Skip to content

Latest commit

 

History

History
109 lines (72 loc) · 6.87 KB

developer-guide.md

File metadata and controls

109 lines (72 loc) · 6.87 KB

Developer Guide

The repo can be built for the following platforms, using the provided setup and the following instructions.

Chip Windows Linux OS X
x64 Instructions Instructions Instructions

Building the repository

The ML.NET repo can be built from a regular, non-admin command prompt. The build produces multiple binaries that make up the ML.NET libraries and the accompanying tests.

Developer Workflow

The dev workflow describes the development process to follow. It is divided into specific tasks that are fast, transparent and easy to understand. The tasks are represented in scripts (cmd/sh) in the root of the repo.

For more information about the different options that each task has, use the argument -help when calling the script. For example:

build -help

Examples

  • Initialize the repo to make build possible (if the build fails because it can't find mf.cpp then perhaps you missed this step)
git submodule update --init
  • Building in release mode for platform x64
build.cmd -configuration Release /p:TargetArchitecture=x64
  • Building the src and then building and running the tests
build.cmd -test

Building individual projects

Note: Before working on individual projects or test projects you must run build from the root once before beginning that work. It is also a good idea to run build whenever you pull a large set of unknown changes into your branch.

Under the src directory is a set of directories, each of which represents a particular assembly in ML.NET.

For example the src\Microsoft.MachineLearning.Core directory holds the source code for the Microsoft.MachineLearning.Core.dll assembly.

You can build the DLL for Microsoft.MachineLearning.Core.dll by going to the src\Microsoft.MachineLearning.Core directory and typing dotnet build.

You can build the tests for Microsoft.MachineLearning.Core.dll by going to test\Microsoft.MachineLearning.Core.Tests directory and typing dotnet test.

Note: We use build/vsts-ci.yml to define our official build

Building in Release or Debug

By default, building from the root or within a project will build the libraries in Debug mode. One can build in Debug or Release mode from the root by doing build.cmd -configuration Release or build.cmd -configuration Debug.

Currently, the full list of supported configurations are:

  • Debug, Release (for .NET Core 3.1 and .NET Framework 4.6.1)

Building other Architectures

We support both 32-bit and 64-bit binaries. To build in 32-bit, use the TargetArchitecture flag as below:

build.cmd -configuration Debug /p:TargetArchitecture=x86

Updating manifest and ep-list files

During development, there may arise a need to update the current baseline core_manifest.json and/or core_ep-list.tsv files. For example, a change in the name or type of a variable in a given class in the API that is not reflected in core_manifest.json will trigger the following failure:

*** Failure: Output and baseline mismatch at line 123 , expected ' ...x... ' but got ' ...y...' : '../Common/EntryPoints/core_manifest.json'

Steps to update core_manifest.json and core_ep-list.tsv:

  1. Unskip the RegenerateEntryPointCatalog unit test in test/Microsoft.ML.Core.Tests/UnitTests/TestEntryPoints.cs. This can be done by temporarily commenting out the skip attribute on the unit test for RegenerateEntryPointCatalog ([Fact(Skip = "Execute this test if you want to regenerate the core_manifest and core_ep_list files")]).
  2. Run the unit tests build.cmd -runTests (alternatively, run the RegenerateEntryPointCatalog unit test natively on Visual Studio through the Test Explorer or through Shortcuts by clicking on RegenerateEntryPointCatalog in test/Microsoft.ML.Core.Tests/UnitTests/TestEntryPoints.cs and pressing Ctrl+R,T).
  3. Verify the changes to core_manifest.json and core_ep-list.tsv are correct.
  4. Re-enable the skip attribute on the RegenerateEntryPointCatalog test.
  5. Commit the updated core_manifest.json and core_ep-list.tsv files to your branch.

Running specifics unit tests on CI

It may be necessary to run only specific unit tests on CI, and perhaps even run these tests back to back multiple times. The steps to run one or more unit tests are as follows:

  1. Set runSpecific: true and innerLoop: false in .vsts-dotnet-ci.yml per each build you'd like to run the specifics tests on CI.
  2. Import Microsoft.ML.TestFrameworkCommon.Attributes in the unit test files that contain specific unit tests to be run.
  3. Add the [TestCategory("RunSpecificTest")] to the unit test(s) you'd like to run specifically.

If you would like to run these specific unit test(s) multiple times, do the following for each unit test to run:

  1. Replace the [Fact] attribute with [Theory, IterationData(X)] where X is the number of times to run the unit test.
  2. Add the int iteration argument to the unit test you'd like to run multiple times.
  3. Use the iteration parameter at least once in the unit test. This may be as simple as printing to console the iteration parameter's value.

These steps are demonstrated in this demonstrative commit.

Running unit tests through VSTest Task & Collecting memory dumps

During development, there may also arise a need to debug hanging tests. In this scenario, it can be beneficial to collect the memory dump while a given test is hanging.

In this case, the given needs to be implemented according to the Microsoft test framework. Please check out the Microsoft test framework walkthrough and the VSTest sample demonstrating the "TestClass", "TestMethod", "DataTestMethod", and "DataRow" attributes.

Once the unit test(s) are implemented according to VSTest and ready to be debugged, the useVSTestTask parameter in build\ci\job-template.yml needs to be set to True. Once these steps are completed and pushed in your pull request, the unit test(s) will run and produce a full memory dump. At the end of a run, the memory dump .dmp file will be available for downloading and inspection in the published artifacts of the build, in the folder TestResults.

Note: this is only supported on Windows builds, as ProcDump is officially only available on Windows.