Scrub Fast Tree Family #2753

wschin · 2019-02-27T01:37:10Z

This PR fixes #2617 (also fixes #2375, fixes #2857) and is also a part of #2613.

Cleaned items:

FastTree.cs
FastTreeArguments.cs
Derived fast tree classes
Gam family

wschin · 2019-02-27T01:38:13Z

docs/samples/Microsoft.ML.Samples/Dynamic/FastTreeRegression.cs

-            // Use the path object to pass to GetLeaf, which will populate path with the IDs of th nodes from root to leaf.
-            List<int> path = default;
-            // Get the ID of the leaf this example ends up in tree 0.
-            var leafID = modelParams.GetLeaf(0, in testRow, ref path);


Removed because our public APIs doesn't provide those functions. #Resolved

But you can still traverse through trees with new structure, right?

In reply to: 260562343 [](ancestors = 260562343)

No, we cannot but this definitely should be hidden here. @[email protected], do you think this feature is required in tree's public APIs?

In reply to: 261011054 [](ancestors = 261011054,260562343)

I think the trees should enable someone to explain what the tree is. I wouldn't make evaluation code part of the API. Any more than we insist that linear model parameters expose "do dot product" as a method on the model parameters. Both things are potentially useful of course, but we don't need it as part of our public API (at least at first).

In reply to: 261417026 [](ancestors = 261417026,261011054,260562343)

Sounds good. I feel evaluation module should be independent to the expression of the trained objects. Thanks you!

In reply to: 261734289 [](ancestors = 261734289,261417026,261011054,260562343)

Ivanidzo4ka · 2019-02-27T01:40:27Z

src/Microsoft.ML.FastTree/FastTreeArguments.cs

-        [Argument(ArgumentType.LastOccurenceWins, HelpText = "The seed of the random number generator", ShortName = "r1")]
-        public int RngSeed = 123;
+        [Argument(ArgumentType.LastOccurenceWins, HelpText = "The seed of the random number generator", ShortName = "r1,RngSeed")]
+        public int RandomSeed = 123;


RandomSeed [](start = 19, length = 10)

I think everywhere else we call it just seed. #Resolved

Sure.

In reply to: 260562756 [](ancestors = 260562756)

Ivanidzo4ka · 2019-02-27T01:40:56Z

src/Microsoft.ML.FastTree/FastTreeArguments.cs

        [TGUI(NotGui = true)]
-        public int FeatureSelectSeed = 123;
+        public int FeatureSelectionRandomSeed = 123;


FeatureSelectionRandomSeed [](start = 19, length = 26)

Do we really need two of them? #Resolved

Kind of. There are two degrees of control that internal people wanted to independently vary or not vary: selection of features and selection of data. I have not kept touch with this set of people over the years but the reasoning at one point was definitely considered important. Let us not change it. #Resolved

Yes sir.

In reply to: 260863640 [](ancestors = 260863640)

Ivanidzo4ka · 2019-02-27T01:41:32Z

src/Microsoft.ML.FastTree/FastTreeArguments.cs

-        [Argument(ArgumentType.LastOccurenceWins, HelpText = "Minimum categorical docs percentage in a bin to consider for a split.", ShortName = "mdop")]
-        public double MinDocsPercentageForCategoricalSplit = 0.001;
+        [Argument(ArgumentType.LastOccurenceWins, HelpText = "Minimum categorical docs percentage in a bin to consider for a split.", ShortName = "mdop,MinDocsPercentageForCategoricalSplit")]
+        public double MinExamplePercentageForCategoricalSplit = 0.001;


Min [](start = 22, length = 3)

Minimum #Resolved

Ivanidzo4ka · 2019-02-27T01:42:39Z

src/Microsoft.ML.FastTree/FastTreeArguments.cs

-        [Argument(ArgumentType.LastOccurenceWins, HelpText = "Minimum categorical docs percentage in a bin to consider for a split.", ShortName = "mdop")]
-        public double MinDocsPercentageForCategoricalSplit = 0.001;
+        [Argument(ArgumentType.LastOccurenceWins, HelpText = "Minimum categorical docs percentage in a bin to consider for a split.", ShortName = "mdop,MinDocsPercentageForCategoricalSplit")]
+        public double MinExamplePercentageForCategoricalSplit = 0.001;


0.001 [](start = 64, length = 5)

is it 0.001% or 0.1% ? #Resolved

0.1%. Docs are updated to all Percentage and Fraction arguments.

