Testing MXNET

Running CPP Tests

1. Install cmake

2. Create a build directory in the root of the mxnet project

   mkdir build
   cd build
   

3. Generate your Makefile and build along with the tests with cmake (specify appropraite flags)

   cmake -DUSE_CUDNN=ON -DUSE_CUDA=ON -DUSE_MKLDNN=ON -DBLAS=Open -DCMAKE_BUILD_TYPE=Debug .. && make
   

4. Run tests

   ctest --verbose
   

5. The following will run all the tests the in cpp directory. To run just your test file replace the following in your tests/CMakeLists.txt

   file(GLOB_RECURSE UNIT_TEST_SOURCE "cpp/*.cc" "cpp/*.h")
   

   with

   file(GLOB_RECURSE UNIT_TEST_SOURCE "cpp/test_main.cc" "cpp/{RELATIVE_PATH_TO_TEST_FILE}")
   

Building with Ninja

Ninja is a build tool (like make) that prioritizes building speed. If you will be building frequently, we recommend you use ninja

1. Download Ninja via package manager or directly from source

   apt-get install ninja-build
   

2. When running cmake, add the -GNinja flag to specify cmake to generate a Ninja build file

   cmake -DUSE_CUDNN=ON -DUSE_CUDA=ON -DUSE_MKLDNN=ON -DBLAS=Open -GNinja -DCMAKE_BUILD_TYPE=Debug ..
   

3. Run the ninja build file with

   ninja
   

Runing Python Tests Within Docker

To run tests inside docker, you first need to install docker and docker-compose on your machine.

On Ubuntu you may install them via sudo apt-get install docker.io docker-compose and set them up via sudo usermod $(whoami) -G docker -a.

Then, to run tests inside docker run the following command

ci/build.py --platform {PLATFORM} /work/runtime_functions.sh {RUNTIME_FUNCTION}

An example for running python tests would be

ci/build.py --platform build_ubuntu_cpu_mkldnn /work/runtime_functions.sh unittest_ubuntu_python3_cpu