static async Task Main(string[] args)
{
var rootDir = Path.GetDirectoryName(Assembly.GetExecutingAssembly().Location);
var squeezeNetModel = SqueezeNetModel.CreateFromFilePath(Path.Combine(rootDir, "squeezenet1.0-9.onnx"));
// Load labels from JSON
var labels = new List <string>();
foreach (var kvp in JsonSerializer.Deserialize <Dictionary <string, string> >(File.ReadAllText(Path.Combine(rootDir, "Labels.json"))))
{
labels.Add(kvp.Value);
}
if (args.Length < 1)
{
return;
}
var filePath = args[0];
// Open image file
SqueezeNetOutput output;
using (var fileStream = File.OpenRead(filePath))
{
// Convert from FileStream to ImageFeatureValue
var decoder = await BitmapDecoder.CreateAsync(fileStream.AsRandomAccessStream());
using var softwareBitmap = await decoder.GetSoftwareBitmapAsync();
using var inputImage = VideoFrame.CreateWithSoftwareBitmap(softwareBitmap);
var imageTensor = ImageFeatureValue.CreateFromVideoFrame(inputImage);
output = await squeezeNetModel.EvaluateAsync(new SqueezeNetInput
{
data_0 = imageTensor
});
}
// Get result, which is a list of floats with all the probabilities for all 1000 classes of SqueezeNet
var resultTensor = output.softmaxout_1;
var resultVector = resultTensor.GetAsVectorView();
// Order the 1000 results with their indexes to know which class is the highest ranked one
List <(int index, float p)> results = new List <(int, float)>();
for (int i = 0; i < resultVector.Count; i++)
{
results.Add((index: i, p: resultVector.ElementAt(i)));
}
results.Sort((a, b) => a.p switch
{
var p when p <b.p => 1,
var p when p> b.p => - 1,
_ => 0
});
public static SqueezeNetModel CreateFromFilePath(string filePath)
{
var learningModel = new SqueezeNetModel
{
model = LearningModel.LoadFromFilePath(filePath)
};
learningModel.session = new LearningModelSession(learningModel.model);
learningModel.binding = new LearningModelBinding(learningModel.session);
return(learningModel);
}
