Introduce VBufferEditor and hide VBuffer.Count #1580

eerhardt · 2018-11-08T18:52:30Z

This PR executes on proposed changes (1) and (5) in #608 (comment).

Keep Length, but hide/private Count

The most drastic proposed change is how to actually mutate a VBuffer.

This introduces VBufferEditor, and uses it in almost all the places where VBuffers are mutated. There are a few stragglers that I will fix in a subsequent PR, but I thought this change was large enough, so I cut it off once I could make .Count private.

Working towards #608.

src/Microsoft.ML.Data/DataView/Transposer.cs

src/Microsoft.ML.Core/Data/MetadataUtils.cs

src/Microsoft.ML.Core/Data/VBuffer.cs

src/Microsoft.ML.Data/Evaluators/RankerEvaluator.cs

TomFinley · 2018-11-10T01:15:29Z

src/Microsoft.ML.Data/Evaluators/EvaluatorUtils.cs

@@ -615,7 +615,7 @@ public static void ReconcileKeyValues(IHostEnvironment env, IDataView[] views, s
            var keyNamesVBuffer = new VBuffer<ReadOnlyMemory<char>>(keyNames.Count, keyNames.Keys.ToArray());
            ValueGetter<VBuffer<ReadOnlyMemory<char>>> keyValueGetter =
                    (ref VBuffer<ReadOnlyMemory<char>> dst) =>
-                        dst = new VBuffer<ReadOnlyMemory<char>>(keyNamesVBuffer.Length, keyNamesVBuffer.Count, keyNamesVBuffer.Values, keyNamesVBuffer.Indices);
+                        keyNamesVBuffer.CopyTo(ref dst);


keyNamesVBuffer [](start = 24, length = 15)

Wow. This was a really bad bug. We shared the underlying output buffers among everyone that could have asked for this!!

The more I read of your PR the more enthusiastic I become. I had thought only of the effort it would save me and others going forward of having to catch all these potential misuses of VBuffer, but now I become alarmed at the sheer volume of stuff I seem to have missed, that will now get caught. :P Or more precisely, made impossible. #Closed

src/Microsoft.ML.Data/Depricated/Vector/VBufferMathUtils.cs

TomFinley · 2018-11-10T01:46:05Z

src/Microsoft.ML.Transforms/GcnTransform.cs

-                                Float mean = Mean(src.Values, src.Count, src.Length);
-                                Float divisor = StdDev(src.Values, src.Count, src.Length, mean);
+                                var srcValues = src.GetValues();
+                                Float mean = Mean(srcValues, src.Length);


Float [](start = 32, length = 5)

Hi @eerhardt, I notice here and elsewhere you've been trying to maintain this Float stuff. This was an artifact of something roughly four years ago where we tried to maintain separate "double" and "single" builds (this of course before we had IDataView and could in principle accomodate either where appropriate), and we've been opportunistically removing them, since the desired final state is we have none of this at all.

By this I mean you should choose whichever is less effort for you. If you find it's causing you effort to keep doing this Float stuff then please don't bother, it is intended to be removed in due time anyway. But if it's more effort to remove it feel free to keep using it. Just whichever you find more helpful to do, is all. #Resolved

Noted. I haven't been going out of my way to change Float => float, but I have made a slightly conscious effort to keep existing code the same. My reasoning was that I didn't want unnecessary changes in my PRs.

In reply to: 232436492 [](ancestors = 232436492)

Makes sense, probably makes merges/rebases easier. #Resolved

TomFinley · 2018-11-10T01:49:37Z

src/Microsoft.ML.TensorFlow/TensorFlow/TensorflowUtils.cs

+                for (int j = 0; j < inputOpsValues.Length; j++)
+                    inputOpsResult[j] = inputOpsValues[j].ToString();
+
+                yield return (name, opType.ToString(), type, inputOpsResult);


This code appears to have a bug in it which has been carried over, that is, it appears to assume that the output vector was dense. I'm not sure you want to fix it or not, but if not, maybe file an issue? #Resolved

It appears that the place that populates this metadata always uses a dense vector. See

machinelearning/src/Microsoft.ML.TensorFlow/TensorFlow/TensorflowUtils.cs

Line 55 in 8ca1c93

dst = new VBuffer<ReadOnlyMemory<char>>(op.NumInputs, inputOps);

.

I can add an assert here, if you'd like. If you want more of a "fix", I can log a bug. Up to you.

In reply to: 232436665 [](ancestors = 232436665)

How about you just call DenseValues() in line 118 and fix the bug?..

