surmeh01 | bceff2f | 2018-03-29 16:29:27 +0100 | [diff] [blame] | 1 | // |
| 2 | // Copyright © 2017 Arm Ltd. All rights reserved. |
| 3 | // See LICENSE file in the project root for full license information. |
| 4 | // |
| 5 | #include "InferenceTestImage.hpp" |
| 6 | #include "MobileNetDatabase.hpp" |
| 7 | |
| 8 | #include <boost/numeric/conversion/cast.hpp> |
| 9 | #include <boost/assert.hpp> |
| 10 | #include <boost/format.hpp> |
| 11 | |
| 12 | #include <iostream> |
| 13 | #include <fcntl.h> |
| 14 | #include <array> |
| 15 | |
| 16 | namespace |
| 17 | { |
| 18 | |
| 19 | inline float Lerp(float a, float b, float w) |
| 20 | { |
| 21 | return w * b + (1.f - w) * a; |
| 22 | } |
| 23 | |
| 24 | inline void PutData(std::vector<float> & data, |
| 25 | const unsigned int width, |
| 26 | const unsigned int x, |
| 27 | const unsigned int y, |
| 28 | const unsigned int c, |
| 29 | float value) |
| 30 | { |
| 31 | data[(3*((y*width)+x)) + c] = value; |
| 32 | } |
| 33 | |
| 34 | std::vector<float> |
| 35 | ResizeBilinearAndNormalize(const InferenceTestImage & image, |
| 36 | const unsigned int outputWidth, |
| 37 | const unsigned int outputHeight) |
| 38 | { |
| 39 | std::vector<float> out; |
| 40 | out.resize(outputWidth * outputHeight * 3); |
| 41 | |
| 42 | // We follow the definition of TensorFlow and AndroidNN: The top-left corner of a texel in the output |
| 43 | // image is projected into the input image to figure out the interpolants and weights. Note that this |
| 44 | // will yield different results than if projecting the centre of output texels. |
| 45 | |
| 46 | const unsigned int inputWidth = image.GetWidth(); |
| 47 | const unsigned int inputHeight = image.GetHeight(); |
| 48 | |
| 49 | // How much to scale pixel coordinates in the output image to get the corresponding pixel coordinates |
| 50 | // in the input image |
| 51 | const float scaleY = boost::numeric_cast<float>(inputHeight) / boost::numeric_cast<float>(outputHeight); |
| 52 | const float scaleX = boost::numeric_cast<float>(inputWidth) / boost::numeric_cast<float>(outputWidth); |
| 53 | |
| 54 | uint8_t rgb_x0y0[3]; |
| 55 | uint8_t rgb_x1y0[3]; |
| 56 | uint8_t rgb_x0y1[3]; |
| 57 | uint8_t rgb_x1y1[3]; |
| 58 | |
| 59 | for (unsigned int y = 0; y < outputHeight; ++y) |
| 60 | { |
| 61 | // Corresponding real-valued height coordinate in input image |
| 62 | const float iy = boost::numeric_cast<float>(y) * scaleY; |
| 63 | |
| 64 | // Discrete height coordinate of top-left texel (in the 2x2 texel area used for interpolation) |
| 65 | const float fiy = floorf(iy); |
| 66 | const unsigned int y0 = boost::numeric_cast<unsigned int>(fiy); |
| 67 | |
| 68 | // Interpolation weight (range [0,1]) |
| 69 | const float yw = iy - fiy; |
| 70 | |
| 71 | for (unsigned int x = 0; x < outputWidth; ++x) |
| 72 | { |
| 73 | // Real-valued and discrete width coordinates in input image |
| 74 | const float ix = boost::numeric_cast<float>(x) * scaleX; |
| 75 | const float fix = floorf(ix); |
| 76 | const unsigned int x0 = boost::numeric_cast<unsigned int>(fix); |
| 77 | |
| 78 | // Interpolation weight (range [0,1]) |
| 79 | const float xw = ix - fix; |
| 80 | |
| 81 | // Discrete width/height coordinates of texels below and to the right of (x0, y0) |
| 82 | const unsigned int x1 = std::min(x0 + 1, inputWidth - 1u); |
| 83 | const unsigned int y1 = std::min(y0 + 1, inputHeight - 1u); |
| 84 | |
| 85 | std::tie(rgb_x0y0[0], rgb_x0y0[1], rgb_x0y0[2]) = image.GetPixelAs3Channels(x0, y0); |
| 86 | std::tie(rgb_x1y0[0], rgb_x1y0[1], rgb_x1y0[2]) = image.GetPixelAs3Channels(x1, y0); |
| 87 | std::tie(rgb_x0y1[0], rgb_x0y1[1], rgb_x0y1[2]) = image.GetPixelAs3Channels(x0, y1); |
| 88 | std::tie(rgb_x1y1[0], rgb_x1y1[1], rgb_x1y1[2]) = image.GetPixelAs3Channels(x1, y1); |
| 89 | |
| 90 | for (unsigned c=0; c<3; ++c) |
| 91 | { |
| 92 | const float ly0 = Lerp(float(rgb_x0y0[c]), float(rgb_x1y0[c]), xw); |
| 93 | const float ly1 = Lerp(float(rgb_x0y1[c]), float(rgb_x1y1[c]), xw); |
| 94 | const float l = Lerp(ly0, ly1, yw); |
| 95 | PutData(out, outputWidth, x, y, c, l/255.0f); |
| 96 | } |
| 97 | } |
| 98 | } |
| 99 | |
| 100 | return out; |
| 101 | } |
| 102 | |
| 103 | } // end of anonymous namespace |
| 104 | |
| 105 | |
| 106 | MobileNetDatabase::MobileNetDatabase(const std::string& binaryFileDirectory, |
| 107 | unsigned int width, |
| 108 | unsigned int height, |
| 109 | const std::vector<ImageSet>& imageSet) |
| 110 | : m_BinaryDirectory(binaryFileDirectory) |
| 111 | , m_Height(height) |
| 112 | , m_Width(width) |
| 113 | , m_ImageSet(imageSet) |
| 114 | { |
| 115 | } |
| 116 | |
| 117 | std::unique_ptr<MobileNetDatabase::TTestCaseData> |
| 118 | MobileNetDatabase::GetTestCaseData(unsigned int testCaseId) |
| 119 | { |
| 120 | testCaseId = testCaseId % boost::numeric_cast<unsigned int>(m_ImageSet.size()); |
| 121 | const ImageSet& imageSet = m_ImageSet[testCaseId]; |
| 122 | const std::string fullPath = m_BinaryDirectory + imageSet.first; |
| 123 | |
| 124 | InferenceTestImage image(fullPath.c_str()); |
| 125 | |
| 126 | // this ResizeBilinear result is closer to the tensorflow one than STB. |
| 127 | // there is still some difference though, but the inference results are |
| 128 | // similar to tensorflow for MobileNet |
| 129 | std::vector<float> resized(ResizeBilinearAndNormalize(image, m_Width, m_Height)); |
| 130 | |
| 131 | const unsigned int label = imageSet.second; |
| 132 | return std::make_unique<TTestCaseData>(label, std::move(resized)); |
| 133 | } |