Use function to create dynamic Bitmap runtime

I would like to build a Image Classificator with ml.net, the classification should be by auto rotate image.
Idea is have X images rotated right, then create dynamic at runtime Z images for each X rotated ad random value than use rotation as Label.
I need to rotate at runtime cause train set is around 2gb, do on disk for random angles will require around 1Tb and is not possible.

```
IDataView shuffledFullImagesDataset = mlContext.Transforms.Conversion.
MapValueToKey(outputColumnName: "LabelAsKey", inputColumnName: nameof(ImageData.Label), keyOrdinality: ValueToKeyMappingEstimator.KeyOrdinality.ByValue)
.Append(mlContext.Transforms.ExtractPixels(outputColumnName: "ImgInput", inputColumnName: "Image"))

public ImageData(string pathFile, int label)
{
PathFile = pathFile;
Label = label;
}

[ImageType(110, 110)]
public Bitmap Image => BitmapHelper.GetRotatedByteArrayToBitmap(File.ReadAllBytes(PathFile), Label);

public readonly int Label;

public readonly string PathFile;
```
But i have error when try to call Image look like not exist.
I found a python script that use keras to create random Imgs rotared at runtime when elaborate each Batch, then will avoid to allocate so much memory, but i didnt found some behavior on ml.net, look like it wanna allocate all imgs before training.