public ImageDataFrame(SiaVariable feature, SiaVariable label)
{
features = feature.Shape.Dimensions.ToArray();
labels = label.Shape.Dimensions[0];
counter = 0;
}
public DataFrameTrainPredict(Function model, Function lossFunc, string lossName, Function metricFunc, string metricName, IList <Learner> learners, SiaVariable featureVariable, SiaVariable labelVariable)
{
this.Model = model;
this.lossFunc = lossFunc;
this.metricFunc = metricFunc;
this.learners = learners;
this.featureVariable = featureVariable;
this.labelVariable = labelVariable;
this.metricName = metricName;
this.lossName = lossName;
}
internal void LoadTextData(CNTK.Variable feature, CNTK.Variable label)
{
int imageSize = feature.Shape.Rank == 1 ? feature.Shape[0] : feature.Shape[0] * feature.Shape[1] * feature.Shape[2];
int numClasses = label.Shape[0];
IList <StreamConfiguration> streamConfigurations = new StreamConfiguration[] { new StreamConfiguration(featureStreamName, imageSize), new StreamConfiguration(labelsStreamName, numClasses) };
miniBatchSource = MinibatchSource.TextFormatMinibatchSource(FileName, streamConfigurations, MinibatchSource.InfinitelyRepeat);
featureVariable = feature;
labelVariable = label;
featureStreamInfo = miniBatchSource.StreamInfo(featureStreamName);
labelStreamInfo = miniBatchSource.StreamInfo(labelsStreamName);
}
public ImageDataFrame(SiaVariable feature, SiaVariable label, string folder, int resize = 0, int numberOfRandomRotation = 0, bool horizontalFlip = false, bool verticalFlip = false)
: this(feature, label)
{
this.folder = folder;
fromFolder = true;
folderMapData = new List <ImageMapInfo>();
DirectoryInfo dir = new DirectoryInfo(folder);
var subfolders = dir.GetDirectories();
int counter = 1;
foreach (var item in subfolders)
{
var files = item.GetFiles().Select(x => (x.FullName)).ToList();
foreach (var file in files)
{
folderMapData.Add(new ImageMapInfo()
{
Filepath = file, Label = counter, RotationAngle = 0, Resize = resize
});
if (numberOfRandomRotation > 0)
{
for (int i = 0; i < numberOfRandomRotation; i++)
{
folderMapData.Add(new ImageMapInfo()
{
Filepath = file, Label = counter, RotationAngle = new Random(30).Next(10, 360), Resize = resize
});
}
}
//if (horizontalFlip)
//{
// folderMapData.Add(new ImageMapInfo() { Filepath = file, Label = counter, RotationAngle = 0, Flip = Emgu.CV.CvEnum.FlipType.Horizontal, Resize = resize });
//}
//if (verticalFlip)
//{
// folderMapData.Add(new ImageMapInfo() { Filepath = file, Label = counter, RotationAngle = 0, Flip = Emgu.CV.CvEnum.FlipType.Vertical, Resize = resize });
//}
}
counter++;
}
Shuffle();
}
public Dictionary <string, List <double> > Train(object trainData, object validationData, int epoches, int batchSize, On_Epoch_Start OnEpochStart, On_Epoch_End OnEpochEnd, On_Batch_Start onBatchStart, On_Batch_End OnBatchEnd, bool shuffle = false)
{
XYFrame train = (XYFrame)trainData;
XYFrame validation = validationData != null ? (XYFrame)validationData : null;
Dictionary <string, List <double> > result = new Dictionary <string, List <double> >();
var trainer = Trainer.CreateTrainer(Model, lossFunc, metricFunc, learners);
int currentEpoch = 1;
Dictionary <string, double> metricsList = new Dictionary <string, double>();
while (currentEpoch <= epoches)
{
if (shuffle)
{
train.Shuffle();
}
metricsList = new Dictionary <string, double>();
OnEpochStart(currentEpoch);
int miniBatchCount = 1;
while (train.NextBatch(miniBatchCount, batchSize))
{
onBatchStart(currentEpoch, miniBatchCount);
Value features = DataFrameUtil.GetValueBatch(train.CurrentBatch.XFrame);
Value labels = DataFrameUtil.GetValueBatch(train.CurrentBatch.YFrame);
