Skip to main content

ResNet50 Object Recognition

This document provides a detailed guide on how to use the NPU hardware acceleration on Sirider S1 to infer the ResNet50 model.

The document is divided into two parts: Quick Start and Detailed Tutorial

Quick Start​

Radxa provides a ready-to-use example for ResNet50 object classification. This allows users to directly run the ResNet50 model on Sirider S1 using AIPU, eliminating the need for complex model and code compilation.
This is ideal for users who want to quickly use AIPU without building the model from scratch. If you're interested in the full workflow, refer to the Detailed Tutorial.

  • Clone the repository:

    git clone https://github.com/zifeng-radxa/siriders1_NPU_example.git

  • Install dependencies:

    tip

    It is recommended to use virtualenv.

    cd siriders1_NPU_example
    pip3 install -r requirements.txt

  • Generate input files for the model:

    python3 input_gen.py --img_path <your_image_path>

  • Run model inference:

    export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:$(pwd)/libs
    ./aipu_test aipu_mlperf_resnet50.bin input_3_224_224.bin

Detailed Tutorial​

To deploy the target model on the Zhouyi Z2 AIPU, you need to follow three steps: model conversion, inference file compilation, and application-level program design.

Model Conversion​

tip

This step is completed on an x86 host. Before starting model conversion, install the Zhouyi SDK according to the Zhouyi AIPU SDK Installation Guide and complete the Configure the nn-compiler Environment.

The nn-compiler can convert models from TensorFlow, ONNX, etc., into model files optimized for hardware-accelerated inference on the Zhouyi AIPU.

  • Navigate to the nn-compiler-user-case-example directory:

    cd siengine/nn-compiler-user-case-example/onnx

  • Generate quantization calibration data:

    python3 generate_calibration_data.py

  • Generate binary input files for model inference:

    python3 generate_input_binary.py

    Output: ./resnet50/input_3_224_224.bin

  • (Optional) Edit build.cfg if needed (default provided in the example):

    vim ./resnet50/build.cfg

  • Build the AIPU model:

    cd ./resnet50
    aipubuild build.cfg

    Output: aipu_mlperf_resnet50.bin

tip

If the aipubuild command is not found, try adding it to your path:

export PATH=$PATH:/root/.local/bin

Inference on the Sirider S1 Board​

Before running inference on the Zhouyi Z2 AIPU, you need to cross-compile an executable file (aiputest) on an x86 host and transfer it to Sirider S1.

Cross-Compiling Executables on x86 Host​

  • Install the cross-compilation toolchain (gcc-linaro-7.5.0):

    tar -xvf gcc-linaro-7.5.0-2019.12-x86_64_aarch64-linux-gnu.tar
    cp -r gcc-linaro-7.5.0-2019.12-x86_64_aarch64-linux-gnu /opt

  • Compile aiputest:

    • Modify the UMDSRC path variable in CMakeLists.txt:

      cd siengine/nn-runtime-user-case-example
      vim CMakeLists.txt
      # Example: Update the path for your SDK installation
      set(UMDSRC "${CMAKE_SOURCE_DIR}/../../AI610-SDK-${AIPU_VERSION}-AIoT/Linux-driver/driver/umd")

    • Build the executable:

      mkdir build && cd build
      cmake -DCMAKE_BUILD_TYPE=Release ..
      make

      Output: siengine/nn-runtime-user-case-example/out/linux/aipu_test

Inference on Sirider S1 Board​

  • Transfer the following files to Sirider S1:

    • aipu_mlperf_resnet50.bin (model file)
    • input_3_224_224.bin (input file)
    • aipu_test (executable)
    • libs directory (dynamic libraries)

  • Run the aiputest program:

    export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:<your_libs_path>
    ./aipu_test aipu_mlperf_resnet50.bin input_3_224_224.bin

  • Sample Output:

    (aiot-focal_overlayfs)root@linux:~/ssd# ./aipu_test aipu_mlperf_resnet50.bin input_3_224_224.bin
    usage: ./aipu_test aipu.bin input0.bin
    aipu_init_context success
    aipu_load_graph_helper success: aipu_mlperf_resnet50.bin
    aipu_create_job success
    Frame #0
    aipu_finish_job success
    No profiler data
    get output tensor 0 success (1/1)
    output_desc zero_point: 0.0000 scale: 5.5835
    idx: 637  fval: 21.4919
    idx: 749  fval: 19.8800
    idx: 415  fval: 16.1189
    idx: 412  fval: 15.0443
    idx: 791  fval: 14.1488
    Frame #1
    aipu_finish_job success
    No profiler data
    get output tensor 0 success (1/1)
    output_desc zero_point: 0.0000 scale: 5.5835
    idx: 637  fval: 21.4919
    idx: 749  fval: 19.8800
    idx: 415  fval: 16.1189
    idx: 412  fval: 15.0443
    idx: 791  fval: 14.1488
    aipu_clean_job success
    aipu_unload_graph success
    aipu_deinit_ctx success

    In this case, the highest confidence index 637 corresponds to "mailbag, postbag" in the ImageNet 1000 categories.

    input.webp