Skip to content

Adding KeyToValueTransformers before finishing exporting to ONNX #4841

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
antoniovs1029 merged 18 commits into from
Feb 25, 2020
Merged

Adding KeyToValueTransformers before finishing exporting to ONNX #4841

Show file tree
Hide file tree
Changes from 11 commits
Commits
Show all changes
18 commits
Select commit Hold shift + click to select a range
1a826dc
Experimentations
antoniovs1029 Feb 14, 2020
009a84f
Removing MyTest method
antoniovs1029 Feb 14, 2020
f11c197
Removing unused comment
antoniovs1029 Feb 14, 2020
3a972f5
Fixing a change a made by mistake in an unrelated test
antoniovs1029 Feb 14, 2020
3e432a1
Added test for debugging purposes
antoniovs1029 Feb 15, 2020
11b84d0
Add support to map back keys from columns that where untouched by the…
antoniovs1029 Feb 15, 2020
a786d9f
Added security checks and removed redundant steps
antoniovs1029 Feb 19, 2020
cfa264d
Added todo comment and formatting
antoniovs1029 Feb 21, 2020
86fb14a
Skip Custom tests on CI
antoniovs1029 Feb 21, 2020
afec6d6
Merge remote-tracking branch 'upstream/master' into is23onnxnimbus
antoniovs1029 Feb 21, 2020
3e5b2bd
Removed my custom tests
antoniovs1029 Feb 21, 2020
28e447d
Merge remote-tracking branch 'upstream/master' into is23onnxnimbus
antoniovs1029 Feb 21, 2020
8864d8e
Added security checks to confirm that the PredictedLabel column has K…
antoniovs1029 Feb 22, 2020
d382373
Merge remote-tracking branch 'upstream/master' into is23onnxnimbus
antoniovs1029 Feb 24, 2020
ff9650e
Added HasKeyValues helper method and changed pass-through columns fil…
antoniovs1029 Feb 24, 2020
953ad7b
Fix extra space
antoniovs1029 Feb 24, 2020
9c63f1a
Merge remote-tracking branch 'upstream/master' into is23onnxnimbus
antoniovs1029 Feb 24, 2020
d70843f
Corrected and refactored GetPassThroughKeyDataViewTypeColumnsNames logic
antoniovs1029 Feb 25, 2020
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
33 changes: 33 additions & 0 deletions src/Microsoft.ML.OnnxConverter/SaveOnnxCommand.cs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -15,6 +15,7 @@
using Microsoft.ML.Internal.Utilities;
using Microsoft.ML.Model.OnnxConverter;
using Microsoft.ML.Runtime;
using Microsoft.ML.Transforms;
using Newtonsoft.Json;
using static Microsoft.ML.Model.OnnxConverter.OnnxCSharpToProtoWrapper;

Expand Down Expand Up @@ -308,6 +309,16 @@ private void Run(IChannel ch)
Host.Assert(scorePipe.Source == end);
end = scorePipe;
transforms.AddLast(scoreOnnx);

if(rawPred.PredictionKind == PredictionKind.BinaryClassification || rawPred.PredictionKind == PredictionKind.MulticlassClassification)
{
if(scorePipe.Schema.GetColumnOrNull("PredictedLabel")?.Type is KeyDataViewType)

@yaeldekel yaeldekel Feb 19, 2020

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

is KeyDataViewType [](start = 84, length = 18)

This still doesn't mean that the KeyToValueMappingTransformer will work - there can be key type columns that don't have key value annotations. I think there is a utility method called something like HasKeyNames that you can use.

