Contents

  • This script is used for testing multiple onnx models using amdMIGraphXnode node extension for inference.

  • The script takes one image for some of the models which use an imagenet dataset and one image for the mnist model.

  • Models tested in this script (needed as input). All models can be found here.

    1. resnet50

    2. googlenet

    3. alexnet

    4. densenet

    5. squeezenet

    6. vgg19

  • Test one or more of the models by specifying the paths to the models which need testing.

  • Tests batch sizes from 1-N (N specified by user)

  • Currently only mode 1 – runs ONNX models

  • Currently only Batch size = 1 works (TO test bigger batch sizes, build and install this branch)

  • Usage:

Usage: ./runMIGraphXTests
            --tensor                    <path to tensor directory>
            --profiler_mode  <range:0-2; default:1> [optional]
              Mode 0 - Run all tests\n"
              Mode 1 - Run all ONNX tests\n"
              Mode 2 - Run all JSON tests\n"
            --profiler_level <range:0-N; default:1> [N = batch size][optional]
            --resnet50       <resnet50-model>
            --googlenet      <googlenet-model>
            --squeezenet     <resnet101-model>
            --alexnet        <resnet152-model>
            --vgg19          <vgg19-model>
            --densenet       <densenet-model>

  • Example: (to run all models for batch sizes 1 to 128)

mkdir build
cd build
cmake ../
make -j
./runMIGraphXTests --tensor <path to Mike tensor folder> --googlenet <path to googlenet model> --alexnet <path to alexnet model> --squeezenet <path to squeezenet model> --resnet50 <path to resnet50 model> --vgg19 <path to vgg19 model> --densenet <path to densenet model> --profiler_level 1
Contents