StochasticDualCoordinateAscent not work For Multiclass after migrate to 0.10.0

[DevLob-zz]

System information

• OS version/distro: W10
• .NET Version (eg., dotnet --info): 4.6.1

Issue

• What did you do? Migrated my code from 0.9.0 to 0.10.0
• What happened? StochasticDualCoordinateAscent Algorithm was working fine for multi class and binary trainer before move to 0.10 after updating its working only for binary and freeze on
  trainingPipeline.Fit(trainingDataView); for Multiclass take a while
• What did you expect?
  it should work fine

Source code / logs

`var mlContext = new MLContext(seed: 1);

#region "STEP 1: Common data loading configuration"

IDataView trainingDataView = GetNormalDataSet(mlContext, allFeatures, mLFeatures);

if (trainingDataView.GetRowCount() == 0)
{

    return;
}

textFeatures = GetTextFeatures(normalFeatures);

numericFeatures = GetNumericFeatures(normalFeatures).ToArray();

#endregion

#region "STEP 2: Common data process configuration with pipeline data transformations"

// STEP 2: Common data process configuration with pipeline data transformations

var textFeaturesProcessPipeline = mlContext.Transforms.Text.FeaturizeText(DefaultColumnNames.Features, textFeatures);

var numericFeaturesProcessPipeline = mlContext.Transforms.Concatenate(DefaultColumnNames.Features, numericFeatures);

var dataProcessPipeline = numericFeaturesProcessPipeline.Append(textFeaturesProcessPipeline).AppendCacheCheckpoint(mlContext);

#endregion

#region "STEP 3: Set the training algorithm, then create and configure the modelBuilder"

ITransformer trainedModel = null;

//"StochasticDualCoordinateAscent"

var trainingPipeline = mlContext.Transforms.Conversion.MapValueToKey(DefaultColumnNames.Label)
.Append(dataProcessPipeline)
.Append(mlContext.MulticlassClassification.Trainers.StochasticDualCoordinateAscent(labelColumn: DefaultColumnNames.Label, featureColumn: DefaultColumnNames.Features))
.Append(mlContext.Transforms.Conversion.MapKeyToValue(DefaultColumnNames.PredictedLabel));

#region STEP 4: Train the model fitting to the DataSet

//Take a while and no responce when call fit method

trainedModel = trainingPipeline.Fit(trainingDataView);

#endregion`

you can see some screen for values and when change from
"StochasticDualCoordinateAscent" to "Naive Bayes" Working fine

what wrong on my code

also those my Data Structure Classes
`[Serializable]
public class NormalTagsModelFeatures
{
//[Column(ordinal: "0", name: "Label")] public string Label;
[LoadColumn(0)]
public string Label;
[LoadColumn(1)]
public float fontSize;
[LoadColumn(2)]
public float isBold;
[LoadColumn(3)]
public float isItalic;
[LoadColumn(4)]
public float isUnderLine;
[LoadColumn(5)]
public float containsDot;
[LoadColumn(6)]
public float containsQuestionMark;
[LoadColumn(7)]
public string fontColor;
[LoadColumn(8)]
public float isAllCaps;
[LoadColumn(9)]
public string tagText;
[LoadColumn(10)]
public string firstWord;
[LoadColumn(11)]
public string FontName;
[LoadColumn(12)]
public float verticalText;
[LoadColumn(13)]
public float trdLeft;
[LoadColumn(14)]
public float trdRight;
[LoadColumn(15)]
public float trdTop;
[LoadColumn(16)]
public float trdBottom;
[LoadColumn(17)]
public float pageNo;

}

public class NormalTagsPrediction
{
    [ColumnName("PredictedLabel")]
    public string PredictedLabel;

    [ColumnName("Score")]
    public float[] Score { get; set; }

}`

Please paste or attach the code or logs or traces that would be helpful to diagnose the issue you are reporting.

[DevLob-zz]

Hint My dataView Contain 250K rows
Also worked fine within Naive Bayes and LogisticRegression

[rogancarr]

@DevLob-zz Thanks for reporting this! I'll take a look at this!

Marking as need info until I repro it.

[DevLob-zz]

is there is info can i provide you with it , thank you for interest

best regards

[rogancarr]

Hi @DevLob-zz ,

I've run a multiclass classification sample, GitHubIssueClassification in 0.9 and 0.10, and they get very similar times, so I don't think the problem is related to the SDCA trainer or it's API. Plus we haven't done any work on SDCA between the releases.

One thing that did change between 0.9 and 0.10 is the order of arguments to transforms. Do you do any featurization in GetNormalDataSet(mlContext, allFeatures, mLFeatures)? Is it possible that your column names have changed between the releases, and the learners are building features on different data?

