Multicolumn mapping for some estimators #3066

artidoro · 2019-03-22T20:26:28Z

Adding multicolumn mapping for some estimators (as per list by @TomFinley and @glebuk):

OneHotEncodingEstimator
TypeConvertingEstimator
KeyToVectorMappingEstimator
ValueToKeyMappingEstimator
OneHotHashEncodingEstimator
MissingValueEstimator
FeatureSelectionCatalog.*
KeyToValueMappingEstiamtor

Leaving out:

TextFeaturizingEstimator (probably requires column specific settings most of the time)
NoramlizingEstiamtor (in experimental nuget)

Let me know if I should add more estimators.

Fixes #3068
Related to #2884

src/Microsoft.ML.Data/Transforms/ExtensionsCatalog.cs

Ivanidzo4ka · 2019-03-22T20:45:37Z

src/Microsoft.ML.Data/Transforms/ExtensionsCatalog.cs

+        /// <summary>
+        /// Name of the column to transform. If set to <see langword="null"/>, the value of the <see cref="OutputColumnName"/> will be used as source.
+        /// </summary>
+        public readonly string InputColumnName;


public readonly string InputColumnName [](start = 7, length = 39)

I'm slightly confuse.
We get rid of ColumnOptions because they were immutable, and now we add another immutable class... #Resolved

Will invite you to check with Tom about this, I am just executing what he asked

In reply to: 268333577 [](ancestors = 268333577)

Ivanidzo4ka · 2019-03-22T20:47:21Z

src/Microsoft.ML.Data/Transforms/ExtensionsCatalog.cs

+    /// <summary>
+    /// Specifies input and output column names for a transformation.
+    /// </summary>
+    public sealed class InputOutputColumnPair


InputOutputColumnPair [](start = 24, length = 21)

what is difference between this one and ColumnOptions? #Resolved

Not much, I would have used the bellow if it were me, but Tom asked it to be different, for some reason!

In reply to: 268334077 [](ancestors = 268334077)

Yes, a very good reason. ColumnOptions is a struct meant to serve a specific transformer base, and that is involved in the type heirarchy, and in particular something that captured all of the individual settings and state for each mapping. Our goals here were comparatively more modest: we just needed to . Well designed code does what it is designed to do, and in the simplest possible way. There is no need for this to be part of an elaborate type hierarchy -- this was in fact the mistake that led to the issue #2884 being filed, that we'd conflated two distinct techniques for the extremely bad reasoning that they both had to do with "column." (And, of course, if something deals with the same sort of object, obviously they belong in the same type heirarchy, right?)

In reply to: 268334389 [](ancestors = 268334389,268334077)

Just in case, I'm not talking about ColumnOptions inside Estimator, (for example HashingEstimator.ColumnOptions.
I'm talking about ColumnOptions in this exact file few lines below (line 41)
Only difference I see is Input/output column names is public in this one instead of private in other, and name of class.
Ok, two difference, one below has implicit converter from tuple.

Can we delete ColumnOptions in this file?

In reply to: 268337035 [](ancestors = 268337035,268334389,268334077)

That's fine. What Artidoro and I had discussed was actually somewhat different. (Or we were talking about two separate things without realizing it.)

In reply to: 268339442 [](ancestors = 268339442,268337035,268334389,268334077)

I guess we were talking about something different, glad we are on the same page.

In reply to: 268760195 [](ancestors = 268760195,268339442,268337035,268334389,268334077)

This seems good to me. The ColumnOptions I had been talking about (which is to say, the vast majority of things with that name) are as I said meant to serve a different purpose.

We should probably rename them back to ColumnInfo at some point but this can be delayed as they are internal... ummm except one. Whoops. Opened #3078. :) Aside from that, yeah.

In reply to: 268771510 [](ancestors = 268771510,268760195,268339442,268337035,268334389,268334077)

artidoro · 2019-03-22T22:59:52Z

Since @ivanbasov asked for it the third commit shows the how this would look like if we removed the ColumnOptions class, and used InputOutputColumnPair.

Let me know what looks better! I will either eliminate this commit or keep it.

Notice that even in commit 3 InputOutputColumnPair is never used inside any transform. #Resolved

codecov · 2019-03-22T23:50:10Z

Codecov Report

Merging #3066 into master will decrease coverage by 0.02%.
The diff coverage is 60.11%.

