HOME

See development workflow page for developer instructions.

---

Here are some useful tutorial documents:

• LOOM - part 1 - intro to 360 cameras
• LOOM - part 2 - developing 3rd party LoomIO plug-ins

---

Getting Started with Neural Network Inference Sample

Setup and run annInferenceServer

1. Recommended server configuration: EPYC(tm) processor with multiple Vega-based GPUs
2. Install Ubuntu 16.04 64-bit
3. Install ROCm from AMD repositories with OpenCL development kit
4. Install protobuf for inference_generator -- just checkout https://github.com/google/protobuf and follow the C++ Installation Instructions
5. Checkout, Build, and Install dependent projects

    % git clone https://github.com/RadeonOpenCompute/rocm-cmake
    % mkdir -p rocm-cmake/build
    % cd rocm-cmake/build
    % cmake -DCMAKE_BUILD_TYPE=Release ..
    % sudo make install
    % cd ../..

    % git clone https://github.com/ROCmSoftwarePlatform/MIOpenGEMM
    % mkdir -p MIOpenGEMM/build
    % cd MIOpenGEMM/build
    % cmake -DCMAKE_BUILD_TYPE=Release ..
    % make -j4
    % sudo make install
    % cd ../..

    % git clone https://github.com/ROCmSoftwarePlatform/MIOpen
    % mkdir -p MIOpen/build
    % cd MIOpen/build
    % cmake -DMIOPEN_BACKEND=OpenCL -DCMAKE_BUILD_TYPE=Release ..
    % make -j4 MIOpenDriver
    % sudo make install
    % cd ../..

6. amdovx-modules: Checkout, Build, and Install -- optionally you can pick the develop branch using -b develop in the git clone command-line.

    % git clone --recursive https://github.com/GPUOpen-ProfessionalCompute-Libraries/amdovx-modules
    % mkdir -p amdovx-modules/build
    % cd amdovx-modules/build
    % cmake -DCMAKE_BUILD_TYPE=Release ..
    % make -j4
    % sudo make install
    % cd ../..

7. Add /opt/rocm/bin to PATH environment variable
8. Add /opt/rocm/lib to LD_LIBRARY_PATH environment variable
9. Make sure that OpenCL library is in LD_LIBRARY_PATH environment variable (e.g., /opt/rocm/opencl/lib/x86_64)
10. Run /opt/rocm/bin/annInferenceServer

Setup and run annInferenceApp

1. Use another workstation for annInferenceApp
2. Build annInferenceApp
3. Run annInferenceApp
   • Connect to the server (use port 28282)
   • Upload CAFFE model (such as ResNet-50 for ImageNet)
     • Select .prototxt, .caffemodel, input dimensions, and other optional parameters
     • Click Upload & Compile
   • Run Inference
     • Select synset txt file with one name of each output label per line
     • Select input image folder (such as ImageNet validation data)
     • Click Run

Neural Network Inference Workflow for CAFFE users

First you need to build a C++ library from your pre-trained CAFFE model.

1. Use inference_generator to generate the C++ code

        % caffe2openvx weights.caffemodel <batchSize> <inputChannels> <inputHeight> <inputWidth>
        % caffe2openvx deploy.prototxt <batchSize> <inputChannels> <inputHeight> <inputWidth>
        (this will create annmodule.h, annmodule.cpp, anntest.cpp, CMakeLists.txt, weights(folder), bias(folder))

2. Use cmake to build the C++ source code (above) to produce libannmodule.so (library) and anntest (test program). See inference_generator for further details.
3. Use the anntest test program for sanity tests. This program optionally takes folder containing weights and bias sub-directories followed by filenames of input tensor data and output tensor data. The tensor data will be raw binary dump of float values.

        Usage: anntest [<folder-containing-weights-and-bias> [<raw-input-tensor.dat> [<raw-output-tensor.dat>]]]

        Example:
          % ./anntest . input-f32.dat output-f32.dat

4. Study the anntest.cpp to understand the application integration.

Integrate annmodule into you application.

1. Add annmodule.h and annmodule.cpp (or libannmodule.so) into your project space.
2. All the dependencies, i.e., openvx, vx_nn, MIOpenGEMM, MIOpen, and OpenCL, will be available in /opt/rocm folder.
3. Use annGetTensorDimensions() in annmodule.h to query and check neural network input and output dimensions
4. Use annCreateGraph() to create and initialize neural network -- this function returns an OpenVX graph object that is ready to process (i.e., to run inference)
5. Just use vxProcessGraph() or vxScheduleGraph() APIs to the run inference

To deploy your application, you need to package the weights and bias folders with your application executable. Note that the application need to specify the path of installed folder to annCreateGraph() function call.