In reply to: 232822331 [](ancestors = 232822331,232436665)

TomFinley

Thank you @eerhardt ! I'll take my extra ration of milk now... 😉

Zruty0 · 2018-11-14T22:45:52Z

src/Microsoft.ML.Core/Utilities/Utils.cs

+            for (int i = 0; i < source.Length; i++)
+            {
+                if (predicate(source[i]))
+                {


{ [](start = 16, length = 1)

remove curlies for 1-line clauses like this and below #Resolved

done.

In reply to: 233650638 [](ancestors = 233650638)

Zruty0 · 2018-11-14T22:50:12Z

src/Microsoft.ML.Core/Data/VBufferMutationContext.cs

+        /// with a larger physical value count than was needed
+        /// because the final value count was not known at creation time.
+        /// </param>
+        /// <returns></returns>


[](start = 12, length = 19)

remove empty #Resolved

fixing.

In reply to: 233651783 [](ancestors = 233651783)

Zruty0 · 2018-11-14T22:53:35Z

src/Microsoft.ML.Core/Data/VBufferMutationContext.cs

+        /// because the final value count was not known at creation time.
+        /// </param>
+        /// <returns></returns>
+        public VBuffer<T> CreateBuffer(int? physicalValuesCount = null)


CreateBuffer [](start = 26, length = 12)

should we assert that this only called once? We don't want it to be called multiple times, right?

Also, if we don't, should we call it something like Commit? I can see you already had the name 'Complete' before, and you left it inside a comment. #Resolved

It's not possible to assert it is only called once because VBufferMutationContext<T> is a ref struct (and actually I need to mark it as readonly). Since it is a ref struct, it shouldn't change state because you can get in a lot of trouble passing it into methods as parameters, or making copies of it on the stack.

I can change the name of CreateBuffer, if you'd like. Do you have a suggestion? Do you like Commit more than CreateBuffer? I didn't like Complete, but I can be persuaded.

In reply to: 233652648 [](ancestors = 233652648)

Zruty0 · 2018-11-14T23:58:39Z

src/Microsoft.ML.Core/Data/VBufferMutationContext.cs

+        /// because the final value count was not known at creation time.
+        /// </param>
+        /// <returns></returns>
+        public VBuffer<T> CreateBuffer(int? physicalValuesCount = null)


physicalValuesCount [](start = 44, length = 19)

how do we use this? #Resolved

There are some cases that don't know the final "count" until they are done iterating through a list of mutations. So they allocate "the max necessary" upfront, but may change this value to a smaller count at the end. It isn't often we need this, but there are a handful of places that NEED it.

Here's an example:

machinelearning/src/Microsoft.ML.Data/Transforms/KeyToVectorTransform.cs

Line 599 in 33cac0c

dst = mutation.CreateBuffer(count);

In reply to: 233667829 [](ancestors = 233667829)

Zruty0 · 2018-11-15T00:08:16Z

src/Microsoft.ML.Data/Depricated/Vector/VBufferMathUtils.cs

-                        Array.Clear(values, 0, src.Length);
-                    dst = new VBuffer<Float>(src.Length, values, dst.Indices);
+                    var mutation = VBufferMutationContext.Create(ref dst, src.Length);
+                    if (!mutation.CreatedNewValues) // We need to clear it


t [](start = 73, length = 1)

you lost the period. #Resolved

OOPS. Will fix.

In reply to: 233669776 [](ancestors = 233669776)

Zruty0 · 2018-11-15T00:15:03Z

src/Microsoft.ML.Data/Utilities/SlotDropper.cs

@@ -211,7 +210,7 @@ public void DropSlots<TDst>(ref VBuffer<TDst> src, ref VBuffer<TDst> dst)
                Contracts.Assert(index <= max);
            }

-            dst = new VBuffer<TDst>(newLength, iiDst, values, indices);
+            dst = mutation.CreateBuffer(iiDst);


iiDst [](start = 40, length = 5)

this looks like the only usage of this parameter.
Can we maybe have a separate version of CreateBuffer / Commit? Like, I don't know, CommitTruncated or something? #Resolved

This isn't the only usage of the parameter. I pointed out one above in the other comment. And there are more, like in EnsembleUtils.cs. #Resolved

Zruty0 · 2018-11-15T00:19:02Z

src/Microsoft.ML.Core/Data/VBufferMutationContext.cs

+    /// An object capable of mutation a <see cref="VBuffer{T}"/> by filling out
+    /// <see cref="Values"/> (and <see cref="Indices"/> if the buffer is not dense).
+    /// </summary>
+    public ref struct VBufferMutationContext<T>


VBufferMutationContext [](start = 22, length = 22)

I'm not a big fan of this name. Is the discussion closed on naming?
Given the usage pattern, maybe we can call it VBufferBuilder or VBufferEditor? #Resolved

I'm not a big fan of the name either. I'm open to suggestions.

