Forums - Preprocessing image for SNPE and Tiny YOLO

Steps to build SNPE neural network and get the output FloatTensor:

1. Create an asset folder in Android/app directory and keep the model file(.dlc) in the asset folder.
        // assetFileName is the file name of .dlc
        InputStream assetInputStream = application.getAssets().open(assetFileName); // Create and build the neural network
        NeuralNetwork network = new SNPE.NeuralNetworkBuilder(application)
                .setDebugEnabled(false)
        //outputLayerNames can be got while converted model to DLC format
                .setOutputLayers(outputLayerNames)
                .setModel(assetInputStream, assetInputStream.available())
                .setPerformanceProfile(NeuralNetwork.PerformanceProfile.DEFAULT)
                .setRuntimeOrder(selectedRuntime) // Runtime.DSP, Runtime.GPU_FLOAT16, Runtime.GPU, Runtime.CPU
                .setCpuFallbackEnabled(needsCpuFallback)
                .build();
        // Close input
        assetInputStream.close();
2. Create an Input Tensor
3. Propagate Input Tensors Through the Network
4. Process the Neural Network Output

Please follow the link below and find sections mentioned in steps 2,3 and 4 for preparing input Tensors and processing output tensors https://developer.qualcomm.com/docs/snpe/android_tutorial.html

with above setup i am only able to achieve 4-5fps on GPU for the yolov4 tiny model. I have tried many variations of the NeuralNetworkBuilder like below:
 

val builder: SNPE.NeuralNetworkBuilder = SNPE.NeuralNetworkBuilder(Cerebro.app)

                .setDebugEnabled(false)
//                .setRuntimeOrder(NeuralNetwork.Runtime.DSP, NeuralNetwork.Runtime.GPU, NeuralNetwork.Runtime.GPU_FLOAT16,  NeuralNetwork.Runtime.CPU)
                .setRuntimeOrder(NeuralNetwork.Runtime.GPU_FLOAT16)
                // CPU fall back is needed for our model to run on the DSP, some ops not supported
                .setCpuFallbackEnabled(true)
                // Determined by the layer names from snpe-dlc-info -i quant.dlc
//                .setOutputLayers("concat_29", "concat_30") // yolov4_quantized.dlc
                .setOutputLayers("concat_21", "concat_22") // fp32.dlc
                .setModel(File(modelPath))
                .setPerformanceProfile(NeuralNetwork.PerformanceProfile.HIGH_PERFORMANCE)
                .setExecutionPriorityHint(NeuralNetwork.ExecutionPriorityHint.HIGH)
                .setUnsignedPD(false)
                .setRuntimeCheckOption(NeuralNetwork.RuntimeCheckOption.BASIC_CHECK)

This same model runs at 25fps on other non-qc platforms.

please help!