@@            Coverage Diff             @@
##           master    #3066      +/-   ##
==========================================
- Coverage   72.53%   72.51%   -0.03%     
==========================================
  Files         806      806              
  Lines      144282   144642     +360     
  Branches    16183    16197      +14     
==========================================
+ Hits       104661   104889     +228     
- Misses      35217    35342     +125     
- Partials     4404     4411       +7

Flag	Coverage Δ
#Debug	`72.51% <60.11%> (-0.03%)`	⬇️
#production	`68.11% <41.59%> (-0.05%)`	⬇️
#test	`88.8% <98.18%> (+0.04%)`	⬆️

Impacted Files	Coverage Δ
src/Microsoft.ML.Transforms/Text/TextCatalog.cs	`41.66% <0%> (-3.79%)`	⬇️
src/Microsoft.ML.Transforms/NormalizerCatalog.cs	`36.36% <0%> (-35.07%)`	⬇️
...c/Microsoft.ML.ImageAnalytics/ExtensionsCatalog.cs	`11.11% <0%> (-18.89%)`	⬇️
...icrosoft.ML.Tests/Transformers/CategoricalTests.cs	`100% <100%> (ø)`	⬆️
...sts/Transformers/KeyToBinaryVectorEstimatorTest.cs	`100% <100%> (ø)`	⬆️
...Microsoft.ML.Tests/Transformers/NormalizerTests.cs	`100% <100%> (ø)`	⬆️
...oft.ML.Tests/Transformers/FeatureSelectionTests.cs	`100% <100%> (ø)`	⬆️
...ios/IrisPlantClassificationWithStringLabelTests.cs	`98.63% <100%> (ø)`	⬆️
...icrosoft.ML.Tests/Transformers/NAIndicatorTests.cs	`100% <100%> (ø)`	⬆️
...crosoft.ML.Tests/Transformers/ValueMappingTests.cs	`100% <100%> (ø)`	⬆️
... and 30 more

glebuk · 2019-03-25T16:10:45Z

src/Microsoft.ML.Data/Transforms/ConversionsExtensionsCatalog.cs

+        /// <param name="columns">Specifies the names of the columns on which to apply the transformation.</param>
+        /// <param name="outputKind">The expected kind of the output column.</param>
+        public static TypeConvertingEstimator ConvertType(this TransformsCatalog.ConversionTransforms catalog,
+            InputOutputColumnPair[] columns,


columns [](start = 36, length = 7)

Don't you need to check for null for the columns arg to avoid null reference exception? #Resolved

I added the checks thanks for pointing out. I only fixed the extensions that are public. Not those that are internal.

In reply to: 268724517 [](ancestors = 268724517)

glebuk · 2019-03-25T16:13:00Z

src/Microsoft.ML.Data/Transforms/ConversionsExtensionsCatalog.cs

-             => new KeyToValueMappingEstimator(CatalogUtils.GetEnvironment(catalog), ColumnOptions.ConvertToValueTuples(columns));
+        /// <param name="catalog">The conversion transform's catalog.</param>
+        /// <param name="columns">Specifies the names of the columns on which to apply the transformation.</param>
+        public static KeyToValueMappingEstimator MapKeyToValue(this TransformsCatalog.ConversionTransforms catalog, InputOutputColumnPair[] columns)


columns [](start = 140, length = 7)

also check for null #Resolved

glebuk · 2019-03-25T16:15:24Z

src/Microsoft.ML.Data/Transforms/ExtensionsCatalog.cs

-        /// Instantiates a <see cref="ColumnOptions"/> from a tuple of input and output column names.
-        /// </summary>
-        public static implicit operator ColumnOptions((string outputColumnName, string inputColumnName) value)
+        public InputOutputColumnPair(string outputColumnName, string inputColumnName = null)


public InputOutputColumnPair(string outputColumnName, string inputColumnName = null) [](start = 8, length = 84)

Add another overload for the case when input name = output name. That would make it a lot clearer vs setting one to null. #ByDesign

Unfortunately I believe that everywhere in the codebase we have the same pattern string outputColumnName, string inputColumnName = null. This is found in all the mlContext extensions for transforms. I can add it here, but I think it would make more sense to stick to the general pattern.

In reply to: 268726871 [](ancestors = 268726871)

We must be consistent. Sorry @glebuk!

In reply to: 268779384 [](ancestors = 268779384,268726871)

Ok then.

