[BUG Report]: Tensorflow.Numpy.NDArray handling exception with .NET 4.8

[abrguyt]

Description

Using .NET 4.8 and the latest nightly (Tensorflow.Keras 0.150.0-BERT-Model-46-g6a2d7e1c-nightly) and simply preparing Tensorflow.NumPy.NDArray data objects for Keras model fitting; it seems NDArray objects are created whose contents cannot be read properly:

float[] floatArray = [1f, 1.1f, 1.11f, 2f, 2.2f, 2.22f];

var targetShape = new Tensorflow.Shape(2, 3);

var tf_nd_array = new Tensorflow.NumPy.NDArray(floatArray, targetShape);

// simply reading value from row 0, column 0 will result in an exception
var value = tf_nd_array[0, 0];

The exception thrown when a value is read from the NDArray object is:

The type initializer for 'OnMethodBoundaryAspectCompile.MethodInfos' threw an exception.
Inner Exception:
FileNotFoundException: Could not load file or assembly 'System.Private.CoreLib, Version=8.0.0.0, Culture=neutral, PublicKeyToken=7cec85d7bea7798e

This exception seems related to the NuGet 'MethodBoundaryAspect.Fody' dependency and a reference to .NET 8, while this all is running under .NET 4.8.

Can anyone confirm what the last version of Tensorflow.Keras is that can run on .NET 4.8 ?

Reproduction Steps

No response

Known Workarounds

No response

Configuration and Other Information

Tensorflow.Keras 0.150.0-BERT-Model-46-g6a2d7e1c-nightly
.NET 4.8
Windows 10

[abrguyt]

I reverted back to just using the 'official' latest Tensorflow.Keras v. 0.15 which doesn't have this issue.

There is one other problem though with the NuGet Tensorflow.Keras v. 0.15 / Tensorflow.NET v 0.150 versions - model.save() does not work and results in an error under .NET 4.8.

However as suggested by https://github.com/SciSharp/TensorFlow.NET/issues/1246#issuecomment-2134968838 just downloading the Github Repo and building and/or repackaging yourself solve this model.save() issue - somehow the DLL files produced when doing a custom build results in different DLL files from those available on NuGet.

So now I got it all working with (custom) packaged Tensorflow.Keras v. 0.15 and .NET 4.8 - thanks for very valuable input on this forum.