giuros01 | 1887081 | 2018-09-13 09:31:40 +0100 | [diff] [blame] | 1 | /* |
| 2 | * Copyright (c) 2018 ARM Limited. |
| 3 | * |
| 4 | * SPDX-License-Identifier: MIT |
| 5 | * |
| 6 | * Permission is hereby granted, free of charge, to any person obtaining a copy |
| 7 | * of this software and associated documentation files (the "Software"), to |
| 8 | * deal in the Software without restriction, including without limitation the |
| 9 | * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or |
| 10 | * sell copies of the Software, and to permit persons to whom the Software is |
| 11 | * furnished to do so, subject to the following conditions: |
| 12 | * |
| 13 | * The above copyright notice and this permission notice shall be included in all |
| 14 | * copies or substantial portions of the Software. |
| 15 | * |
| 16 | * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR |
| 17 | * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, |
| 18 | * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE |
| 19 | * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER |
| 20 | * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, |
| 21 | * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE |
| 22 | * SOFTWARE. |
| 23 | */ |
| 24 | #include "ROIAlignLayer.h" |
| 25 | |
| 26 | #include "arm_compute/core/Types.h" |
| 27 | #include "arm_compute/core/utils/misc/ShapeCalculator.h" |
| 28 | #include "tests/validation/Helpers.h" |
| 29 | |
| 30 | #include <algorithm> |
| 31 | |
| 32 | namespace arm_compute |
| 33 | { |
| 34 | namespace test |
| 35 | { |
| 36 | namespace validation |
| 37 | { |
| 38 | namespace reference |
| 39 | { |
| 40 | namespace |
| 41 | { |
| 42 | /** Average pooling over an aligned window */ |
| 43 | template <typename T> |
| 44 | inline T roi_align_1x1(const T *input, TensorShape input_shape, |
| 45 | float region_start_x, |
| 46 | float bin_size_x, |
| 47 | int grid_size_x, |
| 48 | float region_end_x, |
| 49 | float region_start_y, |
| 50 | float bin_size_y, |
| 51 | int grid_size_y, |
| 52 | float region_end_y, |
| 53 | int pz) |
| 54 | { |
| 55 | if((region_end_x <= region_start_x) || (region_end_y <= region_start_y)) |
| 56 | { |
| 57 | return T(0); |
| 58 | } |
| 59 | else |
| 60 | { |
| 61 | float avg = 0; |
| 62 | // Iterate through the aligned pooling region |
| 63 | for(int iy = 0; iy < grid_size_y; ++iy) |
| 64 | { |
| 65 | for(int ix = 0; ix < grid_size_x; ++ix) |
| 66 | { |
| 67 | // Align the window in the middle of every bin |
| 68 | float y = region_start_y + (iy + 0.5) * bin_size_y / float(grid_size_y); |
| 69 | float x = region_start_x + (ix + 0.5) * bin_size_x / float(grid_size_x); |
| 70 | |
| 71 | // Interpolation in the [0,0] [0,1] [1,0] [1,1] square |
| 72 | const int y_low = y; |
| 73 | const int x_low = x; |
| 74 | const int y_high = y_low + 1; |
| 75 | const int x_high = x_low + 1; |
| 76 | |
| 77 | const float ly = y - y_low; |
| 78 | const float lx = x - x_low; |
| 79 | const float hy = 1. - ly; |
| 80 | const float hx = 1. - lx; |
| 81 | |
| 82 | const float w1 = hy * hx; |
| 83 | const float w2 = hy * lx; |
| 84 | const float w3 = ly * hx; |
| 85 | const float w4 = ly * lx; |
| 86 | |
| 87 | const size_t idx1 = coord2index(input_shape, Coordinates(x_low, y_low, pz)); |
| 88 | T data1 = input[idx1]; |
| 89 | |
| 90 | const size_t idx2 = coord2index(input_shape, Coordinates(x_high, y_low, pz)); |
| 91 | T data2 = input[idx2]; |
| 92 | |
| 93 | const size_t idx3 = coord2index(input_shape, Coordinates(x_low, y_high, pz)); |
| 94 | T data3 = input[idx3]; |
| 95 | |
| 96 | const size_t idx4 = coord2index(input_shape, Coordinates(x_high, y_high, pz)); |
| 97 | T data4 = input[idx4]; |
| 98 | |
| 99 | avg += w1 * data1 + w2 * data2 + w3 * data3 + w4 * data4; |
| 100 | } |
| 101 | } |
| 102 | |
| 103 | avg /= grid_size_x * grid_size_y; |
| 104 | |
| 105 | return T(avg); |
| 106 | } |