In reply to: 268819893 [](ancestors = 268819893,268779384,268726871)

glebuk · 2019-03-25T16:17:12Z

src/Microsoft.ML.Data/Transforms/ExtensionsCatalog.cs

+        /// <summary>
+        /// Name of the column to transform. If set to <see langword="null"/>, the value of the <see cref="OutputColumnName"/> will be used as source.
+        /// </summary>
+        public readonly string InputColumnName;


InputColumnName [](start = 31, length = 15)

Shouldn't it be in reverse - the input be set, but output be optional and equal to input if out is null?? #Closed

As per #2064 we use the outputColumnName as inputColumnName when inputColumnName is null. We are doing this across the code base.

In reply to: 268727730 [](ancestors = 268727730)

glebuk · 2019-03-25T16:19:05Z

src/Microsoft.ML.Transforms/CategoricalCatalog.cs

+                ValueToKeyMappingEstimator.KeyOrdinality keyOrdinality = ValueToKeyMappingEstimator.Defaults.Ordinality,
+                IDataView keyData = null)
+        {
+            var columnOptions = columns.Select(x => new OneHotEncodingEstimator.ColumnOptions(x.OutputColumnName, x.InputColumnName, outputKind, maximumNumberOfKeys, keyOrdinality)).ToArray();


columns [](start = 32, length = 7)

check for null here and elsewhere #Resolved

glebuk · 2019-03-25T16:21:28Z

test/Microsoft.ML.Tests/Transformers/FeatureSelectionTests.cs

-                        ("out1", "VectorFloat"),
-                        ("out2", "VectorDouble")
+                    columns: new[] {
+                        new InputOutputColumnPair("out1", "VectorFloat"),


InputOutputColumnPair [](start = 28, length = 21)

why do we have to use the more verbose initializer here? Ideally we should use the old syntax is possible as it is a lot more compact. #Resolved

Because the other one uses tuples. We have decided not to have tuples in the public surface any longer.

In reply to: 268729946 [](ancestors = 268729946)

glebuk · 2019-03-25T16:22:07Z

    public void ValueMappingValueTypeIsVectorWorkout()

Add a test for when InputOutputColumnPair is null.

Refers to: test/Microsoft.ML.Tests/Transformers/ValueMappingTests.cs:523 in 802e4de. [](commit_id = 802e4de, deletion_comment = False)

TomFinley · 2019-03-25T17:22:43Z

src/Microsoft.ML.Data/Transforms/ExtensionsCatalog.cs

-        /// Instantiates a <see cref="ColumnOptions"/> from a tuple of input and output column names.
-        /// </summary>
-        public static implicit operator ColumnOptions((string outputColumnName, string inputColumnName) value)
+        public InputOutputColumnPair(string outputColumnName, string inputColumnName = null)


outputColumnName [](start = 44, length = 16)

Check non-empty on outputColumName probably. #Resolved

TomFinley · 2019-03-25T17:22:54Z

src/Microsoft.ML.Data/Transforms/ExtensionsCatalog.cs

+        /// <summary>
+        /// Name of the column to transform. If set to <see langword="null"/>, the value of the <see cref="OutputColumnName"/> will be used as source.
+        /// </summary>
+        public readonly string InputColumnName;


InputColumnName [](start = 31, length = 15)

Should these be properties? #Resolved

Ok making these properties!

In reply to: 268761458 [](ancestors = 268761458)

TomFinley

Thank you for working on this @artidoro !! I see the central thing still isn't using properties, but I am not certain that is absolutely essential. Might be nice though if you get to it.

Ivanidzo4ka

artidoro · 2019-03-25T21:05:03Z

Thank you for reviewing, I am updating now with the latest changes.

artidoro force-pushed the multicolumn branch 3 times, most recently from d08a2f9 to 0896590 Compare March 22, 2019 20:39

artidoro changed the title ~~WIP: Multicolumn mapping for some estimators~~ Multicolumn mapping for some estimators Mar 22, 2019

artidoro requested review from sfilipi, Ivanidzo4ka, TomFinley and abgoswam March 22, 2019 20:39

artidoro self-assigned this Mar 22, 2019

artidoro added this to the 0319 milestone Mar 22, 2019

Ivanidzo4ka reviewed Mar 22, 2019

View reviewed changes

src/Microsoft.ML.Data/Transforms/ExtensionsCatalog.cs Outdated Show resolved Hide resolved

artidoro force-pushed the multicolumn branch from 0896590 to a3419f4 Compare March 22, 2019 20:45