In reply to: 260563129 [](ancestors = 260563129)

wschin · 2019-02-27T01:42:51Z

src/Microsoft.ML.FastTree/FastTree.cs

+        /// <summary>
+        /// Copy the weights of all training features to <paramref name="weights"/>.
+        /// </summary>
+        public void GetFeatureWeights(ref VBuffer<float> weights)


Not sure if this should return IReadOnlyList<float>.

Ivanidzo4ka · 2019-02-27T01:43:53Z

src/Microsoft.ML.FastTree/FastTreeArguments.cs

-        [Argument(ArgumentType.LastOccurenceWins, HelpText = "Maximum number of distinct values (bins) per feature", ShortName = "mb")]
-        public int MaxBins = 255;  // save one for undefs
+        [Argument(ArgumentType.LastOccurenceWins, HelpText = "Maximum number of distinct values (bins) per feature", ShortName = "mb,MaxBins")]
+        public int MaxBinCountPerFeature = 255;  // save one for undefs


Max [](start = 19, length = 3)

Maximum #Resolved

Ivanidzo4ka · 2019-02-27T01:44:38Z

    /// Print execution time breakdown to stdout.

are we really print into stdout or we use our logging mechanism? #Resolved

Refers to: src/Microsoft.ML.FastTree/FastTreeArguments.cs:301 in 3257470. [](commit_id = 3257470, deletion_comment = False)

Ivanidzo4ka · 2019-02-27T01:45:48Z

src/Microsoft.ML.FastTree/FastTreeArguments.cs

-        [Argument(ArgumentType.AtMostOnce, HelpText = "The fraction of features (chosen randomly) to use on each split", ShortName = "sf")]
-        public Double SplitFraction = 1;
+        [Argument(ArgumentType.AtMostOnce, HelpText = "The fraction of features (chosen randomly) to use on each split", ShortName = "sf,SplitFraction")]
+        public Double FeatureFractionPerSplit = 1;

        /// <summary>
        /// Smoothing paramter for tree regularization.


paramter [](start = 22, length = 8)

parameter #Resolved

Ivanidzo4ka · 2019-02-27T01:46:58Z

    public bool PrintTrainValidGraph;

is this remnants of TLC TV ?
Do we need to expose them? #Resolved

Refers to: src/Microsoft.ML.FastTree/FastTreeArguments.cs:410 in d64a4ae. [](commit_id = d64a4ae, deletion_comment = False)

TomFinley · 2019-02-27T17:39:55Z

    public bool PrintTrainValidGraph;
is this remnants of TLC TV ?
Do we need to expose them?

Refers to: src/Microsoft.ML.FastTree/FastTreeArguments.cs:410 in d64a4ae. [](commit_id = d64a4ae, deletion_comment = False)

It has nothing to do with TLC TV. It is however important for command line applications, less so for API. Let us make internal for now. #Resolved

Ivanidzo4ka · 2019-02-28T01:10:31Z

    public RegressionTreeEnsemble TrainedTreeEnsemble { get; }

I've looked into RegressionTree.cs and some things like:
NumLeaves, NumNodes, GtChild, LteChild doesn't sound like great names
Shall we address that in this PR as well? #Resolved

Refers to: src/Microsoft.ML.FastTree/FastTree.cs:3393 in 3257470. [](commit_id = 3257470, deletion_comment = False)

wschin · 2019-02-28T23:14:39Z

    public RegressionTreeEnsemble TrainedTreeEnsemble { get; }

Fixed as well. Feel strange if they are in two PRs.

In reply to: 468095313 [](ancestors = 468095313)

Refers to: src/Microsoft.ML.FastTree/FastTree.cs:3393 in 3257470. [](commit_id = 3257470, deletion_comment = False)

wschin · 2019-02-28T23:40:38Z

    public bool PrintTrainValidGraph;

I will do internal-best-friend.

In reply to: 467689803 [](ancestors = 467689803)

Refers to: src/Microsoft.ML.FastTree/FastTreeArguments.cs:410 in d64a4ae. [](commit_id = d64a4ae, deletion_comment = False)

wschin · 2019-02-28T23:41:03Z

Sure.

In reply to: 467958748 [](ancestors = 467958748)

wschin · 2019-02-28T23:46:06Z

    /// Print execution time breakdown to stdout.

