API reference - XML documentation template for transforms

[sfilipi]

The XML documentation for the transforms should contain information about the schema: requirements about the type of the columns to work on, and information about the type of the columns produced.

@shmoradims

[wschin]

I'd suggest to document GetOutputSchema function and transformer's XML can reference GetOutputSchema's document.

[singlis]

Referencing #3127 as this feels same or related, maybe we could track this work under one issue? I think there are going to be a number of subsections that will need to be tracked. And should this go on the transforms? Or the extensions?

[sfilipi]

I'd suggest to document GetOutputSchema function and transformer's XML can reference GetOutputSchema's document.

It won't work, because GetOutputSchema is on the base class, and is calling GetOutputSchemaCore of every class, which is internal.
Also the types of the columns for the input columns need to be documented.

[sfilipi]

Proposal for the transforms XML template, mirroring the trainers template from issue: #3218,

1- XML on Transform extension method:

• Summary - Create an <see cref the estimator>estimator that <Short transform description>

• Parameters:
  inputColumnName: Name of column to transform. If set to <see langword="null"/>, the value of the <paramref name="outputColumnName"/> will be used as source.
  The data type on this column should be <ref required data type.> | The data type on this column can be any type.
  outputColumnName: Name of the column resulting from the transformation of <paramref name="inputColumnName"/>.
  This column's data type will be <info from GetOutputSchema here>.

• Example

2 - XML of the Estimator, Transformer

• Summary - One line description of what the transform does.
• Remarks - More details about the transform and its implementation.
  Does it do a pass through the data? Yes/No
  Input column type
  Output column type.
  Additional NuGet: "Link to NuGet" OR None of all that are included already in Microsoft.ML
• SeeAlso cref the extension method for an example usage.

cc @natke @shmoradims @glebuk @singlis

[natke]

Looks good @sfilipi. One small comment: for the summary of the extension method, would we want to say: "Create an xxx transform ..."?

[artidoro]

Shouldn't it say "Create an xxx estimator" instead of "transform"?
@sfilipi

[sfilipi]

Shouldn't it say "Create an xxx estimator" instead of "transform"?
@sfilipi

@natke suggested that we reference the components by their functionality: trainers and transforms.
In that specific case we'll keep it to estimator - i updated the template above to reflect that.

in the Remarks of the actual estimator we're keeping transform, as a term.

[shmoradims]

• Do we need the remarks? Can the summary just say: "Create an <cref=the estimator>"?
• Let's add a short description of the transform in the summary, similar to trainers.
• Onnx -> ONNX

[sfilipi]

Transform priority is in this old issue. Use it as guide to prioritize more popular transforms. Also, reuse descriptions from here.

List of transforms to sign up for:

MLContext.Transforms (root)	API Owner	Priority	Status
CopyColumns	Senja		Done PrToV1: #3348
Concatenate	Artidoro	0	Done
DropColumns	Artidoro	3	Done
SelectColumns	Artidoro	3	Done
Normalize.MinMax	Scott		Done #3432
Normalize.MeanVariance	Scott		Done  #3432
Normalize.LogMeanVariance	Scott		Done  #3432
Normalize.Binning	Scott		Done #3432
Normalize.SupervisedBinning	Scott		Done #3432
CustomMapping	Artidoro		Done
IndicateMissingValues	Ivan		Done #3386
ReplaceMissingValues	Ivan		Done #3386
ConvertToGrayscale	Ivan		In PR #3376
LoadImages	Ivan		In PR #3376
ExtractPixels	Ivan		In PR #3376
ResizeImages	Ivan		In PR #3376
ConvertToImage	Ivan		In PR #3376
IidChangePointEstimator	Wei-Sheng		Done #3444
IidSpikeEstimator	Wei-Sheng		Done #3444
SsaChangePointEstimator	Wei-Sheng		Done #3444
SsaSpikeEstimator	Wei-Sheng		Done #3444
ApplyOnnxModel	Gani		Done  #3387
DnnFeaturizeImage	Senja
NormalizeGlobalContrast	Artidoro		Done
NormalizeLpNorm	Artidoro		Done
ApproximatedKernelMap	Yael		Done #3377
CalculateFeatureContribution	Yael		Done #3384

[sfilipi]

Other catalogs:

MLContext.Transforms.Categorical	API Owner	Priority	Status
OneHotEncoding	Najeeb	0	Done #3388
OneHotHashEncoding	Najeeb	1	Done #3388

MLContext.Transforms.Conversion	API Owner	Priority	Status
Hash	Senja		Done
ConvertType	Senja		Done
MapKeyToValue	Senja		Done
MapKeyToVector	Senja		Done
MapValueToKey	Senja		Done
MapKeyToBinaryVector	Senja		Done

MLContext.Transforms.FeatureSelection	API Owner	Priority	Status
SelectFeaturesBasedOnMutualInformation	Senja need a better example to show MI computation. something like this		Done
SelectFeaturesBasedOnCount	Senja		Done

MLContext.Transforms.Text	API Owner	Priority	Status
FeaturizeText	Senja		Done #3438
TokenizeCharacters	Artidoro	2	Done #3418
NormalizeText	Artidoro	2	Done#3418
ExtractWordEmbeddings	Artidoro		Done #3418
TokenizeWords	Artidoro	2	Done #3418
ProduceNgrams	Artidoro	2	Done #3418
RemoveDefaultStopWords	Ivan		Done #3413
RemoveStopWords	Ivan		Done #3413
ProduceWordBags	Ivan		Done #3440
ProduceHashedWordBags	Ivan		Done #3440
ProduceHashedNgrams	Ivan		Done #3419
LatentDirichletAllocation	Senja		Done #3442

For the Data catalog, all API's documentations needs to be augmented with suggestions for when would one use this API.

MLContext.Data	API Owner	Priority	Status
LoadFromEnumerable	Najeeb		Done #3417
CreateEnumerable	Najeeb	The second overload of this API is a P4 scenario. the use case for that API would be: users has a model which has slot names preserved for the features, and when they load the models, they also get the schema out of the loaded model and pass that schema, together with the TRow type they want to load the data to this API. This API will then populate the Annotations (former metadata) for the feature column.	Done #3417
BootstrapSample	Najeeb		Done (previously by Rogan)
Cache	Najeeb		Done (previously by Rogan)
FilterRowsByColumn	Najeeb		Done (previously by Rogan)
FilterRowsByKeyColumnFraction	Najeeb		Done (previously by Rogan)
FilterRowsByMissingValues	Wei-Sheng		Done (previously by Rogan)
ShuffleRows	Wei-Sheng		Done (previously by Rogan)
SkipRows	Wei-Sheng #3415		Done (previously by Rogan)
TakeRows	Wei-Sheng #3415		Done (previously by Rogan)

Other	API Owner	Priority	Status
Permutation Feature Importance	Shahab		Doen by @codemzs
MLContext.Model (root)	Shahab		#3451

[shmoradims]

Finished 4/21 9:29pm. Great team work.