Ivanidzo4ka reviewed Mar 22, 2019

View reviewed changes

artidoro added 2 commits March 22, 2019 15:56

multicolumn mapping for some estimators

6207a54

review comment

b4c3ea0

artidoro force-pushed the multicolumn branch from 5272c0c to 0be49b8 Compare March 22, 2019 22:57

merging ML.ColumnOptions and ML.InputOutputColumnPair

802e4de

artidoro force-pushed the multicolumn branch from 0be49b8 to 802e4de Compare March 22, 2019 23:14

glebuk reviewed Mar 25, 2019

View reviewed changes

TomFinley reviewed Mar 25, 2019

View reviewed changes

TomFinley approved these changes Mar 25, 2019

View reviewed changes

Ivanidzo4ka approved these changes Mar 25, 2019

View reviewed changes

review comments, adding check for non null

4e14295

artidoro merged commit 5f9be36 into dotnet:master Mar 25, 2019

ghost locked as resolved and limited conversation to collaborators Mar 23, 2022

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Multicolumn mapping for some estimators #3066

Multicolumn mapping for some estimators #3066

artidoro commented Mar 22, 2019 •

edited

Loading

Ivanidzo4ka Mar 22, 2019 •

edited by artidoro

Loading

artidoro Mar 22, 2019

Ivanidzo4ka Mar 22, 2019 •

edited by artidoro

Loading

artidoro Mar 22, 2019

TomFinley Mar 22, 2019

Ivanidzo4ka Mar 22, 2019

TomFinley Mar 25, 2019

artidoro Mar 25, 2019

TomFinley Mar 25, 2019

artidoro commented Mar 22, 2019 •

edited

Loading

codecov bot commented Mar 22, 2019 •

edited

Loading

glebuk Mar 25, 2019 •

edited by artidoro

Loading

artidoro Mar 25, 2019

glebuk Mar 25, 2019 •

edited by artidoro

Loading

glebuk Mar 25, 2019 •

edited

Loading

artidoro Mar 25, 2019

TomFinley Mar 25, 2019

glebuk Mar 25, 2019

glebuk Mar 25, 2019 •

edited

Loading

artidoro Mar 25, 2019

glebuk Mar 25, 2019 •

edited by artidoro

Loading

glebuk Mar 25, 2019 •

edited by artidoro

Loading

artidoro Mar 25, 2019

glebuk commented Mar 25, 2019

TomFinley Mar 25, 2019 •

edited by artidoro

Loading

TomFinley Mar 25, 2019 •

edited by artidoro

Loading

artidoro Mar 25, 2019

TomFinley left a comment •

edited

Loading

Ivanidzo4ka left a comment

artidoro commented Mar 25, 2019

Multicolumn mapping for some estimators #3066

Multicolumn mapping for some estimators #3066

Conversation

artidoro commented Mar 22, 2019 • edited Loading

Ivanidzo4ka Mar 22, 2019 • edited by artidoro Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Ivanidzo4ka Mar 22, 2019 • edited by artidoro Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

artidoro commented Mar 22, 2019 • edited Loading

codecov bot commented Mar 22, 2019 • edited Loading

Codecov Report

glebuk Mar 25, 2019 • edited by artidoro Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

glebuk Mar 25, 2019 • edited by artidoro Loading

Choose a reason for hiding this comment

glebuk Mar 25, 2019 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

glebuk Mar 25, 2019 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

glebuk Mar 25, 2019 • edited by artidoro Loading

Choose a reason for hiding this comment

glebuk Mar 25, 2019 • edited by artidoro Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

glebuk commented Mar 25, 2019

TomFinley Mar 25, 2019 • edited by artidoro Loading

Choose a reason for hiding this comment

TomFinley Mar 25, 2019 • edited by artidoro Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

TomFinley left a comment • edited Loading

Choose a reason for hiding this comment

Ivanidzo4ka left a comment

Choose a reason for hiding this comment

artidoro commented Mar 25, 2019

artidoro commented Mar 22, 2019 •

edited

Loading

Ivanidzo4ka Mar 22, 2019 •

edited by artidoro

Loading

Ivanidzo4ka Mar 22, 2019 •

edited by artidoro

Loading

artidoro commented Mar 22, 2019 •

edited

Loading

codecov bot commented Mar 22, 2019 •

edited

Loading

glebuk Mar 25, 2019 •

edited by artidoro

Loading

glebuk Mar 25, 2019 •

edited by artidoro

Loading

glebuk Mar 25, 2019 •

edited

Loading

glebuk Mar 25, 2019 •

edited

Loading

glebuk Mar 25, 2019 •

edited by artidoro

Loading

glebuk Mar 25, 2019 •

edited by artidoro

Loading

TomFinley Mar 25, 2019 •

edited by artidoro

Loading

TomFinley Mar 25, 2019 •

edited by artidoro

Loading

TomFinley left a comment •

edited

Loading