public void BuildAndTrain()
{
var featurizerModelLocation = inputModelLocation;
ConsoleWriteHeader("Read model");
Console.WriteLine($"Model location: {featurizerModelLocation}");
Console.WriteLine($"Images folder: {imagesFolder}");
Console.WriteLine($"Training file: {dataLocation}");
Console.WriteLine($"Default parameters: image size=({ImageNetSettings.imageWidth},{ImageNetSettings.imageHeight}), image mean: {ImageNetSettings.mean}");
var data = mlContext.Data.ReadFromTextFile <ImageNetData>(path: dataLocation, hasHeader: false);
var pipeline = mlContext.Transforms.Conversion.MapValueToKey(outputColumnName: LabelTokey, inputColumnName: DefaultColumnNames.Label)
.Append(mlContext.Transforms.LoadImages(imagesFolder, (ImageReal, nameof(ImageNetData.ImagePath))))
.Append(mlContext.Transforms.Resize(outputColumnName: ImageReal, imageWidth: ImageNetSettings.imageWidth, imageHeight: ImageNetSettings.imageHeight, inputColumnName: ImageReal))
.Append(mlContext.Transforms.ExtractPixels(new ImagePixelExtractorTransformer.ColumnInfo(name: "input", inputColumnName: ImageReal, interleave: ImageNetSettings.channelsLast, offset: ImageNetSettings.mean)))
.Append(mlContext.Transforms.ScoreTensorFlowModel(modelLocation: featurizerModelLocation, outputColumnNames: new[] { "softmax2_pre_activation" }, inputColumnNames: new[] { "input" }))
.Append(mlContext.MulticlassClassification.Trainers.LogisticRegression(labelColumn: LabelTokey, featureColumn: "softmax2_pre_activation"))
.Append(mlContext.Transforms.Conversion.MapKeyToValue((PredictedLabelValue, DefaultColumnNames.PredictedLabel)));
// Train the model
ConsoleWriteHeader("Training classification model");
ITransformer model = pipeline.Fit(data);
// Process the training data through the model
// This is an optional step, but it's useful for debugging issues
var trainData = model.Transform(data);
var loadedModelOutputColumnNames = trainData.Schema
.Where(col => !col.IsHidden).Select(col => col.Name);
var trainData2 = mlContext.CreateEnumerable <ImageNetPipeline>(trainData, false, true).ToList();
trainData2.ForEach(pr => ConsoleWriteImagePrediction(pr.ImagePath, pr.PredictedLabelValue, pr.Score.Max()));
// Get some performance metric on the model using training data
var classificationContext = new MulticlassClassificationCatalog(mlContext);
ConsoleWriteHeader("Classification metrics");
var metrics = classificationContext.Evaluate(trainData, label: LabelTokey, predictedLabel: DefaultColumnNames.PredictedLabel);
Console.WriteLine($"LogLoss is: {metrics.LogLoss}");
Console.WriteLine($"PerClassLogLoss is: {String.Join(" , ", metrics.PerClassLogLoss.Select(c => c.ToString()))}");
// Save the model to assets/outputs
ConsoleWriteHeader("Save model to local file");
ModelHelpers.DeleteAssets(outputModelLocation);
using (var f = new FileStream(outputModelLocation, FileMode.Create))
mlContext.Model.Save(model, f);
Console.WriteLine($"Model saved: {outputModelLocation}");
}
private static IDataView GetMulticlassMetrics(
IHostEnvironment env,
IPredictor predictor,
RoleMappedData roleMappedData,
PermutationFeatureImportanceArguments input)
{
var roles = roleMappedData.Schema.GetColumnRoleNames();
var featureColumnName = roles.Where(x => x.Key.Value == RoleMappedSchema.ColumnRole.Feature.Value).First().Value;
var labelColumnName = roles.Where(x => x.Key.Value == RoleMappedSchema.ColumnRole.Label.Value).First().Value;
var pred = new MulticlassPredictionTransformer <IPredictorProducing <VBuffer <float> > >(
env, predictor as IPredictorProducing <VBuffer <float> >, roleMappedData.Data.Schema, featureColumnName, labelColumnName);
var multiclassCatalog = new MulticlassClassificationCatalog(env);
var permutationMetrics = multiclassCatalog
.PermutationFeatureImportance(pred,
roleMappedData.Data,
labelColumnName: labelColumnName,
useFeatureWeightFilter: input.UseFeatureWeightFilter,
numberOfExamplesToUse: input.NumberOfExamplesToUse,
permutationCount: input.PermutationCount);
var slotNames = GetSlotNames(roleMappedData.Schema);
Contracts.Assert(slotNames.Length == permutationMetrics.Length,
"Mismatch between number of feature slots and number of features permuted.");
List <MulticlassMetrics> metrics = new List <MulticlassMetrics>();
for (int i = 0; i < permutationMetrics.Length; i++)
{
if (string.IsNullOrWhiteSpace(slotNames[i]))
{
continue;
}
var pMetric = permutationMetrics[i];
metrics.Add(new MulticlassMetrics
{
FeatureName = slotNames[i],
MacroAccuracy = pMetric.MacroAccuracy.Mean,
MacroAccuracyStdErr = pMetric.MacroAccuracy.StandardError,
MicroAccuracy = pMetric.MicroAccuracy.Mean,
MicroAccuracyStdErr = pMetric.MicroAccuracy.StandardError,
LogLoss = pMetric.LogLoss.Mean,
LogLossStdErr = pMetric.LogLoss.StandardError,
LogLossReduction = pMetric.LogLossReduction.Mean,
LogLossReductionStdErr = pMetric.LogLossReduction.StandardError,
TopKAccuracy = pMetric.TopKAccuracy.Mean,
TopKAccuracyStdErr = pMetric.TopKAccuracy.StandardError,
PerClassLogLoss = pMetric.PerClassLogLoss.Select(x => x.Mean).ToArray(),
PerClassLogLossStdErr = pMetric.PerClassLogLoss.Select(x => x.StandardError).ToArray()
});;
}
// Convert unknown size vectors to known size.
