// Create a model from an ONNX 1.2 file
public static async Task <CustomVisionModel> CreateONNXModel(StorageFile file)
{
LearningModel learningModel = null;
try
{
learningModel = await LearningModel.LoadFromStorageFileAsync(file);
}
catch (Exception ex)
{
throw ex;
}
var inputFeatures = learningModel.InputFeatures;
ImageFeatureDescriptor inputImageDescription = inputFeatures?.FirstOrDefault(feature => feature.Kind == LearningModelFeatureKind.Image) as ImageFeatureDescriptor;
uint inputImageWidth = 0, inputImageHeight = 0;
if (inputImageDescription != null)
{
inputImageHeight = inputImageDescription.Height;
inputImageWidth = inputImageDescription.Width;
}
return(new CustomVisionModel()
{
_learningModel = learningModel,
_session = new LearningModelSession(learningModel),
InputImageWidth = (int)inputImageWidth,
InputImageHeight = (int)inputImageHeight
});
}
public void debugModelIO()
{
string _inName, _outName;
foreach (var inputF in _learningModel.InputFeatures)
{
Debug.WriteLine($"input | kind:{inputF.Kind}, name:{inputF.Name}, type:{inputF.GetType()}");
int i = 0;
ImageFeatureDescriptor imgDesc = inputF as ImageFeatureDescriptor;
TensorFeatureDescriptor tfDesc = inputF as TensorFeatureDescriptor;
_inWidth = (uint)(imgDesc == null ? tfDesc.Shape[3] : imgDesc.Width);
_inHeight = (uint)(imgDesc == null ? tfDesc.Shape[2] : imgDesc.Height);
_inName = inputF.Name;
Debug.WriteLine($"N: {(imgDesc == null ? tfDesc.Shape[0] : 1)}, " +
$"Channel: {(imgDesc == null ? tfDesc.Shape[1].ToString() : imgDesc.BitmapPixelFormat.ToString())}, " +
$"Height:{(imgDesc == null ? tfDesc.Shape[2] : imgDesc.Height)}, " +
$"Width: {(imgDesc == null ? tfDesc.Shape[3] : imgDesc.Width)}");
}
foreach (var outputF in _learningModel.OutputFeatures)
{
Debug.WriteLine($"output | kind:{outputF.Kind}, name:{outputF.Name}, type:{outputF.GetType()}");
int i = 0;
ImageFeatureDescriptor imgDesc = outputF as ImageFeatureDescriptor;
TensorFeatureDescriptor tfDesc = outputF as TensorFeatureDescriptor;
_outWidth = (uint)(imgDesc == null ? tfDesc.Shape[3] : imgDesc.Width);
_outHeight = (uint)(imgDesc == null ? tfDesc.Shape[2] : imgDesc.Height);
_outName = outputF.Name;
Debug.WriteLine($"N: {(imgDesc == null ? tfDesc.Shape[0] : 1)}, " +
$"Channel: {(imgDesc == null ? tfDesc.Shape[1].ToString() : imgDesc.BitmapPixelFormat.ToString())}, " +
$"Height:{(imgDesc == null ? tfDesc.Shape[2] : imgDesc.Height)}, " +
$"Width: {(imgDesc == null ? tfDesc.Shape[3] : imgDesc.Width)}");
}
}
/// <summary>
/// Load the labels and model and initialize WinML
/// </summary>
/// <returns></returns>
private async Task LoadModelAsync(string modelFileName)
{
Debug.WriteLine("LoadModelAsync");
_evaluationLock.Wait();
{
m_binding = null;
m_model = null;
m_session = null;
_isReadyForEval = false;
try
{
// Start stopwatch
_perfStopwatch.Restart();
// Load Model
StorageFile modelFile = await StorageFile.GetFileFromApplicationUriAsync(new Uri($"ms-appx:///Assets/{modelFileName}.onnx"));
m_model = await LearningModel.LoadFromStorageFileAsync(modelFile);
// Stop stopwatch
_perfStopwatch.Stop();
// Setting preferred inference device given user's intent
m_inferenceDeviceSelected = _useGPU ? LearningModelDeviceKind.DirectXHighPerformance : LearningModelDeviceKind.Cpu;
m_session = new LearningModelSession(m_model, new LearningModelDevice(m_inferenceDeviceSelected));
// Debugging logic to see the input and output of ther model and retrieve dimensions of input/output variables
// ### DEBUG ###
foreach (var inputF in m_model.InputFeatures)
{
Debug.WriteLine($"input | kind:{inputF.Kind}, name:{inputF.Name}, type:{inputF.GetType()}");
int i = 0;
ImageFeatureDescriptor imgDesc = inputF as ImageFeatureDescriptor;
TensorFeatureDescriptor tfDesc = inputF as TensorFeatureDescriptor;
m_inWidth = (uint)(imgDesc == null ? tfDesc.Shape[3] : imgDesc.Width);
m_inHeight = (uint)(imgDesc == null ? tfDesc.Shape[2] : imgDesc.Height);
m_inName = inputF.Name;
Debug.WriteLine($"N: {(imgDesc == null ? tfDesc.Shape[0] : 1)}, " +
$"Channel: {(imgDesc == null ? tfDesc.Shape[1].ToString() : imgDesc.BitmapPixelFormat.ToString())}, " +
$"Height:{(imgDesc == null ? tfDesc.Shape[2] : imgDesc.Height)}, " +
$"Width: {(imgDesc == null ? tfDesc.Shape[3] : imgDesc.Width)}");
}
foreach (var outputF in m_model.OutputFeatures)
{
Debug.WriteLine($"output | kind:{outputF.Kind}, name:{outputF.Name}, type:{outputF.GetType()}");
int i = 0;
ImageFeatureDescriptor imgDesc = outputF as ImageFeatureDescriptor;
TensorFeatureDescriptor tfDesc = outputF as TensorFeatureDescriptor;
m_outWidth = (uint)(imgDesc == null ? tfDesc.Shape[3] : imgDesc.Width);
m_outHeight = (uint)(imgDesc == null ? tfDesc.Shape[2] : imgDesc.Height);
m_outName = outputF.Name;
Debug.WriteLine($"N: {(imgDesc == null ? tfDesc.Shape[0] : 1)}, " +
$"Channel: {(imgDesc == null ? tfDesc.Shape[1].ToString() : imgDesc.BitmapPixelFormat.ToString())}, " +
$"Height:{(imgDesc == null ? tfDesc.Shape[2] : imgDesc.Height)}, " +
$"Width: {(imgDesc == null ? tfDesc.Shape[3] : imgDesc.Width)}");
}
// ### END OF DEBUG ###
// Create output frame
_outputFrame?.Dispose();
_outputFrame = new VideoFrame(BitmapPixelFormat.Bgra8, (int)m_outWidth, (int)m_outHeight);
Debug.WriteLine($"Elapsed time: {_perfStopwatch.ElapsedMilliseconds} ms");
_isReadyForEval = true;
}
catch (Exception ex)
{
NotifyUser($"error: {ex.Message}", NotifyType.ErrorMessage);
Debug.WriteLine($"error: {ex.Message}");
}
}
_evaluationLock.Release();
}
/// <summary>
/// Crop image given a imageVariableDescription
/// </summary>
/// <param name="inputVideoFrame"></param>
/// <returns></returns>
public static IAsyncOperation <VideoFrame> CenterCropImageAsync(VideoFrame inputVideoFrame, ImageFeatureDescriptor imageVariableDescription)
{
return(CenterCropImageAsync(inputVideoFrame, imageVariableDescription.Width, imageVariableDescription.Height));
}
