NuGet packages with native assets in a notebook

Describe the bug

When using the Microsoft.ML NuGet package in a notebook, you get an exception when ML.NET calls into a native assembly.

Repro

Run the following notebook:

#r "nuget:Microsoft.ML"
    
using Microsoft.ML;
using Microsoft.ML.Data;

class IrisData
{
    public IrisData(float sepalLength, float sepalWidth, float petalLength, float petalWidth)
    {
        SepalLength = sepalLength;
        SepalWidth = sepalWidth;
        PetalLength = petalLength;
        PetalWidth = petalWidth;
    }
    public float SepalLength;
    public float SepalWidth;
    public float PetalLength;
    public float PetalWidth;
}

var data = new []
{
    new IrisData(1.4f, 1.3f, 2.5f, 4.5f), 
    new IrisData(2.4f, 0.3f, 9.5f, 3.4f), 
    new IrisData(3.4f, 4.3f, 1.6f, 7.5f), 
    new IrisData(3.9f, 5.3f, 1.5f, 6.5f), 
};

MLContext mlContext = new MLContext();
var pipeline = mlContext.Transforms
    .Concatenate("Features", "SepalLength", "SepalWidth", "PetalLength", "PetalWidth")
    .Append(mlContext.Clustering.Trainers.KMeans("Features", numberOfClusters: 2));

try
{
    pipeline.Fit(mlContext.Data.LoadFromEnumerable(data));
}
catch (Exception e)
{
    Console.WriteLine(e);
}

This shouldn’t throw an exception, but it does:

System.DllNotFoundException: Unable to load shared library 'CpuMathNative' or one of its dependencies. In order to help diagnose loading problems, consider setting the DYLD_PRINT_LIBRARIES environment variable: dlopen(libCpuMathNative, 1): image not found
   at Microsoft.ML.Internal.CpuMath.Thunk.SumSqU(Single* ps, Int32 c)
   at Microsoft.ML.Internal.CpuMath.CpuMathUtils.SumSq(ReadOnlySpan`1 src)
   at Microsoft.ML.Numeric.VectorUtils.NormSquared(VBuffer`1& a)
   at Microsoft.ML.Trainers.KMeansBarBarInitialization.Initialize(IHost host, Int32 numThreads, IChannel ch, Factory cursorFactory, Int32 k, Int32 dimensionality, VBuffer`1[] centroids, Int64 accelMemBudgetMb, Int64& missingFeatureCount, Int64& totalTrainingInstances)
   at Microsoft.ML.Trainers.KMeansTrainer.TrainCore(IChannel ch, RoleMappedData data, Int32 dimensionality)
   at Microsoft.ML.Trainers.KMeansTrainer.TrainModelCore(TrainContext context)
   at Microsoft.ML.Trainers.TrainerEstimatorBase`2.TrainTransformer(IDataView trainSet, IDataView validationSet, IPredictor initPredictor)
   at Microsoft.ML.Trainers.TrainerEstimatorBase`2.Fit(IDataView input)
   at Microsoft.ML.Data.EstimatorChain`1.Fit(IDataView input)
   at Submission#6.<<Initialize>>d__0.MoveNext()

Did this error occur while using dotnet try or online?

• dotnet-try
• online

What kind of error was it?

• User Interface (UI): For example the output never displayed
• Service Error: For example “The service is temporarily unavailable. We are working on it”
• Other:

Screenshots

Please complete the following:

• OS
  • Windows 10
  • macOS
  • Linux (Please specify distro)
  • iOS
  • Android
• Browser
  • Chrome
  • Edge
  • Safari

[ericstj]

How we would create that .deps.json file during dotnet-interactive is not clear to me. My current thinking is that we could get by with inspecting the project.assets.json file that we get from a NuGet restore + the RID fallback graph I described above.

Does dotnet-interactive run a restore on the packages it loads? If it did, it shouldn’t be too much more to generate a deps file for that assembly. According to https://docs.microsoft.com/en-us/dotnet/core/tutorials/creating-app-with-plugin-support#load-plugins

It resolves assemblies and native libraries to their relative paths based on the .deps.json file for the class library whose path was passed to the AssemblyDependencyResolver constructor. The custom AssemblyLoadContext enables plugins to have their own dependencies, and the AssemblyDependencyResolver makes it easy to correctly load the dependencies.

Here’s a quick sample I threw together that produces a deps file for a package.

<Project Sdk="Microsoft.NET.Sdk" DefaultTargets="GenerateBuildDependencyFile">

  <PropertyGroup>
    <TargetFramework>net6.0</TargetFramework>
    <IncludeMainProjectInDepsFile>false</IncludeMainProjectInDepsFile>
    <NoBuild>true</NoBuild>
    <ProjectDepsFileName>Microsoft.ML.deps.json</ProjectDepsFileName>
  </PropertyGroup>

  <ItemGroup>
    <PackageReference Include="Microsoft.ML" Version="1.5.5" />
  </ItemGroup>

</Project>

[eerhardt]

Yep, that’s the one! That’s how you can get the RID of the current machine. The user shouldn’t need to supply it at all.