telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 1 | // |
| 2 | // Copyright © 2017 Arm Ltd. All rights reserved. |
David Beck | ecb56cd | 2018-09-05 12:52:57 +0100 | [diff] [blame] | 3 | // SPDX-License-Identifier: MIT |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 4 | // |
| 5 | #pragma once |
| 6 | |
| 7 | #include <armnn/Exceptions.hpp> |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 8 | #include <VerificationHelpers.hpp> |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 9 | |
| 10 | #include <array> |
| 11 | #include <cstdint> |
| 12 | #include <vector> |
| 13 | #include <utility> |
| 14 | |
| 15 | class InferenceTestImageException : public armnn::Exception |
| 16 | { |
| 17 | public: |
| 18 | using Exception::Exception; |
| 19 | }; |
| 20 | |
| 21 | class InferenceTestImageLoadFailed : public InferenceTestImageException |
| 22 | { |
| 23 | public: |
| 24 | using InferenceTestImageException::InferenceTestImageException; |
| 25 | }; |
| 26 | |
| 27 | class InferenceTestImageOutOfBoundsAccess : public InferenceTestImageException |
| 28 | { |
| 29 | public: |
| 30 | using InferenceTestImageException::InferenceTestImageException; |
| 31 | }; |
| 32 | |
| 33 | class InferenceTestImageResizeFailed : public InferenceTestImageException |
| 34 | { |
| 35 | public: |
| 36 | using InferenceTestImageException::InferenceTestImageException; |
| 37 | }; |
| 38 | |
| 39 | class InferenceTestImageWriteFailed : public InferenceTestImageException |
| 40 | { |
| 41 | public: |
| 42 | using InferenceTestImageException::InferenceTestImageException; |
| 43 | }; |
| 44 | |
| 45 | class UnknownImageChannelLayout : public InferenceTestImageException |
| 46 | { |
| 47 | public: |
| 48 | using InferenceTestImageException::InferenceTestImageException; |
| 49 | }; |
| 50 | |
| 51 | class InferenceTestImage |
| 52 | { |
| 53 | public: |
| 54 | enum class WriteFormat |
| 55 | { |
| 56 | Png, |
| 57 | Bmp, |
| 58 | Tga |
| 59 | }; |
| 60 | |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 61 | // Common names used to identify a channel in a pixel. |
| 62 | enum class ResizingMethods |
| 63 | { |
| 64 | STB, |
| 65 | BilinearAndNormalized, |
| 66 | }; |
| 67 | |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 68 | explicit InferenceTestImage(const char* filePath); |
| 69 | |
| 70 | InferenceTestImage(InferenceTestImage&&) = delete; |
| 71 | InferenceTestImage(const InferenceTestImage&) = delete; |
| 72 | InferenceTestImage& operator=(const InferenceTestImage&) = delete; |
| 73 | InferenceTestImage& operator=(InferenceTestImage&&) = delete; |
| 74 | |
| 75 | unsigned int GetWidth() const { return m_Width; } |
| 76 | unsigned int GetHeight() const { return m_Height; } |
| 77 | unsigned int GetNumChannels() const { return m_NumChannels; } |
| 78 | unsigned int GetNumElements() const { return GetWidth() * GetHeight() * GetNumChannels(); } |
| 79 | unsigned int GetSizeInBytes() const { return GetNumElements() * GetSingleElementSizeInBytes(); } |
| 80 | |
| 81 | // Returns the pixel identified by the given coordinates as a 3-channel value. |
| 82 | // Channels beyond the third are dropped. If the image provides less than 3 channels, the non-existent |
| 83 | // channels of the pixel will be filled with 0. Channels are returned in RGB order (that is, the first element |
| 84 | // of the tuple corresponds to the Red channel, whereas the last element is the Blue channel). |
| 85 | std::tuple<uint8_t, uint8_t, uint8_t> GetPixelAs3Channels(unsigned int x, unsigned int y) const; |
| 86 | |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 87 | void StbResize(InferenceTestImage& im, const unsigned int newWidth, const unsigned int newHeight); |
| 88 | |
| 89 | |
| 90 | std::vector<float> Resize(unsigned int newWidth, |
| 91 | unsigned int newHeight, |
| 92 | const armnn::CheckLocation& location, |
| 93 | const ResizingMethods meth = ResizingMethods::STB, |
| 94 | const std::array<float, 3>& mean = {{0.0, 0.0, 0.0}}, |
FinnWilliamsArm | af8b72d | 2019-05-22 14:50:55 +0100 | [diff] [blame] | 95 | const std::array<float, 3>& stddev = {{1.0, 1.0, 1.0}}, |
| 96 | const float scale = 255.0f); |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 97 | |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 98 | void Write(WriteFormat format, const char* filePath) const; |
| 99 | |
| 100 | private: |
| 101 | static unsigned int GetSingleElementSizeInBytes() |
| 102 | { |
| 103 | return sizeof(decltype(std::declval<InferenceTestImage>().m_Data[0])); |
| 104 | } |
| 105 | |
| 106 | std::vector<uint8_t> m_Data; |
| 107 | unsigned int m_Width; |
| 108 | unsigned int m_Height; |
| 109 | unsigned int m_NumChannels; |
| 110 | }; |
| 111 | |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 112 | // Common names used to identify a channel in a pixel. |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 113 | enum class ImageChannel |
| 114 | { |
| 115 | R, |
| 116 | G, |
| 117 | B |
| 118 | }; |
| 119 | |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 120 | // Channel layouts handled by the test framework. |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 121 | enum class ImageChannelLayout |
| 122 | { |
| 123 | Rgb, |
| 124 | Bgr |
| 125 | }; |
| 126 | |
| 127 | // Reads the contents of an inference test image as 3-channel pixels whose channel values have been normalized (scaled) |
| 128 | // and now lie in the range [0,1]. Channel data is stored according to the ArmNN layout (CHW). The order in which |
| 129 | // channels appear in the resulting vector is defined by the provided layout. |
| 130 | std::vector<float> GetImageDataInArmNnLayoutAsNormalizedFloats(ImageChannelLayout layout, |
Matteo Martincigh | 9a5f9f2 | 2019-10-31 11:02:47 +0000 | [diff] [blame] | 131 | const InferenceTestImage& image); |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 132 | |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 133 | // Reads the contents of an inference test image as 3-channel pixels, whose value is the result of subtracting the mean |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 134 | // from the values in the original image. Channel data is stored according to the ArmNN layout (CHW). The order in |
| 135 | // which channels appear in the resulting vector is defined by the provided layout. The order of the channels of the |
| 136 | // provided mean should also match the given layout. |
| 137 | std::vector<float> GetImageDataInArmNnLayoutAsFloatsSubtractingMean(ImageChannelLayout layout, |
Matteo Martincigh | 9a5f9f2 | 2019-10-31 11:02:47 +0000 | [diff] [blame] | 138 | const InferenceTestImage& image, |
| 139 | const std::array<float, 3>& mean); |
surmeh01 | bceff2f | 2018-03-29 16:29:27 +0100 | [diff] [blame] | 140 | |
| 141 | // Reads the contents of an inference test image as 3-channel pixels and returns the image data as normalized float |
| 142 | // values. The returned image stay in the original order (HWC) order. The C order may be changed according to the |
| 143 | // supplied layout value. |
| 144 | std::vector<float> GetImageDataAsNormalizedFloats(ImageChannelLayout layout, |
| 145 | const InferenceTestImage& image); |