Added onnx export support for VectorWhitening

[kere-nel]

No description provided.

[codecov]

Codecov Report

❗ No coverage uploaded for pull request base (master@fa75f4f). Click here to learn what that means.
The diff coverage is 97.01%.

@@            Coverage Diff            @@
##             master    #4577   +/-   ##
=========================================
  Coverage          ?   75.13%           
=========================================
  Files             ?      909           
  Lines             ?   160334           
  Branches          ?    17267           
=========================================
  Hits              ?   120466           
  Misses            ?    35050           
  Partials          ?     4818

Flag	Coverage Δ
#Debug	75.13% <97.01%> (?)
#production	70.53% <95%> (?)
#test	90.3% <100%> (?)

Impacted Files	Coverage Δ
test/Microsoft.ML.Tests/OnnxConversionTest.cs	95.66% <100%> (ø)
src/Microsoft.ML.Mkl.Components/VectorWhitening.cs	81.76% <95%> (ø)

[yaeldekel]

for (int i = 0; i < parameters.Rank; i++) [](start = 20, length = 41)

If I understand correctly, this loop copies the first rank*dimension elements from model to principalComponents in the same order. Do we really need to copy?
I noticed that ctx.AddInitializer takes an IEnumerable, perhaps it would be enough to call it with model.Take(rank*dimension)?

[kere-nel]

I think the problem with Take is that when the rank is higher than the dimension, ML.NET simply treats the extra eigenvectors as rows of zeros. For ONNX, the model tensor has to be resized. I can introduce a check for both ML.NET and ONNX to check that the rank <= dimension?

[kere-nel]

/azp run

[azure-pipelines]

Azure Pipelines successfully started running 2 pipeline(s).