trainer.TrainMinibatch(new Dictionary <SiaVariable, Value>()
{
{ featureVariable, features }, { labelVariable, labels }
}, GlobalParameters.Device);
OnBatchEnd(currentEpoch, miniBatchCount, trainer.TotalNumberOfSamplesSeen(), trainer.PreviousMinibatchLossAverage(), new Dictionary <string, double>()
{
{ metricName, trainer.PreviousMinibatchEvaluationAverage() }
});
miniBatchCount++;
}
if (!result.ContainsKey("loss"))
{
result.Add("loss", new List <double>());
}
if (!result.ContainsKey(metricName))
{
result.Add(metricName, new List <double>());
}
double lossValue = trainer.PreviousMinibatchLossAverage();
double metricValue = trainer.PreviousMinibatchEvaluationAverage();
result["loss"].Add(lossValue);
result[metricName].Add(metricValue);
metricsList.Add(metricName, metricValue);
if (validation != null)
{
if (!result.ContainsKey("val_loss"))
{
result.Add("val_loss", new List <double>());
}
if (!result.ContainsKey("val_" + metricName))
{
result.Add("val_" + metricName, new List <double>());
}
int evalMiniBatchCount = 1;
List <double> totalEvalBatchLossList = new List <double>();
List <double> totalEvalMetricValueList = new List <double>();
while (validation.NextBatch(evalMiniBatchCount, batchSize))
{
SiaVariable actualVariable = CNTKLib.InputVariable(labelVariable.Shape, DataType.Float);
var evalLossFunc = Losses.Get(lossName, labelVariable, actualVariable);
var evalMetricFunc = Metrics.Get(metricName, labelVariable, actualVariable);
Value actual = EvaluateInternal(validation.CurrentBatch.XFrame);
Value expected = DataFrameUtil.GetValueBatch(validation.CurrentBatch.YFrame);
var inputDataMap = new Dictionary <SiaVariable, Value>()
{
{ labelVariable, expected }, { actualVariable, actual }
};
var outputDataMap = new Dictionary <SiaVariable, Value>()
{
{ evalLossFunc.Output, null }
};
evalLossFunc.Evaluate(inputDataMap, outputDataMap, GlobalParameters.Device);
var evalLoss = outputDataMap[evalLossFunc.Output].GetDenseData <float>(evalLossFunc.Output).Select(x => x.First()).ToList();
totalEvalBatchLossList.Add(evalLoss.Average());
inputDataMap = new Dictionary <SiaVariable, Value>()
{
{ labelVariable, expected }, { actualVariable, actual }
};
outputDataMap = new Dictionary <SiaVariable, Value>()
{
{ evalMetricFunc.Output, null }
};
evalMetricFunc.Evaluate(inputDataMap, outputDataMap, GlobalParameters.Device);
var evalMetric = outputDataMap[evalMetricFunc.Output].GetDenseData <float>(evalMetricFunc.Output).Select(x => x.First()).ToList();
totalEvalMetricValueList.Add(evalMetric.Average());
evalMiniBatchCount++;
}
result["val_loss"].Add(totalEvalBatchLossList.Average());
metricsList.Add("val_loss", totalEvalBatchLossList.Average());
result["val_" + metricName].Add(totalEvalMetricValueList.Average());
metricsList.Add("val_" + metricName, totalEvalMetricValueList.Average());
}
OnEpochEnd(currentEpoch, trainer.TotalNumberOfSamplesSeen(), lossValue, metricsList);
currentEpoch++;
}
return(result);
}
public Dictionary <string, List <double> > Train(object trainData, object validationData, int epoches, int batchSize, On_Epoch_Start OnEpochStart, On_Epoch_End OnEpochEnd, On_Batch_Start onBatchStart, On_Batch_End OnBatchEnd, bool shuffle = false)
{
ImageDataGenerator train = (ImageDataGenerator)trainData;
ImageDataGenerator validation = validationData != null ? (ImageDataGenerator)validationData : null;
Dictionary <string, List <double> > result = new Dictionary <string, List <double> >();
var trainer = Trainer.CreateTrainer(Model, lossFunc, metricFunc, learners);
int currentEpoch = 1;
Dictionary <string, double> metricsList = new Dictionary <string, double>();
