padding& padding, int dim) const; 64 }; 65 66 // Common code between implementations of Conv?DBackpropInput and 67 // Conv?DBackpropFilter. Verifies that the dimensions all match, and computes 68 // sizes/padding for the spatial dimensions. Does not support explicit padding. 69 Status ConvBackpropComputeDimensions(StringPiece label, int num_spatial_dims, 70 const TensorShape& input_shape, 71 const TensorShape& filter_shape, 72 const TensorShape& out_backprop_shape, 73 const std::vector<int32>& strides, 74 Padding padding, TensorFormat data_format, 75 ConvBackpropDimensions* dims); 76 77 // The V2 version computes the same outputs with arbitrary dilation rate and 78 // supports explicit padding. 79 // TODO(b/67112639): Merge V2 versions and the original versions eventually. 80 Status ConvBackpropComputeDimensionsV2( 81 StringPiece label, int num_spatial_dims, const TensorShape& input_shape, 82 const TensorShape& filter_shape, const TensorShape& out_backprop_shape, 83 const gtl::ArraySlice<int32>& dilations, const std::vector<int32>& strides, 84 Padding padding, absl::Span<const int64_t> explicit_paddings, 85 TensorFormat data_format, ConvBackpropDimensions* dims); 86 87 // Computes the shape of the in_backprop. 88 Status Conv2DBackpropComputeInputShape(const Tensor& input_sizes, 89 const TensorShape& filter_shape, 90 const TensorShape& out_backprop_shape, 91 const TensorFormat& data_format, 92 TensorShape* input_shape); 93 } // namespace tensorflow 94 95 #endif // TENSORFLOW_CORE_KERNELS_CONV_GRAD_SHAPE_UTILS_H_ 96