[DevLob-zz]

No the code never changed between 0.9 and 0.10 and still work fine for Naive Bayes and Logistic Regression for Multi class
SDCA Still work fine for Binary Classification but seem to deep into infinite loop wile Multi class

[DevLob-zz]

the code stop and no response in Fit Method in this line however if code going to other case same code just change algorithm working well

[DevLob-zz]

any help would be appreciated

[Ivanidzo4ka]

mlContext has Log event.

 mlContext.Log += MlContext_Log;
private static void MlContext_Log(object sender, LoggingEventArgs e)
{
    Console.WriteLine(e.Message);
}

If we can look on this messages, it can be helpful to diagnose at what state it get stuck at least.
Do you get any success if you setup NumThreads in SDCA to 1?

We had this change #2152 which can be reason why it worked in 0.9 but stopped in 0.10, but as @rogancarr mentioned it works for us on some datasets, so your case can be quite tricky. It's hard to fix problem without ability to reproduce it.

[DevLob-zz]

mlContext has Log event.

 mlContext.Log += MlContext_Log;
private static void MlContext_Log(object sender, LoggingEventArgs e)
{
    Console.WriteLine(e.Message);
}

If we can look on this messages, it can be helpful to diagnose at what state it get stuck at least.
Do you get any success if you setup NumThreads in SDCA to 1?

We had this change #2152 which can be reason why it worked in 0.9 but stopped in 0.10, but as @rogancarr mentioned it works for us on some datasets, so your case can be quite tricky. It's hard to fix problem without ability to reproduce it.
this what the logger stop at

Do you get any success if you setup NumThreads in SDCA to 1?
how can i set this value do you mean this
var mlContext = new MLContext(seed: 1,conc:1);
if yes i did and same
It's hard to fix problem without ability to reproduce it.
yes sure i will try to debug through The ML Source
and keep you posted

best regards

[DevLob-zz]

after integrate my code with the source i can see that the code stick here

and he stick here as he get numThreads = 2 so if i can set it to 1 i think he will never stick again
NumofCycle = 750000

[DevLob-zz]

i replace the trainer line with
var trainner = mlContext.MulticlassClassification.Trainers.StochasticDualCoordinateAscent(new SdcaMultiClassTrainer.Options
{ FeatureColumn = DefaultColumnNames.Features ,
LabelColumn = DefaultColumnNames.Label,
NumThreads = 1
});

and same issue

this is my program .cs code i usesd

using System.Collections.Generic;
using Microsoft.Data.DataView;
using Microsoft.ML.Core.Data;
using Microsoft.ML.Data;
using Microsoft.ML.Samples.Dynamic;
using Microsoft.ML.Trainers;

namespace Microsoft.ML.Samples
{
    internal static class Program
    {
        static void Main(string[] args)
        {
            List<NormalTagsModelFeatures> noramalTagsTrainingData = new List<NormalTagsModelFeatures>();
            noramalTagsTrainingData.Add(
                        new NormalTagsModelFeatures()
                        {
                            //Label = TagListDictionary.GetTagId(label),
                            Label = "Tag Test",
                            pageNo =  -1,
                            fontSize =  -1,
                            isBold =  -1,
                            isItalic =  -1,
                            isUnderLine =  -1,
                            containsDot =  -1,
                            containsQuestionMark =  -1,
                            isAllCaps =  -1,
                            fontColor =  "Ss",
                            FontName =  "Ss",
                            tagText =  "dd  dd",
                            firstWord = "dd",
                            trdBottom =  -1,
                            trdLeft =  -1,
                            trdRight = -1,
                            trdTop = -1,
                            verticalText = -1,
                        }
                    );
            List<string> textFeatures = new List<string>() { "firstWord", "tagText", "FontName", "fontColor" };
            string[] numericFeatures = new string[] { "pageNo", "fontSize", "isBold", "isItalic", "isUnderLine",
                "containsDot", "containsQuestionMark","isAllCaps","trdBottom","trdLeft","trdRight","trdTop","verticalText" } ;
            var mlContext = new MLContext(seed: 1);
            IDataView trainingDataView =mlContext.Data.ReadFromEnumerable(noramalTagsTrainingData);
            var textFeaturesProcessPipeline = mlContext.Transforms.Text.FeaturizeText(DefaultColumnNames.Features, textFeatures,new Transforms.Text.TextFeaturizingEstimator.Options());
            var numericFeaturesProcessPipeline = mlContext.Transforms.Concatenate(DefaultColumnNames.Features, numericFeatures);
            var dataProcessPipeline = numericFeaturesProcessPipeline.Append(textFeaturesProcessPipeline).AppendCacheCheckpoint(mlContext);
            var trainner = mlContext.MulticlassClassification.Trainers.StochasticDualCoordinateAscent(new SdcaMultiClassTrainer.Options
            {
                FeatureColumn = DefaultColumnNames.Features,
                LabelColumn = DefaultColumnNames.Label,
                NumThreads = 1
            });
            trainner.AppendCacheCheckpoint(mlContext);
            var trainingPipeline = mlContext.Transforms.Conversion.MapValueToKey(DefaultColumnNames.Label)
           .Append(dataProcessPipeline)
           .Append(trainner)
           .Append(mlContext.Transforms.Conversion.MapKeyToValue(DefaultColumnNames.PredictedLabel));
            ITransformer trainedModel = trainingPipeline.Fit(trainingDataView);
            //TakeRows.Example();
        }
    }