int imageSize = featureVariable.Shape.Rank == 1 ? featureVariable.Shape[0] : featureVariable.Shape[0] * featureVariable.Shape[1] * featureVariable.Shape[2];
int numClasses = labelVariable.Shape[0];
IList <StreamConfiguration> streamConfigurations = new StreamConfiguration[] { new StreamConfiguration("features", imageSize), new StreamConfiguration("labels", numClasses) };
if (train.GenType == ImageGenType.FromTextFile)
{
train.LoadTextData(featureVariable, labelVariable);
if (validation != null)
{
validation.LoadTextData(featureVariable, labelVariable);
}
}
while (currentEpoch <= epoches)
{
metricsList.Clear();
OnEpochStart(currentEpoch);
int miniBatchCount = 1;
List <double> miniBatchLosses = new List <double>();
while (!train.NextBatch(batchSize))
{
onBatchStart(currentEpoch, miniBatchCount);
trainer.TrainMinibatch(new Dictionary <SiaVariable, Value> {
{ featureVariable, train.CurrentBatchX }, { labelVariable, train.CurrentBatchY }
}, true, GlobalParameters.Device);
OnBatchEnd(currentEpoch, miniBatchCount, trainer.TotalNumberOfSamplesSeen(), trainer.PreviousMinibatchLossAverage(), new Dictionary <string, double>()
{
{ metricName, trainer.PreviousMinibatchEvaluationAverage() }
});
miniBatchLosses.Add(trainer.PreviousMinibatchLossAverage());
miniBatchCount++;
}
if (!result.ContainsKey("loss"))
{
result.Add("loss", new List <double>());
}
if (!result.ContainsKey(metricName))
{
result.Add(metricName, new List <double>());
}
double lossValue = miniBatchLosses.Average();
double metricValue = trainer.PreviousMinibatchEvaluationAverage();
result["loss"].Add(lossValue);
result[metricName].Add(metricValue);
metricsList.Add(metricName, metricValue);
if (validation != null)
{
if (!result.ContainsKey("val_loss"))
{
result.Add("val_loss", new List <double>());
}
if (!result.ContainsKey("val_" + metricName))
{
result.Add("val_" + metricName, new List <double>());
}
List <double> totalEvalBatchLossList = new List <double>();
List <double> totalEvalMetricValueList = new List <double>();
while (validation.NextBatch(batchSize))
{
SiaVariable actualVariable = CNTKLib.InputVariable(labelVariable.Shape, DataType.Float);
var evalLossFunc = Losses.Get(lossName, labelVariable, actualVariable);
var evalMetricFunc = Metrics.Get(metricName, labelVariable, actualVariable);
Value actual = EvaluateInternal(validation.CurrentBatchX);
Value expected = validation.CurrentBatchY;
var inputDataMap = new Dictionary <SiaVariable, Value>()
{
{ labelVariable, expected }, { actualVariable, actual }
};
var outputDataMap = new Dictionary <SiaVariable, Value>()
{
{ evalLossFunc.Output, null }
};
evalLossFunc.Evaluate(inputDataMap, outputDataMap, GlobalParameters.Device);
var evalLoss = outputDataMap[evalLossFunc.Output].GetDenseData <float>(evalLossFunc.Output).Select(x => x.First()).ToList();
totalEvalBatchLossList.Add(evalLoss.Average());
inputDataMap = new Dictionary <SiaVariable, Value>()
{
{ labelVariable, expected }, { actualVariable, actual }
};
outputDataMap = new Dictionary <SiaVariable, Value>()
{
{ evalMetricFunc.Output, null }
};
evalMetricFunc.Evaluate(inputDataMap, outputDataMap, GlobalParameters.Device);
var evalMetric = outputDataMap[evalMetricFunc.Output].GetDenseData <float>(evalMetricFunc.Output).Select(x => x.First()).ToList();
totalEvalMetricValueList.Add(evalMetric.Average());
}
result["val_loss"].Add(totalEvalBatchLossList.Average());
metricsList.Add("val_loss", totalEvalBatchLossList.Average());
result["val_" + metricName].Add(totalEvalMetricValueList.Average());
metricsList.Add("val_" + metricName, totalEvalMetricValueList.Average());
}
OnEpochEnd(currentEpoch, trainer.TotalNumberOfSamplesSeen(), lossValue, metricsList);
currentEpoch++;
}
return(result);
}
