| 17 | |
| 18 | template <typename TDataType> |
| 19 | unsigned int ImagePreprocessor<TDataType>::GetLabelAndResizedImageAsFloat(unsigned int testCaseId, |
| 20 | std::vector<float> & result) |
| 21 | { |
| 22 | testCaseId = testCaseId % armnn::numeric_cast<unsigned int>(m_ImageSet.size()); |
| 23 | const ImageSet& imageSet = m_ImageSet[testCaseId]; |
| 24 | const std::string fullPath = m_BinaryDirectory + imageSet.first; |
| 25 | |
| 26 | InferenceTestImage image(fullPath.c_str()); |
| 27 | |
| 28 | // this ResizeBilinear result is closer to the tensorflow one than STB. |
| 29 | // there is still some difference though, but the inference results are |
| 30 | // similar to tensorflow for MobileNet |
| 31 | |
| 32 | result = image.Resize(m_Width, m_Height, CHECK_LOCATION(), |
| 33 | InferenceTestImage::ResizingMethods::BilinearAndNormalized, |
| 34 | m_Mean, m_Stddev, m_Scale); |
| 35 | |
| 36 | // duplicate data across the batch |
| 37 | for (unsigned int i = 1; i < m_BatchSize; i++) |
| 38 | { |
| 39 | result.insert(result.end(), result.begin(), result.begin() + armnn::numeric_cast<int>(GetNumImageElements())); |
| 40 | } |
| 41 | |
| 42 | if (m_DataFormat == DataFormat::NCHW) |
| 43 | { |
| 44 | const armnn::PermutationVector NHWCToArmNN = { 0, 2, 3, 1 }; |
| 45 | armnn::TensorShape dstShape({m_BatchSize, 3, m_Height, m_Width}); |
| 46 | std::vector<float> tempImage(result.size()); |
| 47 | armnnUtils::Permute(dstShape, NHWCToArmNN, result.data(), tempImage.data(), sizeof(float)); |
| 48 | result.swap(tempImage); |
| 49 | } |
| 50 | |
| 51 | return imageSet.second; |
| 52 | } |
| 53 | |
| 54 | template <> |
| 55 | std::unique_ptr<ImagePreprocessor<float>::TTestCaseData> |