public ReturnResult <Model> StartTraining([FromBody] TrainInput input)
{
var results = Model.TrainModel(
Db,
input,
Int32.Parse(User.FindFirstValue(ClaimTypes.NameIdentifier)),
Configuration["ModelBuilderUrl"]);
return(results);
}
public static ReturnResult <Model> Run([HttpTrigger(AuthorizationLevel.Anonymous, "post", Route = null)] HttpRequest req, ILogger log)
{
var dataFilePath = Path.Combine(Path.GetTempPath(), Guid.NewGuid().ToString());
try
{
db.BeginTransaction();
MLContext context = new MLContext();
TrainInput input = null;
using (StreamReader reader = new StreamReader(req.Body))
{
input = JsonConvert.DeserializeObject <TrainInput>(reader.ReadToEnd());
}
File.WriteAllText(dataFilePath, input.Data);
IDataView LoadedData = null;
var columnData = new List <TextLoader.Column>();
foreach (var c in input.Columns)
{
//data type 1 is for ignore
if (c.Type != 1)
{
var newColData = new TextLoader.Column()
{
DataKind = (DataKind)c.Type,
Name = c.ColumnName,
Source = new TextLoader.Range[] { new TextLoader.Range(c.ColumnIndex) }
};
columnData.Add(newColData);
}
}
LoadedData = context.Data.LoadFromTextFile(
dataFilePath,
columnData.ToArray(),
separatorChar: input.Separator,
hasHeader: input.HasHeaders,
allowQuoting: true
);
LoadedData = context.Data.ShuffleRows(LoadedData);
/*
* Multiclass will be used in the case of binary experiments and multiclass experiments.
* This is because multiclass can accept all types as an output column. This will
* allow less interaction with the user and a better user experience.
*/
double bestRunMetric = 0;
ITransformer bestModel = null;
if (input.ModelType == TrainInput.ModelTypes.Multiclass)
{
ExperimentResult <MulticlassClassificationMetrics> Results = null;
var settings = new MulticlassExperimentSettings()
{
MaxExperimentTimeInSeconds = 20
};
var training = context.Auto().CreateMulticlassClassificationExperiment(settings);
Results = training.Execute(LoadedData, labelColumnName: input.LabelColumn);
bestRunMetric = Results.BestRun.ValidationMetrics.MacroAccuracy;
bestModel = Results.BestRun.Model;
}
else if (input.ModelType == TrainInput.ModelTypes.Binary)
{
ExperimentResult <BinaryClassificationMetrics> Results = null;
var settings = new BinaryExperimentSettings()
{
MaxExperimentTimeInSeconds = 20
};
var training = context.Auto().CreateBinaryClassificationExperiment(settings);
Results = training.Execute(LoadedData, labelColumnName: input.LabelColumn);
bestRunMetric = Results.BestRun.ValidationMetrics.Accuracy;
bestModel = Results.BestRun.Model;
}
else if (input.ModelType == TrainInput.ModelTypes.Regression)
{
ExperimentResult <RegressionMetrics> Results = null;
var settings = new RegressionExperimentSettings()
{
MaxExperimentTimeInSeconds = 20
};
var training = context.Auto().CreateRegressionExperiment(settings);
Results = training.Execute(LoadedData, labelColumnName: input.LabelColumn);
bestRunMetric = Results.BestRun.ValidationMetrics.RSquared;
bestModel = Results.BestRun.Model;
if (bestRunMetric < 0)
{
bestRunMetric = 0;
}
}
else
{
throw new Exception("Invalid model type");
}
var modelFileId = 0;
using (MemoryStream ms = new MemoryStream())
{
context.Model.Save(bestModel, LoadedData.Schema, ms);
//Save model to the database
FileStore modelSave = new FileStore()
{
Data = ms.ToArray()
};
modelFileId = FileStore.InsertUpdate(db, modelSave).Item.FileStoreId;
}
var resultModel = new Model()
{
FileStoreId = modelFileId,
Accuracy = bestRunMetric,
Rows = input.Data.Trim().Split('\n').Length
};
db.CompleteTransaction();
return(new ReturnResult <Model>()
{
Success = true,
Item = resultModel
});
}
catch (Exception e)
{
db.AbortTransaction();
log.LogError(e.Message);
return(new ReturnResult <Model>()
{
Success = false,
Exception = e
});
}
}
public void RunTestBad()
{
Mock <ILogger> Logger = new Mock <ILogger>();
Mock <HttpRequest> request = new Mock <HttpRequest>();
Mock <IDatabase> Db = new Mock <IDatabase>();
TrainInput trainingInput = new TrainInput()
{
Columns = new List <ColumnInformation>()
{
new ColumnInformation()
{
ColumnIndex = 0,
ColumnName = "Id",
Type = 1 //ignore
},
new ColumnInformation()
{
ColumnIndex = 1,
ColumnName = "SepalLengthCm",
Type = (int)DataKind.Single
},
new ColumnInformation()
{
ColumnIndex = 2,
ColumnName = "SepalWidthCm",
Type = (int)DataKind.Single
},
new ColumnInformation()
{
ColumnIndex = 3,
ColumnName = "PetalLengthCm",
Type = (int)DataKind.Single
},
new ColumnInformation()
{
ColumnIndex = 4,
ColumnName = "PetalWidthCm",
Type = (int)DataKind.Single
},
new ColumnInformation()
{
ColumnIndex = 5,
ColumnName = "Species",
Type = (int)DataKind.Single // Wrong data type
}
},
Data = File.ReadAllText("Iris.csv"),
HasHeaders = true,
Description = "test",
LabelColumn = "Species",
ModelType = TrainInput.ModelTypes.Multiclass,
Separator = ',',
Title = "Iris Test"
};
var dataString = JsonConvert.SerializeObject(trainingInput);
MemoryStream dataStream = new MemoryStream(Encoding.ASCII.GetBytes(dataString));
request.Setup(x => x.Body).Returns(dataStream);
Train.db = Db.Object;
var results = Train.Run(request.Object, Logger.Object);
Assert.IsFalse(results.Success);
}
