public static void Serialize(ProxySerialize serialize, LossMulticlassLogPerPixel net)
{
if (serialize == null)
{
throw new ArgumentNullException(nameof(serialize));
}
if (net == null)
{
throw new ArgumentNullException(nameof(net));
}
net.ThrowIfDisposed();
var error = NativeMethods.LossMulticlassLogPerPixel_serialize_proxy(net.NetworkType,
serialize.NativePtr,
net.NativePtr,
out var errorMessage);
switch (error)
{
case NativeMethods.ErrorType.DnnNotSupportNetworkType:
throw new NotSupportNetworkTypeException(net.NetworkType);
case NativeMethods.ErrorType.GeneralSerialization:
throw new SerializationException(StringHelper.FromStdString(errorMessage, true));
}
}
public static void Serialize(this IDictionary <string, LossMulticlassLogPerPixel> maps,
ProxySerialize serialize,
int networkType = 0)
{
if (serialize == null)
{
throw new ArgumentNullException(nameof(serialize));
}
serialize.ThrowIfDisposed();
StdString[] keys = null;
try
{
keys = new StdString[maps.Count];
var values = new IntPtr[maps.Count];
var index = 0;
foreach (var kvp in maps)
{
keys[index] = new StdString(kvp.Key);
values[index] = kvp.Value.NativePtr;
index++;
}
var keysArray = keys.Select(s => s.NativePtr).ToArray();
var error = NativeMethods.LossMulticlassLogPerPixel_serialize_proxy_map(networkType,
serialize.NativePtr,
keysArray,
values,
maps.Count,
out var errorMessage);
Dnn.Cuda.ThrowCudaException(error);
switch (error)
{
case NativeMethods.ErrorType.DnnNotSupportNetworkType:
throw new NotSupportNetworkTypeException(networkType);
case NativeMethods.ErrorType.GeneralSerialization:
throw new SerializationException(StringHelper.FromStdString(errorMessage, true));
}
}
finally
{
keys?.DisposeElement();
}
}
private static int Main(string[] args)
{
if (args.Length < 1)
{
Console.WriteLine("To run this program you need a copy of the PASCAL VOC2012 dataset.");
Console.WriteLine();
Console.WriteLine("You call this program like this: ");
Console.WriteLine("./dnn_instance_segmentation_train_ex /path/to/VOC2012 [det-minibatch-size] [seg-minibatch-size] [class-1] [class-2] [class-3] ...");
return(1);
}
try
{
Console.WriteLine("\nSCANNING PASCAL VOC2012 DATASET");
Console.WriteLine();
var listing = PascalVOC2012.GetPascalVoc2012TrainListing(args[0]).ToArray();
Console.WriteLine($"images in entire dataset: {listing.Length}");
if (listing.Length == 0)
{
Console.WriteLine("Didn't find the VOC2012 dataset. ");
return(1);
}
// mini-batches smaller than the default can be used with GPUs having less memory
var argc = args.Length;
var detMiniBatchSize = argc >= 2 ? int.Parse(args[1]) : 35;
var segMiniBatchSize = argc >= 3 ? int.Parse(args[2]) : 100;
Console.WriteLine($"det mini-batch size: {detMiniBatchSize}");
Console.WriteLine($"seg mini-batch size: {segMiniBatchSize}");
var desiredClassLabels = new List <string>();
for (var arg = 3; arg < argc; ++arg)
{
desiredClassLabels.Add(args[arg]);
}
if (!desiredClassLabels.Any())
{
desiredClassLabels.Add("bicycle");
desiredClassLabels.Add("car");
desiredClassLabels.Add("cat");
}
Console.Write("desired classlabels:");
foreach (var desiredClassLabel in desiredClassLabels)
{
Console.Write($" {desiredClassLabel}");
}
Console.WriteLine();
// extract the MMOD rects
Console.Write("\nExtracting all truth instances...");
var truthInstances = LoadAllTruthInstances(listing);
Console.WriteLine(" Done!");
Console.WriteLine();
if (listing.Length != truthInstances.Count)
{
throw new ApplicationException();
}
var originalTruthImages = new List <TruthImage>();
for (int i = 0, end = listing.Length; i < end; ++i)
{
originalTruthImages.Add(new TruthImage
{
Info = listing[i],
TruthInstances = truthInstances[i]
});
}
var truthImagesFilteredByClass = FilterBasedOnClassLabel(originalTruthImages, desiredClassLabels);
Console.WriteLine($"images in dataset filtered by class: {truthImagesFilteredByClass.Count}");
IgnoreSomeTruthBoxes(truthImagesFilteredByClass);
var truthImages = FilterImagesWithNoTruth(truthImagesFilteredByClass);
Console.WriteLine($"images in dataset after ignoring some truth boxes: {truthImages.Count}");
// First train an object detector network (loss_mmod).
Console.WriteLine("\nTraining detector network:");
var detNet = TrainDetectionNetwork(truthImages, (uint)detMiniBatchSize);
// Then train mask predictors (segmentation).
var segNetsByClass = new Dictionary <string, LossMulticlassLogPerPixel>();
// This flag controls if a separate mask predictor is trained for each class.
// Note that it would also be possible to train a separate mask predictor for
// class groups, each containing somehow similar classes -- for example, one
// mask predictor for cars and buses, another for cats and dogs, and so on.
const bool separateSegNetForEachClass = true;
if (separateSegNetForEachClass)
{
foreach (var classLabel in desiredClassLabels)
{
// Consider only the truth images belonging to this class.
var classImages = FilterBasedOnClassLabel(truthImages, new[] { classLabel });
Console.WriteLine($"\nTraining segmentation network for class {classLabel}:");
segNetsByClass[classLabel] = TrainSegmentationNetwork(classImages, (uint)segMiniBatchSize, classLabel);
}
}
else
{
Console.WriteLine("Training a single segmentation network:");
segNetsByClass[""] = TrainSegmentationNetwork(truthImages, (uint)segMiniBatchSize, "");
}
Console.WriteLine("Saving networks");
using (var proxy = new ProxySerialize(InstanceSegmentationNetFilename))
{
LossMmod.Serialize(proxy, detNet);
segNetsByClass.Serialize(proxy, 4);
}
}
catch (Exception e)
{
Console.WriteLine(e);
}
return(0);
}