We use ch defined in

           using (var ch = Host.Start("Training"))
            {
                ch.CheckValue(trainData, nameof(trainData));
                trainData.CheckBinaryLabel();
                trainData.CheckFeatureFloatVector();
                trainData.CheckOptFloatWeight();
                FeatureCount = trainData.Schema.Feature.Value.Type.GetValueCount();
                ConvertData(trainData);
                TrainCore(ch);
            }

In reply to: 467689294 [](ancestors = 467689294)

Refers to: src/Microsoft.ML.FastTree/FastTreeArguments.cs:301 in d64a4ae. [](commit_id = d64a4ae, deletion_comment = False)

singlis · 2019-03-01T01:57:24Z

src/Microsoft.ML.Data/Prediction/Calibrator.cs

@@ -14,7 +14,6 @@
 using Microsoft.ML.CommandLine;
 using Microsoft.ML.Data;
 using Microsoft.ML.EntryPoints;
-using Microsoft.ML.Internal.Internallearn;


Are you merged with master? I thought that InternalLearn was removed in a prior change. #Resolved

Yep, maybe rebase did something.

In reply to: 261456781 [](ancestors = 261456781)

singlis · 2019-03-01T01:59:24Z

            bestIteration = Ensemble.NumTrees;

Changing this to NumberOfTrees is probably out of scope for this PR right? #Resolved

Refers to: src/Microsoft.ML.FastTree/BoostingFastTree.cs:143 in 3257470. [](commit_id = 3257470, deletion_comment = False)

wschin · 2019-03-01T02:00:10Z

            bestIteration = Ensemble.NumTrees;

Let me double-check. Thanks!

In reply to: 468512482 [](ancestors = 468512482)

Refers to: src/Microsoft.ML.FastTree/BoostingFastTree.cs:143 in 3257470. [](commit_id = 3257470, deletion_comment = False)

singlis · 2019-03-01T02:05:05Z

        string featureColumn,

suggestion:
I know this constructor is not exposed publicly, but I renamed these parameters (featureColumn->featureColumnName, etc) in lightgbm so that they aligned with the parameters that are used in the catalog function. And technically they will align with the parameters in the Options class. #Resolved

Refers to: src/Microsoft.ML.FastTree/FastTree.cs:104 in 3257470. [](commit_id = 3257470, deletion_comment = False)

singlis · 2019-03-01T02:17:20Z

src/Microsoft.ML.FastTree/FastTree.cs

@@ -3230,6 +3231,9 @@ private void ToCSharp(InternalRegressionTree tree, TextWriter writer, int node,
            }
        }

+        /// <summary>
+        /// Copy the weights of all training features to <paramref name="weights"/>.
+        /// </summary>


</ [](start = 12, length = 2)

Missing pararm for weights #Resolved

Thanks. I will add

/// <param name="weights">a <see cref="VBuffer{T}"/> where feature weights would be assigned to. /// The i-th element in <paramref name="weights"/> stores the weight of the i-th feature.</param>

In reply to: 261459910 [](ancestors = 261459910)

singlis · 2019-03-01T02:28:52Z

src/Microsoft.ML.FastTree/FastTreeClassification.cs

-            int numTrees = Defaults.NumTrees,
-            int minDatapointsInLeaves = Defaults.MinDocumentsInLeaves,
-            double learningRate = Defaults.LearningRates)
+            int numLeaves = Defaults.NumberOfLeaves,


numLeaves [](start = 16, length = 9)

same suggestion here to use the full names for the arguments. #Resolved

Sure.

In reply to: 261461521 [](ancestors = 261461521)

singlis · 2019-03-01T02:40:02Z

    internal BinaryClassificationGamTrainer(IHostEnvironment env, Options options)

Should this be renamed to GamBinaryClassificationTrainer? I know there is an issue that is tracking consistent naming with our algorithms, so probably not worth doing in this pr.
#WontFix

Refers to: src/Microsoft.ML.FastTree/GamClassification.cs:52 in 3257470. [](commit_id = 3257470, deletion_comment = False)

singlis · 2019-03-01T02:46:07Z

        /// wrong, this method will throw.