VBufferBuilder is already taken by BufferBuilder<T>. So I don't think we should go with that.

(BTW: some of these "design" discussions would have been great to have been had up front on #608, instead of in the PR 😝) #Resolved

well, well. I didn't actually understand what you're planning to do until I saw it implemented.

BufferBuilder is internal anyway, right?

In reply to: 233672374 [](ancestors = 233672374)

I would much rather rename BufferBuilder to something ridiculous like 'value accumulator', and have the good name for the more common (and public) use case.

In reply to: 233675246 [](ancestors = 233675246,233672374)

BufferBuilder is internal anyway, right?

It isn't right now.

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

Line 17 in 2d2ed27

public sealed class BufferBuilder<T>

I don't really view this thing as a "builder" anyway. For one there is no "Build" method on it. VBufferEditor would be more appropriate IMO. Do you like that?

VBufferMutation

VBufferMutationContext

VBufferEditor

Any others to consider? #Resolved

Zruty0 · 2018-11-15T00:20:52Z

src/Microsoft.ML.FastTree/BinFile/BinFinder.cs

        {
-            if (values.Count == 0)
+            var valueValues = values.GetValues();


valueValues [](start = 16, length = 11)

valueValues goes against feng shui.
Maybe explicitValues or something? #Resolved

Changed to explicitValues and explicitValuesCount

In reply to: 233672186 [](ancestors = 233672186)

Zruty0 · 2018-11-15T00:24:52Z

src/Microsoft.ML.FastTree/BinFile/BinFinder.cs


            // Get histogram of values
-            Array.Sort(valArray, 0, values.Count);
+            Contracts.Assert(valueBuffer.Length >= valuesCount);
+            valueValues.CopyTo(valueBuffer);


valueValues.CopyTo(valueBuffer); [](start = 12, length = 32)

so, we haven't done this before. And FastTree.cs line 1463 actually assumes that bin finder is destructive. Maybe we should reconcile one way or another?

I'm not following exactly what you want changed here.

I changed this for 2 reasons:

You can no longer get an array from a VBuffer. So you can't call Array.Sort on it.

Breaking the passed in VBuffer is pretty bad IMO. So I came up with a different way that doesn't destroy its usefulness. Upstream, we can get rid of the other copy, once this REVIEW comment is taken care of:

machinelearning/src/Microsoft.ML.FastTree/FastTree.cs

Line 2410 in 057c4b9

// REVIEW: Change this, as well as the bin finding code and bin upper bounds, to be Float instead of Double.

Once those are floats, then you no longer have to copy upstream. You can just pass in the values, and this method won't muck with the VBuffer.

Zruty0 · 2018-11-15T00:25:52Z

src/Microsoft.ML.HalLearners/OlsLinearRegression.cs

@@ -279,8 +279,9 @@ private OlsLinearRegressionPredictor TrainCore(IChannel ch, FloatLabelCursor.Fac
                ch.Check(FloatUtils.IsFinite(beta[i]), "Non-finite values detected in OLS solution");

            var weights = VBufferUtils.CreateDense<float>(beta.Length - 1);
+            var weightsMutation = VBufferMutationContext.CreateFromBuffer(ref weights);


weights [](start = 78, length = 7)

why create a VBuffer and immediately mutate? #Resolved

You mean, why not just create an array, mutate that, and then create the buffer from the array? I can change that. #Resolved

Zruty0 · 2018-11-15T00:30:13Z

src/Microsoft.ML.ImageAnalytics/ImagePixelExtractorTransform.cs

-                        var vf = values as float[];
-                        var vb = values as byte[];
-                        Contracts.Assert(vf != null || vb != null);
+                        Span<float> vf = typeof(TValue) == typeof(float) ? MemoryMarshal.Cast<TValue, float>(mutation.Values) : default;


vf [](start = 36, length = 2)

I feel that a more common pattern is to have 2 getters: one for float case and one for byte case.
Maybe add a // REVIEW: split the getter into 2 specialized getters, one for float case and one for byte case. ? #Resolved

Done.

In reply to: 233673911 [](ancestors = 233673911)

Zruty0 · 2018-11-15T00:58:43Z

src/Microsoft.ML.Transforms/MissingValueDroppingTransformer.cs


            if (newCount == 0)
            {
-                dst = new VBuffer<TDst>(0, dst.Values, dst.Indices);
+                dst = VBufferMutationContext.Create(ref dst, 0)
+                    .CreateBuffer();


Create followed immediately by CreateBuffer seems to be a common pattern too. Maybe add VBufferMutationContext.Resize for that? #Resolved

That method seems like it would belong in VBufferUtils. I will add it there.