var metric = metrics.First();
SchemaDefinition schema = SchemaDefinition.Create(typeof(MulticlassMetrics));
ConvertVectorToKnownSize(nameof(metric.PerClassLogLoss), metric.PerClassLogLoss.Length, ref schema);
ConvertVectorToKnownSize(nameof(metric.PerClassLogLossStdErr), metric.PerClassLogLossStdErr.Length, ref schema);
var dataOps = new DataOperationsCatalog(env);
var result = dataOps.LoadFromEnumerable(metrics, schema);
return(result);
}
// </SnippetInceptionSettings>
// Build and train model
public static void ReuseAndTuneInceptionModel(MLContext mlContext, string dataLocation, string imagesFolder, string inputModelLocation, string outputModelLocation)
{
Console.WriteLine("Read model");
Console.WriteLine($"Model location: {inputModelLocation}");
Console.WriteLine($"Images folder: {_trainImagesFolder}");
Console.WriteLine($"Training file: {dataLocation}");
Console.WriteLine($"Default parameters: image size=({InceptionSettings.ImageWidth},{InceptionSettings.ImageHeight}), image mean: {InceptionSettings.Mean}");
// <SnippetLoadData>
var data = mlContext.Data.ReadFromTextFile <ImageData>(path: dataLocation, hasHeader: true);
// </SnippetLoadData>
// <SnippetMapValueToKey1>
var estimator = mlContext.Transforms.Conversion.MapValueToKey(outputColumnName: LabelTokey, inputColumnName: DefaultColumnNames.Label)
// </SnippetMapValueToKey1>
// The image transforms transform the images into the model's expected format.
// <SnippetImageTransforms>
.Append(mlContext.Transforms.LoadImages(_trainImagesFolder, (ImageReal, nameof(ImageData.ImagePath))))
.Append(mlContext.Transforms.Resize(outputColumnName: ImageReal, imageWidth: InceptionSettings.ImageWidth, imageHeight: InceptionSettings.ImageHeight, inputColumnName: ImageReal))
.Append(mlContext.Transforms.ExtractPixels(new ImagePixelExtractorTransformer.ColumnInfo(name: "input", inputColumnName: ImageReal, interleave: InceptionSettings.ChannelsLast, offset: InceptionSettings.Mean)))
// </SnippetImageTransforms>
// The ScoreTensorFlowModel transform scores the TensorFlow model and allows communication
// <SnippetScoreTensorFlowModel>
.Append(mlContext.Transforms.ScoreTensorFlowModel(modelLocation: inputModelLocation, outputColumnNames: new[] { "softmax2_pre_activation" }, inputColumnNames: new[] { "input" }))
// </SnippetScoreTensorFlowModel>
// <SnippetAddTrainer>
.Append(mlContext.MulticlassClassification.Trainers.LogisticRegression(labelColumn: LabelTokey, featureColumn: "softmax2_pre_activation"))
// </SnippetAddTrainer>
// <SnippetMapValueToKey2>
.Append(mlContext.Transforms.Conversion.MapKeyToValue((PredictedLabelValue, DefaultColumnNames.PredictedLabel)));
// </SnippetMapValueToKey2>
// Train the model
Console.WriteLine("=============== Training classification model ===============");
// Create and train the model based on the dataset that has been loaded, transformed.
// <SnippetTrainModel>
ITransformer model = estimator.Fit(data);
// </SnippetTrainModel>
// Process the training data through the model
// This is an optional step, but it's useful for debugging issues
// <SnippetTransformData>
var predictions = model.Transform(data);
// </SnippetTransformData>
// Create enumerables for both the ImageData and ImagePrediction DataViews
// for displaying results
// <SnippetEnumerateDataViews>
var imageData = mlContext.CreateEnumerable <ImageData>(data, false, true);
var imagePredictionData = mlContext.CreateEnumerable <ImagePrediction>(predictions, false, true);
// </SnippetEnumerateDataViews>
// Read the tags.tsv file and add the filepath to the image file name
// before loading into ImageData
// <SnippetCallPairAndDisplayResults1>
PairAndDisplayResults(imageData, imagePredictionData);
// </SnippetCallPairAndDisplayResults1>
// Get some performance metrics on the model using training data
Console.WriteLine("=============== Classification metrics ===============");
// <SnippetEvaluate>
var regressionContext = new MulticlassClassificationCatalog(mlContext);
var metrics = regressionContext.Evaluate(predictions, label: LabelTokey, predictedLabel: DefaultColumnNames.PredictedLabel);
// </SnippetEvaluate>
//<SnippetDisplayMetrics>
Console.WriteLine($"LogLoss is: {metrics.LogLoss}");
Console.WriteLine($"PerClassLogLoss is: {String.Join(" , ", metrics.PerClassLogLoss.Select(c => c.ToString()))}");
//</SnippetDisplayMetrics>
// Save the model to assets/outputs
Console.WriteLine("=============== Save model to local file ===============");
// <SnippetSaveModel>
using (var fileStream = new FileStream(outputModelLocation, FileMode.Create))
mlContext.Model.Save(model, fileStream);
// </SnippetSaveModel>
Console.WriteLine($"Model saved: {outputModelLocation}");
}
