Hi,
I followed the code in the demo and ran it on my phone, but the results I got are inconsistent with the results I got using the benchmark tool. I want to ask for help.
My device is Snapdragon 888, and the compilation environment is ubuntu20.04.
My code:
zdl::DlSystem::TensorShape tensorShape;
tensorShape = snpe->getInputDimensions();
size_t batchSize = tensorShape.getDimensions()[0];
std::cout << "Batch size for the container is " << batchSize << std::endl;
bool execStatus = false;
std::vector<float> inputVec;
cv::Mat inputImage = cv::imread(image_path, cv::IMREAD_COLOR);
cv::Mat floatImage;
inputImage.convertTo(floatImage, CV_32F, 1 / 255.0, 0);
std::unique_ptr<zdl::DlSystem::ITensor> inputTensor =
zdl::SNPE::SNPEFactory::getTensorFactory().createTensor(tensorShape);
std::copy(inputVec.begin(), inputVec.end(), inputTensor->begin());
execStatus = snpe->execute(inputTensor.get(), outputTensorMap);
Dear customer,
Which SNPE version you have tested?
What's type of data format you feeds?
SNPE demo code provided the conversion step to customer if you feed with float data. I assume that you need to change the core flow of float to fixed point and reverse.
BR.
Wei