| 107 | } |
| 108 | |
| 109 | /** Clamp the value between lower and upper */ |
| 110 | template <typename T> |
| 111 | T clamp(T value, T lower, T upper) |
| 112 | { |
| 113 | return std::max(lower, std::min(value, upper)); |
| 114 | } |
| 115 | } // namespace |
| 116 | template <typename T> |
| 117 | SimpleTensor<T> roi_align_layer(const SimpleTensor<T> &src, const std::vector<ROI> &rois, const ROIPoolingLayerInfo &pool_info) |
| 118 | { |
| 119 | const size_t num_rois = rois.size(); |
| 120 | DataType dst_data_type = src.data_type(); |
| 121 | |
| 122 | TensorShape input_shape = src.shape(); |
| 123 | TensorShape output_shape(pool_info.pooled_width(), pool_info.pooled_height(), src.shape()[2], num_rois); |
| 124 | SimpleTensor<T> dst(output_shape, dst_data_type); |
| 125 | |
| 126 | // Iterate over every pixel of the input image |
| 127 | for(size_t px = 0; px < pool_info.pooled_width(); px++) |
| 128 | { |
| 129 | for(size_t py = 0; py < pool_info.pooled_height(); py++) |
| 130 | { |
| 131 | for(size_t pw = 0; pw < num_rois; pw++) |
| 132 | { |
| 133 | ROI roi = rois[pw]; |
| 134 | const int roi_batch = roi.batch_idx; |
| 135 | |
| 136 | const float roi_anchor_x = roi.rect.x * pool_info.spatial_scale(); |
| 137 | const float roi_anchor_y = roi.rect.y * pool_info.spatial_scale(); |
| 138 | const float roi_dims_x = std::max(roi.rect.width * pool_info.spatial_scale(), 1.0f); |
| 139 | const float roi_dims_y = std::max(roi.rect.height * pool_info.spatial_scale(), 1.0f); |
| 140 | ; |
| 141 | |
| 142 | float bin_size_x = roi_dims_x / pool_info.pooled_width(); |
| 143 | float bin_size_y = roi_dims_y / pool_info.pooled_height(); |
| 144 | float region_start_x = px * bin_size_x + roi_anchor_x; |
| 145 | float region_start_y = py * bin_size_y + roi_anchor_y; |
| 146 | float region_end_x = (px + 1) * bin_size_x + roi_anchor_x; |
| 147 | float region_end_y = (py + 1) * bin_size_y + roi_anchor_y; |
| 148 | |
| 149 | region_start_x = clamp(region_start_x, 0.0f, float(input_shape[0])); |
| 150 | region_start_y = clamp(region_start_y, 0.0f, float(input_shape[1])); |
| 151 | region_end_x = clamp(region_end_x, 0.0f, float(input_shape[0])); |
| 152 | region_end_y = clamp(region_end_y, 0.0f, float(input_shape[1])); |
| 153 | |
| 154 | const int roi_bin_grid_x = (pool_info.sampling_ratio() > 0) ? pool_info.sampling_ratio() : int(ceil(bin_size_x)); |
| 155 | const int roi_bin_grid_y = (pool_info.sampling_ratio() > 0) ? pool_info.sampling_ratio() : int(ceil(bin_size_y)); |
| 156 | |
| 157 | // Move input and output pointer across the fourth dimension |
| 158 | const size_t input_stride_w = input_shape[0] * input_shape[1] * input_shape[2]; |
| 159 | const size_t output_stride_w = output_shape[0] * output_shape[1] * output_shape[2]; |
| 160 | const T *input_ptr = src.data() + roi_batch * input_stride_w; |
| 161 | T *output_ptr = dst.data() + px + py * output_shape[0] + pw * output_stride_w; |
| 162 | |
| 163 | for(int pz = 0; pz < int(input_shape[2]); ++pz) |
| 164 | { |
| 165 | // For every pixel pool over an aligned region |
| 166 | *(output_ptr + pz * output_shape[0] * output_shape[1]) = roi_align_1x1(input_ptr, input_shape, |
| 167 | region_start_x, |
| 168 | bin_size_x, |
| 169 | roi_bin_grid_x, |
| 170 | region_end_x, |
| 171 | region_start_y, |
| 172 | bin_size_y, |
| 173 | roi_bin_grid_y, |
| 174 | region_end_y, pz); |
| 175 | } |
| 176 | } |
| 177 | } |
| 178 | } |
| 179 | return dst; |
| 180 | } |
| 181 | template SimpleTensor<float> roi_align_layer(const SimpleTensor<float> &src, const std::vector<ROI> &rois, const ROIPoolingLayerInfo &pool_info); |
| 182 | template SimpleTensor<half> roi_align_layer(const SimpleTensor<half> &src, const std::vector<ROI> &rois, const ROIPoolingLayerInfo &pool_info); |
| 183 | } // namespace reference |
| 184 | } // namespace validation |
| 185 | } // namespace test |
| 186 | } // namespace arm_compute |