I dont think this comment is right - this doesnt look like it will throw. #Resolved

Refers to: src/Microsoft.ML.FastTree/GamModelParameters.cs:865 in 3257470. [](commit_id = 3257470, deletion_comment = False)

singlis · 2019-03-01T02:53:28Z

src/Microsoft.ML.StaticPipe/TreeTrainersStatic.cs

-            int numTrees = Defaults.NumTrees,
-            int minDatapointsInLeaves = Defaults.MinDocumentsInLeaves,
-            double learningRate = Defaults.LearningRates,
+            int numLeaves = Defaults.NumberOfLeaves,


numLeaves [](start = 16, length = 9)

Shouldnt these parameters also be fully named? #Resolved

Thanks! I didn't realize they are FastTree static APIs.

In reply to: 261464899 [](ancestors = 261464899)

Ivanidzo4ka · 2019-03-02T06:22:07Z

src/Microsoft.ML.FastTree/RegressionTree.cs

@@ -63,14 +63,14 @@ public abstract class RegressionTreeBase
        /// bitwise complement operator in C#; for details, see
        /// https://docs.microsoft.com/en-us/dotnet/csharp/language-reference/operators/bitwise-complement-operator.
        /// </summary>
-        public IReadOnlyList<int> LteChild => _lteChild;
+        public IReadOnlyList<int> LessThanOrEqualToThresholdChildren => _lteChild;


LessThanOrEqualToThresholdChildren [](start = 34, length = 34)

It looks like across lightgbm, catboost and whatever they use in scikit this one called: leftChildrends, or leftChild, another one is rightChild. #Resolved

Will do leftChild for this.

In reply to: 261814959 [](ancestors = 261814959)

wschin · 2019-03-04T16:37:47Z

        public int LambdaMartMaxTruncation = 100;

I think so. Can we remove "k" in top-k accuracy? No, we can't.

In reply to: 468890315 [](ancestors = 468890315)

Refers to: src/Microsoft.ML.FastTree/FastTreeArguments.cs:93 in ee67d50. [](commit_id = ee67d50, deletion_comment = False)

wschin · 2019-03-04T21:10:50Z

        // REVIEW: Hiding sorting for now. Should be an enum or component factory.

Will make it internal.

In reply to: 468890113 [](ancestors = 468890113)

Refers to: src/Microsoft.ML.FastTree/FastTreeArguments.cs:83 in ee67d50. [](commit_id = ee67d50, deletion_comment = False)

wschin · 2019-03-04T21:11:58Z

        public bool DistanceWeight2;

As suggested by @[email protected], this will be internal.

In reply to: 468890167 [](ancestors = 468890167)

Refers to: src/Microsoft.ML.FastTree/FastTreeArguments.cs:105 in ee67d50. [](commit_id = ee67d50, deletion_comment = False)

wschin · 2019-03-04T21:12:33Z

        public string CustomGains = "0,3,7,15,31";

Will be included in the next iteration. (iter 9)

In reply to: 468890070 [](ancestors = 468890070)

Refers to: src/Microsoft.ML.FastTree/FastTreeArguments.cs:78 in ee67d50. [](commit_id = ee67d50, deletion_comment = False)

Ivanidzo4ka · 2019-03-04T23:51:52Z

src/Microsoft.ML.FastTree/FastTreeArguments.cs

@@ -129,12 +273,12 @@ internal override void Check(IExceptionContext ectx)
 #if OLD_DATALOAD
                ectx.CheckUserArg(0 <= secondaryMetricShare && secondaryMetricShare <= 1, "secondaryMetricShare", "secondaryMetricShare must be between 0 and 1.");
 #endif
-                ectx.CheckUserArg(0 < LambdaMartMaxTruncation, nameof(LambdaMartMaxTruncation), "lambdaMartMaxTruncation must be positive.");
+                ectx.CheckUserArg(0 < NdcgTruncationLevel, nameof(NdcgTruncationLevel), "lambdaMartMaxTruncation must be positive.");


lambdaMartMaxTruncation [](start = 89, length = 23)

no need for this one. you already use nameof #Resolved

Ivanidzo4ka · 2019-03-04T23:58:32Z

src/Microsoft.ML.FastTree/FastTreeArguments.cs

+            /// <summary>
+            /// Early stopping metrics.
+            /// </summary>
+            public EarlyStoppingMetric EarlyStoppingMetric


EarlyStoppingMetric [](start = 19, length = 19)

From comments for previous option it looks like it should be applicable only for regression and ranking problem. Does it actually works for binary classification? #Pending

I don't exactly know but conceptually yes. It can be used to measure gradient norm.