In reply to: 233679182 [](ancestors = 233679182)

Zruty0 · 2018-11-15T01:00:12Z

src/Microsoft.ML.Transforms/Microsoft.ML.Transforms.csproj

@@ -4,6 +4,7 @@
    <TargetFramework>netstandard2.0</TargetFramework>
    <IncludeInPackage>Microsoft.ML</IncludeInPackage>
    <DefineConstants>CORECLR</DefineConstants>
+    <AllowUnsafeBlocks>true</AllowUnsafeBlocks>


AllowUnsafeBlocks [](start = 5, length = 17)

just for one GEMV call... that's annoying. #Resolved

actually nevermind, it's going into HalLearners anyway

In reply to: 233679433 [](ancestors = 233679433)

Zruty0 · 2018-11-15T01:06:29Z

src/Microsoft.ML.Transforms/Text/NgramTransform.cs

-                            writer.WriteLine("# Number of Ngrams terms = {0}", ngramsNames.Count);
-                            for (int j = 0; j < ngramsNames.Count; j++)
-                                writer.WriteLine("{0}\t{1}", j, ngramsNames.Values[j]);
+                            var ngramNameValues = ngramsNames.GetValues();


ngramNameValues [](start = 32, length = 15)

explicitNgramNames? #Resolved

fixed.

In reply to: 233680569 [](ancestors = 233680569)

Zruty0 · 2018-11-15T01:07:12Z

src/Microsoft.ML.Transforms/Text/WordTokenizeTransform.cs

-                        var values = dst.Values;
-                        if (terms.Count > 0)
+                        var mutation = VBufferMutationContext.Create(ref dst, terms.Count);
+                        for (int i = 0; i < terms.Count; i++)
                        {


{ [](start = 24, length = 1)

remove #Resolved

Zruty0

…ivate.

…he underlying arrays in the column buffer in Transposer. Also enable a Transposer test, since it passes.

eerhardt requested review from codemzs, Ivanidzo4ka, TomFinley, yaeldekel and Zruty0 November 8, 2018 18:52

eerhardt commented Nov 9, 2018

View reviewed changes

src/Microsoft.ML.Data/DataView/Transposer.cs Show resolved Hide resolved

TomFinley reviewed Nov 9, 2018

View reviewed changes

src/Microsoft.ML.Core/Data/MetadataUtils.cs Show resolved Hide resolved

TomFinley reviewed Nov 10, 2018

View reviewed changes

src/Microsoft.ML.Core/Data/VBuffer.cs Show resolved Hide resolved

TomFinley reviewed Nov 10, 2018

View reviewed changes

src/Microsoft.ML.Data/Evaluators/RankerEvaluator.cs Show resolved Hide resolved

TomFinley reviewed Nov 10, 2018

View reviewed changes

src/Microsoft.ML.Data/Depricated/Vector/VBufferMathUtils.cs Show resolved Hide resolved

TomFinley reviewed Nov 10, 2018

View reviewed changes

eerhardt force-pushed the MasterVBufferPlan branch from 5102889 to 33cac0c Compare November 13, 2018 02:44

TomFinley approved these changes Nov 14, 2018

View reviewed changes

Zruty0 reviewed Nov 14, 2018

View reviewed changes

Zruty0 reviewed Nov 15, 2018

View reviewed changes

Zruty0 approved these changes Nov 15, 2018

View reviewed changes

eerhardt added 11 commits November 15, 2018 15:28

Implement VBuffer master plan WIP #1

4149cdc

Getting everything to build and tests passing

d031fa1

Keep moving to the master plan of VBuffer.

8adb2af

Remove out bools on VBufferMutationContext.Create.

2eff656

Remove the rest of the VBuffer.Count usages in ML.Data

35a0ccc

Remove the rest of the VBuffer.Count usages and make VBuffer.Count pr…

001f7d8

…ivate.

Clean up VBufferMutationContext to match the spec.

01654a7

Fix two failing tests.

d9905f3

Fix FastTreeBinaryClassificationCategoricalSplitTest by remembering t…

4cafe93

…he underlying arrays in the column buffer in Transposer. Also enable a Transposer test, since it passes.

Fix up VectorWhitening after rebase.

1d786ef

Respond to PR feedback.

1cb1411

eerhardt force-pushed the MasterVBufferPlan branch from 0a9747b to 1cb1411 Compare November 15, 2018 21:55

eerhardt added 3 commits November 15, 2018 16:39

Rename CreateBuffer to Commit and CommitTruncated.

