ML.NET v1.4 doesn't work on Azure Functions with .NET Core 2.1

[eerhardt]

When creating an Azure Function on netcoreapp2.1 and using ML.NET v1.4, it is failing to load the CpuMathNative library.

Between ML.NET 1.3. and 1.4 we’ve updated our “CpuMath” package in order to support .NET Core 3.0. We are now taking advantage of the new C# Hardware Intrinsics (SIMD) APIs in .NET Core 3.0. In order to make that change, we had to refactor our NuGet package a bit, and that refactoring seems to break Azure Functions.

In 1.4, our native assembly has moved inside the NuGet package.

Before

runtimes\win-x64\native\CpuMathNative.dll

After

runtimes\win-x64\nativeassets\netstandard2.0\CpuMathNative.dll

This change was needed because we don’t need the native assembly when running on netcoreapp3.0. So taking advantage of the runtimes\{RID}\nativeassets\{TFM} functionality in NuGet allowed us to not ship an unnecessary native binary in the customer’s app. For more information on this support in NuGet, see NuGet/Home#2782.

When running an Azure Functions app, it isn’t handling the nativeassets folder correctly. Thus, the CpuMathNative assembly isn’t loaded, and the Azure Function fails.

Repro steps

Provide the steps required to reproduce the problem:

1. Clone and check out https://github.com/dotnet/machinelearning-samples/tree/46fb968a9af744f0f97741a0c1556f708bfb277d/samples/csharp/end-to-end-apps/ScalableMLModelOnAzureFunction.
2. Load the Azure Function project.
3. F5
4. From Powershell: Invoke-RestMethod "http://localhost:7071/api/AnalyzeSentiment" -Method Post -Body (@{SentimentText="This is a very bad steak"} | ConvertTo-Json) -ContentType "application/json"

Expected behavior

The app should work and the function should return "Negative" since bad steak is a negative sentiment.

Actual behavior

Invoke-RestMethod : The remote server returned an error: (500) Internal Server Error.
At line:1 char:1
+ Invoke-RestMethod "http://localhost:7071/api/AnalyzeSentiment" -Metho ...
+ ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
    + CategoryInfo          : InvalidOperation: (System.Net.HttpWebRequest:HttpWebRequest) [Invoke-RestMethod], WebException
    + FullyQualifiedErrorId : WebCmdletWebResponseException,Microsoft.PowerShell.Commands.InvokeRestMethodCommand

From Debugging in Visual Studio the exception is:

System.DllNotFoundException: 'Unable to load DLL 'CpuMathNative' or one of its dependencies: The specified module could not be found. (Exception from HRESULT: 0x8007007E)'

Known workarounds

The only workaround I know of is to use ML.NET 1.3.

Related information

I played locally with a change to the Azure Function native assembly loading code, and I was able to get this fixed when the function.deps.json file is present. If you want to take a look, my change is here.

[eerhardt]

@fabiocav - I'd like to take a stab at fixing this. Can you assign it to me?