In reply to: 262295829 [](ancestors = 262295829)

Ivanidzo4ka · 2019-03-05T00:02:59Z

    public int EarlyStoppingMetrics;

internal. otherwise i would see two places in which I can set this up. this one and EarlyStoppingMetric #Resolved

Refers to: src/Microsoft.ML.FastTree/FastTreeArguments.cs:633 in 157aedc. [](commit_id = 157aedc, deletion_comment = False)

Ivanidzo4ka · 2019-03-05T00:10:30Z

    public string BaselineScoresFormula;

make this one internal and next two below it.
In order to work with FreeForms we need to ship "NeuralTreeEvaluator.dll" which we can't expose. #Resolved

Refers to: src/Microsoft.ML.FastTree/FastTreeArguments.cs:737 in 157aedc. [](commit_id = 157aedc, deletion_comment = False)

ganik · 2019-03-05T20:20:37Z

test/BaselineOutput/Common/EntryPoints/core_manifest.json

@@ -9096,18 +9057,6 @@
          "IsNullable": false,
          "Default": false
        },
-        {
-          "Name": "MaxTreesAfterCompression",
-          "Type": "Int",


why this got deleted? #Resolved

Not used at all.

In reply to: 262674075 [](ancestors = 262674075)

…otnet#2613

Ivanidzo4ka

wschin added the API Issues pertaining the friendly API label Feb 27, 2019

wschin self-assigned this Feb 27, 2019

wschin commented Feb 27, 2019

View reviewed changes

Ivanidzo4ka reviewed Feb 27, 2019

View reviewed changes

wschin commented Feb 27, 2019

View reviewed changes

Ivanidzo4ka reviewed Feb 27, 2019

View reviewed changes

wschin force-pushed the scrub-tree branch from d8957fd to aa0b9c2 Compare February 27, 2019 20:59

wschin changed the title ~~[WIP] Scrub Fast Tree Family~~ Scrub Fast Tree Family Feb 28, 2019

wschin force-pushed the scrub-tree branch from eaaa8f4 to e6aa43d Compare March 1, 2019 00:07

singlis reviewed Mar 1, 2019

View reviewed changes

Ivanidzo4ka reviewed Mar 2, 2019

View reviewed changes

Ivanidzo4ka reviewed Mar 4, 2019

View reviewed changes

ganik reviewed Mar 5, 2019

View reviewed changes

Ivanidzo4ka mentioned this pull request Mar 5, 2019

GAM parameter names are non-standard #2857

Closed

wschin added 12 commits March 5, 2019 13:20

Clean FastTree.cs and FastTreeArguments.cs following instructions in d…

38e03c4

…otnet#2613

Go through all places once

099768d

Address comments

b90dcd3

Address comments and update entry-point catalog

ab58a46

Update F# tests

01d6830

Address comments

f323b00

Address comments

8d27bab

Remove extra short names for entry point BC

dd3c8a2

Address some comments

e9745b6

Improve early-stopping metric field

8e93762

Update entry points

004d4cc

Address comments

1b4474c

wschin force-pushed the scrub-tree branch from 157aedc to 1b4474c Compare March 5, 2019 21:21

Ivanidzo4ka approved these changes Mar 5, 2019

View reviewed changes

wschin merged commit 95019e5 into dotnet:master Mar 5, 2019

wschin deleted the scrub-tree branch March 5, 2019 22:18

Ivanidzo4ka mentioned this pull request Mar 8, 2019

FastTree EarlyStoppingMetric is an int but only accepts specific values #2521

Closed

daholste mentioned this pull request Mar 25, 2019

Cannot access leaf nodes of trained trees #3083

Closed

ghost locked as resolved and limited conversation to collaborators Mar 24, 2022

Scrub Fast Tree Family #2753

Scrub Fast Tree Family #2753

Conversation

wschin commented Feb 27, 2019 • edited Loading

wschin Feb 27, 2019 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

wschin Feb 28, 2019 • edited Loading

Choose a reason for hiding this comment

TomFinley Mar 1, 2019 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Ivanidzo4ka Feb 27, 2019 • edited by wschin Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Ivanidzo4ka Feb 27, 2019 • edited by wschin Loading

Choose a reason for hiding this comment

TomFinley Feb 27, 2019 • edited by wschin Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Ivanidzo4ka Feb 27, 2019 • edited by wschin Loading

