CustomPipelineColumn

[Ivanidzo4ka]

Fixes #1043

[Ivanidzo4ka]

@TomFinley Is this what you meant by

should be another subclass out of PipelineColumn for non-vector structural types.

[Zruty0]

[Zruty0]

[](start = 8, length = 32)

remove empty #Resolved

[TomFinley]

Hi @Ivanidzo4ka. Note that you must also add this here, so that an analysis instance is actually created.

machinelearning/src/Microsoft.ML.Data/StaticPipe/StaticPipeInternalUtils.cs

Line 97 in 1ae6070

public static object MakeAnalysisInstanceCore<T>(Rec rec)

But this is interesting: your build actually passed. I guess no where do we actually create an analysis instance over types.

Let's have a test that will fail, since really we should have caught it in a test. This might be a good starting point...

machinelearning/test/Microsoft.ML.StaticPipelineTesting/ImageAnalyticsTests.cs

Line 26 in 1ae6070

ctx => ctx.LoadText(0).LoadAsImage().AsGrayscale().Resize(10, 8).ExtractPixels());

The goal is to have some sort of case where we have an analysis instance that includes this new type. I think this would do it:

var reader = TextLoader.CreateReader(env,
    ctx => ctx.LoadText(0).LoadAsImage());
var est = reader.MakeNewEstimator().Append(r => r.AsGrayscale().Resize(10, 8).ExtractPixels());
reader.Append(est);

That makes the test more verbose in what is from a user's perspective kind of pointless, but it would I think catch this problem.

[TomFinley]

Thanks @Ivanidzo4ka ! If we could make the analysis instance creation work with this new type, and also perhaps have a test that would have caught the problem, then I'll be super happy with this. 😄

[TomFinley]

Thanks @Ivanidzo4ka !