@antoniovs1029 antoniovs1029 Feb 21, 2020

Copy link
Copy Markdown
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

The HasKeyValues method requires me to specify the keyValueItemType. Does anyone know if there's a way to avoid having to specify it, since in here I don't care about the keyValueItemType, as long as it has key value annotations.? @ganik @harishsk (if I leave it null, it defaults to be TextDataViewType)

machinelearning/src/Microsoft.ML.Data/Data/SchemaAnnotationsExtensions.cs

Lines 48 to 63 in b34c3b6

public static bool HasKeyValues(this DataViewSchema.Column column, PrimitiveDataViewType keyValueItemType = null)
{
// False if type is neither a key type, or a vector of key types.
if (!(column.Type.GetItemType() is KeyDataViewType keyType))
return false;
if (keyValueItemType == null)
keyValueItemType = TextDataViewType.Instance;
var metaColumn = column.Annotations.Schema.GetColumnOrNull(AnnotationUtils.Kinds.KeyValues);
return
metaColumn != null
&& metaColumn.Value.Type is VectorDataViewType vectorType
&& keyType.Count == (ulong)vectorType.Size
&& keyValueItemType.Equals(vectorType.ItemType);
}

@ganik ganik Feb 21, 2020

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

How about you do something like:

ReadOnlyMemory tmp = default;
if (input.Data.Schema.TryGetAnnotation(TextDataViewType.Instance, AnnotationUtils.Kinds.ScoreValueKind, i, ref tmp) && ReadOnlyMemoryUtils.EqualsStr(AnnotationUtils.Const.ScoreValueKind.PredictedLabel, tmp))
{

In reply to: 382339571 [](ancestors = 382339571)

@ganik ganik Feb 21, 2020

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

More general advice: search for "ScoreValueKind.PredictedLabel" and see how in other places we check for such column

In reply to: 382745605 [](ancestors = 382745605,382339571)

@harishsk harishsk Feb 21, 2020

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

You can look at the implementation of HasKeyValues and filter out the checks that are relying on keyValueItemType . e.g.

            if (!(column.Type.GetItemType() is KeyDataViewType keyType))
                return false;
            var metaColumn = column.Annotations.Schema.GetColumnOrNull(AnnotationUtils.Kinds.KeyValues);
            return
                metaColumn != null
                && metaColumn.Value.Type is VectorDataViewType vectorType
                && keyType.Count == (ulong)vectorType.Size)

In reply to: 382339571 [](ancestors = 382339571)

@antoniovs1029 antoniovs1029 Feb 22, 2020

Copy link
Copy Markdown
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I've decided to go with @harishsk suggestion. Please, @yaeldekel let me know if you know about another way of doing this.

In reply to: 382804341 [](ancestors = 382804341,382339571)

Comment thread
antoniovs1029 marked this conversation as resolved.
Outdated
{
var outputData = new KeyToValueMappingTransformer(Host, "PredictedLabel").Transform(scorePipe);
end = outputData;
transforms.AddLast(outputData as ITransformCanSaveOnnx);
}
}
}
else
{
Expand All @@ -322,6 +333,28 @@ private void Run(IChannel ch)
nameof(Arguments.LoadPredictor), "We were explicitly told to load the predictor but one was not present.");
}

// Convert back to values the KeyDataViewType columns that appear both in input and output
// (i.e those that remained untouched by the model).
var outputNames = new HashSet<string>();
foreach (var col in end.Schema)
if (col.Type is KeyDataViewType && col.IsHidden == false)
outputNames.Add(col.Name);

var inputNames = new HashSet<string>();
foreach (var col in source.Schema)
if (col.Type is KeyDataViewType && col.IsHidden == false)
inputNames.Add(col.Name);

outputNames.IntersectWith(inputNames);

foreach (var name in outputNames)
{
// MYTODO: Add in here any check necessary to see if the column actually has KeyValue Annotations
var outputData = new KeyToValueMappingTransformer(Host, name).Transform(end);
end = outputData;
transforms.AddLast(end as ITransformCanSaveOnnx);
}

var model = ConvertTransformListToOnnxModel(ctx, ch, source, end, transforms, _inputsToDrop, _outputsToDrop);

using (var file = Host.CreateOutputFile(_outputModelPath))
Expand Down
1 change: 0 additions & 1 deletion test/Microsoft.ML.Tests/OnnxConversionTest.cs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -864,7 +864,6 @@ public void LoadingPredictorModelAndOnnxConversionTest()
Done();
}


[Fact]
public void RemoveVariablesInPipelineTest()
{
Expand Down