Choose a reason for hiding this comment

Ivanidzo4ka Feb 27, 2019 • edited by wschin Loading

Choose a reason for hiding this comment

wschin Feb 28, 2019 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Ivanidzo4ka Feb 27, 2019 • edited by wschin Loading

Choose a reason for hiding this comment

Ivanidzo4ka commented Feb 27, 2019 • edited by wschin Loading

Ivanidzo4ka Feb 27, 2019 • edited by wschin Loading

Choose a reason for hiding this comment

Ivanidzo4ka commented Feb 27, 2019 • edited by wschin Loading

TomFinley commented Feb 27, 2019 • edited by wschin Loading

Ivanidzo4ka commented Feb 28, 2019 • edited by wschin Loading

wschin commented Feb 28, 2019 • edited Loading

wschin commented Feb 28, 2019

wschin commented Feb 28, 2019

wschin commented Feb 28, 2019

singlis Mar 1, 2019 • edited by wschin Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

singlis commented Mar 1, 2019 • edited by wschin Loading

wschin commented Mar 1, 2019

singlis commented Mar 1, 2019 • edited by wschin Loading

singlis Mar 1, 2019 • edited by wschin Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

singlis Mar 1, 2019 • edited by wschin Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

singlis commented Mar 1, 2019 • edited by wschin Loading

singlis commented Mar 1, 2019 • edited by wschin Loading

singlis Mar 1, 2019 • edited by wschin Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Ivanidzo4ka Mar 2, 2019 • edited by wschin Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

wschin commented Mar 4, 2019

wschin commented Mar 4, 2019

wschin commented Mar 4, 2019

wschin commented Mar 4, 2019

Ivanidzo4ka Mar 4, 2019 • edited by wschin Loading

Choose a reason for hiding this comment

Ivanidzo4ka Mar 4, 2019 • edited by wschin Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Ivanidzo4ka commented Mar 5, 2019 • edited by wschin Loading

Ivanidzo4ka commented Mar 5, 2019 • edited by wschin Loading

ganik Mar 5, 2019 • edited by wschin Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Ivanidzo4ka left a comment

Choose a reason for hiding this comment

wschin commented Feb 27, 2019 •

edited

Loading

wschin Feb 27, 2019 •

edited

Loading

wschin Feb 28, 2019 •

edited

Loading

TomFinley Mar 1, 2019 •

edited

Loading

Ivanidzo4ka Feb 27, 2019 •

edited by wschin

Loading

Ivanidzo4ka Feb 27, 2019 •

edited by wschin

Loading

TomFinley Feb 27, 2019 •

edited by wschin

Loading

Ivanidzo4ka Feb 27, 2019 •

edited by wschin

Loading

Ivanidzo4ka Feb 27, 2019 •

edited by wschin

Loading

wschin Feb 28, 2019 •

edited

Loading

Ivanidzo4ka Feb 27, 2019 •

edited by wschin

Loading

Ivanidzo4ka commented Feb 27, 2019 •

edited by wschin

Loading

Ivanidzo4ka Feb 27, 2019 •

edited by wschin

Loading

Ivanidzo4ka commented Feb 27, 2019 •

edited by wschin

Loading

TomFinley commented Feb 27, 2019 •

edited by wschin

Loading

Ivanidzo4ka commented Feb 28, 2019 •

edited by wschin

Loading

wschin commented Feb 28, 2019 •

edited

Loading

singlis Mar 1, 2019 •

edited by wschin

Loading

singlis commented Mar 1, 2019 •

edited by wschin

Loading

singlis commented Mar 1, 2019 •

edited by wschin

Loading

singlis Mar 1, 2019 •

edited by wschin

Loading

singlis Mar 1, 2019 •

edited by wschin

Loading

singlis commented Mar 1, 2019 •

edited by wschin

Loading

singlis commented Mar 1, 2019 •

edited by wschin

Loading

singlis Mar 1, 2019 •

edited by wschin

Loading

Ivanidzo4ka Mar 2, 2019 •

edited by wschin

Loading

Ivanidzo4ka Mar 4, 2019 •

edited by wschin

Loading

Ivanidzo4ka Mar 4, 2019 •

edited by wschin

Loading

Ivanidzo4ka commented Mar 5, 2019 •

edited by wschin

Loading

Ivanidzo4ka commented Mar 5, 2019 •

edited by wschin

Loading

ganik Mar 5, 2019 •

edited by wschin

Loading