28d48a1

Rename VBufferMutationContext to VBufferEditor.

faeb7d4

Rename mutation to editor.

e16ab68

eerhardt merged commit cb9effc into dotnet:master Nov 16, 2018

eerhardt changed the title ~~Introduce VBufferMutationContext and hide VBuffer.Count~~ Introduce VBufferEditor and hide VBuffer.Count Nov 16, 2018

eerhardt mentioned this pull request Nov 16, 2018

Change FastTree BinFinder to use floats and remove one data copy #1648

Open

eerhardt deleted the MasterVBufferPlan branch November 16, 2018 02:43

TomFinley mentioned this pull request Mar 29, 2019

Update VBuffer documentation #3136

Merged

ghost locked as resolved and limited conversation to collaborators Mar 26, 2022

Introduce VBufferEditor and hide VBuffer.Count #1580

Introduce VBufferEditor and hide VBuffer.Count #1580

Conversation

eerhardt commented Nov 8, 2018 • edited Loading

TomFinley Nov 10, 2018 • edited by eerhardt Loading

Choose a reason for hiding this comment

TomFinley Nov 10, 2018 • edited by eerhardt Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

TomFinley Nov 13, 2018 • edited by eerhardt Loading

Choose a reason for hiding this comment

TomFinley Nov 10, 2018 • edited by eerhardt Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

TomFinley left a comment

Choose a reason for hiding this comment

Zruty0 Nov 14, 2018 • edited by eerhardt Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Zruty0 Nov 14, 2018 • edited by eerhardt Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Zruty0 Nov 14, 2018 • edited by eerhardt Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Zruty0 Nov 14, 2018 • edited by eerhardt Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Zruty0 Nov 15, 2018 • edited by eerhardt Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Zruty0 Nov 15, 2018 • edited by eerhardt Loading

Choose a reason for hiding this comment

eerhardt Nov 15, 2018 • edited Loading

Choose a reason for hiding this comment

Zruty0 Nov 15, 2018 • edited by eerhardt Loading

Choose a reason for hiding this comment

eerhardt Nov 15, 2018 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

eerhardt Nov 15, 2018 • edited Loading

Choose a reason for hiding this comment

Zruty0 Nov 15, 2018 • edited by eerhardt 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

Zruty0 Nov 15, 2018 • edited by eerhardt Loading

Choose a reason for hiding this comment

eerhardt Nov 15, 2018 • edited Loading

Choose a reason for hiding this comment

Zruty0 Nov 15, 2018 • edited by eerhardt Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Zruty0 Nov 15, 2018 • edited by eerhardt Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Zruty0 Nov 15, 2018 • edited by eerhardt Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Zruty0 Nov 15, 2018 • edited by eerhardt Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Zruty0 Nov 15, 2018 • edited by eerhardt Loading

Choose a reason for hiding this comment

Zruty0 left a comment

Choose a reason for hiding this comment

eerhardt commented Nov 8, 2018 •

edited

Loading

TomFinley Nov 10, 2018 •

edited by eerhardt

Loading

TomFinley Nov 10, 2018 •

edited by eerhardt

Loading

TomFinley Nov 13, 2018 •

edited by eerhardt

Loading

TomFinley Nov 10, 2018 •

edited by eerhardt

Loading

Zruty0 Nov 14, 2018 •

edited by eerhardt

Loading

Zruty0 Nov 14, 2018 •

edited by eerhardt

Loading

Zruty0 Nov 14, 2018 •

edited by eerhardt

Loading

Zruty0 Nov 14, 2018 •

edited by eerhardt

Loading

Zruty0 Nov 15, 2018 •

edited by eerhardt

Loading

Zruty0 Nov 15, 2018 •

edited by eerhardt

Loading

eerhardt Nov 15, 2018 •

edited

Loading

Zruty0 Nov 15, 2018 •

edited by eerhardt

Loading

eerhardt Nov 15, 2018 •

edited

Loading

eerhardt Nov 15, 2018 •

edited

Loading

Zruty0 Nov 15, 2018 •

edited by eerhardt

Loading

Zruty0 Nov 15, 2018 •

edited by eerhardt

Loading

eerhardt Nov 15, 2018 •

edited

Loading

Zruty0 Nov 15, 2018 •

edited by eerhardt

Loading

Zruty0 Nov 15, 2018 •

edited by eerhardt

Loading

Zruty0 Nov 15, 2018 •

edited by eerhardt

Loading

Zruty0 Nov 15, 2018 •

edited by eerhardt

Loading

Zruty0 Nov 15, 2018 •

edited by eerhardt

Loading