    public class NormalTagsModelFeatures
    {
        //[Column(ordinal: "0", name: "Label")] public string Label;
        [LoadColumn(0)]
        public string Label;
        [LoadColumn(1)]
        public float fontSize;
        [LoadColumn(2)]
        public float isBold;
        [LoadColumn(3)]
        public float isItalic;
        [LoadColumn(4)]
        public float isUnderLine;
        [LoadColumn(5)]
        public float containsDot;
        [LoadColumn(6)]
        public float containsQuestionMark;
        [LoadColumn(7)]
        public string fontColor;
        [LoadColumn(8)]
        public float isAllCaps;
        [LoadColumn(9)]
        public string tagText;
        [LoadColumn(10)]
        public string firstWord;
        [LoadColumn(11)]
        public string FontName;
        [LoadColumn(12)]
        public float verticalText;
        [LoadColumn(13)]
        public float trdLeft;
        [LoadColumn(14)]
        public float trdRight;
        [LoadColumn(15)]
        public float trdTop;
        [LoadColumn(16)]
        public float trdBottom;
        [LoadColumn(17)]
        public float pageNo;

    }

    public class NormalTagsPrediction
    {
        [Column(ordinal: "0", name: "PredictedLabel")]
        public string Label;


        [ColumnName("Score")]
        public float[] Score { get; set; }

    }
}

[DevLob-zz]

maxIterations = 1,500,000

what is the best value for maxIterations can i set and not affect the accuracy

[DevLob-zz]

var trainner = mlContext.MulticlassClassification.Trainers.StochasticDualCoordinateAscent(new SdcaMultiClassTrainer.Options
{
FeatureColumn = DefaultColumnNames.Features,
LabelColumn = DefaultColumnNames.Label,
NumThreads = 1,
MaxIterations= 1000
});

so this solve my issue but

which MaxIterations will be recommenced based on your

[wschin]

Number of iterations is data-dependent. You can try 10, 20, 40, 80, 160, ..., 640 to find the value leading to the best test accuracy. In addition, a small regularization coefficient may lead to overfitting so you need to terminate the training very early (e.g., just 1, 2, 4, 8, 16 iterations are enough).

[Ivanidzo4ka]

It doesn't justify our auto selecting code to put 1 500 000 iterations over dataset.

[DevLob-zz]

the code seem to be wait while here and never out from while when testing on 500K rows

i wait on this iteration more than one hour and never move to next

is we can add timeout or something that guarantee that he will out from while Loop

[eerhardt]

@rogancarr @Ivanidzo4ka @TomFinley - the above deadlock looks to be the same as #1095.

[artidoro]

Hello @DevLob-zz thank you for sharing all this information.
I have a few questions for you:

1. Does it just take a long time but eventually finish training, or is it stuck training?
2. Is the issue showing only on v0.10 or did you have the same issue on v0.9? Are you building off of master? Or using the nuget for v0.10?
3. Is this reproducible 100% of the time?

The above questions can be very helpful in identifying what's wrong.

Also, I am trying to reproduce this issue locally to fix it, and if you are willing to share the dataset I can take a closer look, no problem of course otherwise.

[DevLob-zz]

Hello @DevLob-zz thank you for sharing all this information.
I have a few questions for you:

1. Does it just take a long time but eventually finish training, or is it stuck training?
   it never finish when it stuck
2. Is the issue showing only on v0.10 or did you have the same issue on v0.9? Are you building off of master? Or using the nuget for v0.10?
   for 0.9 everything is OK but for 0.10 i issued this
   i try by nuget and the master this what show me this dead lock
3. Is this reproducible 100% of the time?
   No

The above questions can be very helpful in identifying what's wrong.

Also, I am trying to reproduce this issue locally to fix it, and if you are willing to share the dataset I can take a closer look, no problem of course otherwise.

here is a program.cs that i used to generate more than 400K and try to generate the model

Hint : it work fine with small data set

Program.zip

[luisquintanilla]

Closing this issue assuming it's fixed due to lack of activity and no new reported issues on the matter. Thanks for the discussion / feedback.