public static TensorProto CreateTensorFromImage(int[][] imageData, float revertsBits)
{
var imageFeatureShape = new TensorShapeProto();
imageFeatureShape.Dim.Add(new TensorShapeProto.Types.Dim()
{
Size = 1
});
imageFeatureShape.Dim.Add(new TensorShapeProto.Types.Dim()
{
Size = imageData.Length * imageData.Length
});
var imageTensorBuilder = new TensorProto();
imageTensorBuilder.Dtype = DataType.DtFloat;
imageTensorBuilder.TensorShape = imageFeatureShape;
for (int i = 0; i < imageData.Length; ++i)
{
for (int j = 0; j < imageData.Length; ++j)
{
imageTensorBuilder.FloatVal.Add(imageData[i][j] / revertsBits);
}
}
return(imageTensorBuilder);
}
private TensorProto GetTensorProto(byte[] byteArray)
{
ByteString imageData = ByteString.CopyFrom(byteArray);
TensorProto tensorProto = new TensorProto();
Dim dimBatch = new Dim()
{
Name = "batch", Size = 1
};
Dim dimData = new Dim()
{
Name = "data", Size = 1
};
RepeatedField <Dim> repeatedField = new RepeatedField <Dim>();
repeatedField.Add(dimBatch);
repeatedField.Add(dimData);
TensorShapeProto tensorShape = new TensorShapeProto();
tensorShape.Dim = repeatedField;
tensorProto.TensorShape = tensorShape;
tensorProto.Dtype = DataType.DtString;
tensorProto.StringVal.Add(imageData);
return(tensorProto);
}
public static int[] ConvertSymbolicShapeToBarracuda(TensorShapeProto shape, string onnxLayout)
{
// TODO: use dimension denotation from TensorShapeProto to figure, if this particular tensor has specific data layout
// https://github.com/onnx/onnx/blob/master/docs/DimensionDenotation.md
var onnxShape = shape.AsInts();
return(ConvertSymbolicShapeToBarracuda(onnxShape, onnxLayout));
}
public static int[] ConvertSymbolicShapeToBarracuda(TensorShapeProto shape, string onnxLayout)
{
// TODO: use dimension denotation from TensorShapeProto to figure, if this particular tensor has specific data layout
// https://github.com/onnx/onnx/blob/master/docs/DimensionDenotation.md
var onnxShape = shape.Dim.Select(v => v.DimValue <int.MinValue?int.MinValue : v.DimValue> int.MaxValue ? int.MaxValue : (int)v.DimValue).ToArray();
return(ConvertSymbolicShapeToBarracuda(onnxShape, onnxLayout));
}
public async Task <IEnumerable <Detection> > Predict(Bitmap bmp)
{
if (_client == null)
{
throw new ApplicationException(nameof(_client));
}
// Desired image format
const int channels = 3;
const int width = 300;
const int height = 300;
const PixelFormat format = PixelFormat.Format24bppRgb;
var shape = new TensorShapeProto
{
Dim = { new []
{
new TensorShapeProto.Types.Dim {
Name = "", Size = 1
},
new TensorShapeProto.Types.Dim {
Name = nameof(height), Size = height
},
new TensorShapeProto.Types.Dim {
Name = nameof(width), Size = width
},
new TensorShapeProto.Types.Dim {
Name = nameof(channels), Size = channels
},
} }
};
var proto = new TensorProto
{
TensorShape = shape,
Dtype = Tensorflow.DataType.DtUint8,
TensorContent = ToByteString(bmp, channels, width, height, format)
};
var request = new PredictRequest
{
ModelSpec = new ModelSpec {
Name = _model
}
};
request.Inputs.Add("data", proto);
// Send requenst for inference
PredictResponse response = await _client.PredictAsync(request);
return(ToDetections(response));
}
public static TensorProto TensorProtoFromImage(byte[] imageData)
{
var imageFeatureShape = new TensorShapeProto();
imageFeatureShape.Dim.Add(new TensorShapeProto.Types.Dim()
{
Size = 1
});
var imageTensorBuilder = new TensorProto();
imageTensorBuilder.Dtype = DataType.DtFloat;
imageTensorBuilder.TensorShape = imageFeatureShape;
imageTensorBuilder.StringVal.Add(Google.Protobuf.ByteString.CopyFrom(imageData));
imageTensorBuilder.Dtype = DataType.DtString;
return(imageTensorBuilder);
}
public static TensorProto CreateTensorFromBuffer(IntPtr pixelBuffer, int height, int width, int channels)
{
var imageFeatureShape = new TensorShapeProto();
imageFeatureShape.Dim.Add(new TensorShapeProto.Types.Dim()
{
Size = 1
});
imageFeatureShape.Dim.Add(new TensorShapeProto.Types.Dim()
{
Size = height
});
imageFeatureShape.Dim.Add(new TensorShapeProto.Types.Dim()
{
Size = width
});
imageFeatureShape.Dim.Add(new TensorShapeProto.Types.Dim()
{
Size = channels
});
var imageTensorBuilder = new TensorProto();
imageTensorBuilder.Dtype = DataType.DtUint8;
imageTensorBuilder.TensorShape = imageFeatureShape;
unsafe
{
for (int i = 0; i < height; i++)
{
byte *pI = (byte *)pixelBuffer.ToPointer() + i * width * channels; //pointer to start of row
for (int j = 0; j < width; j++)
{
imageTensorBuilder.IntVal.Add(pI[j * channels]);
imageTensorBuilder.IntVal.Add(pI[j * channels + 1]);
imageTensorBuilder.IntVal.Add(pI[j * channels + 2]);
}
}
}
return(imageTensorBuilder);
}
public static TensorProto CreateTensorFromImage(uint[] pixels, int height, int width, int channels)
{
var imageFeatureShape = new TensorShapeProto();
imageFeatureShape.Dim.Add(new TensorShapeProto.Types.Dim()
{
Size = 1
});
imageFeatureShape.Dim.Add(new TensorShapeProto.Types.Dim()
{
Size = width
});
imageFeatureShape.Dim.Add(new TensorShapeProto.Types.Dim()
{
Size = height
});
imageFeatureShape.Dim.Add(new TensorShapeProto.Types.Dim()
{
Size = channels
});
var imageTensorBuilder = new TensorProto();
imageTensorBuilder.Dtype = DataType.DtUint8;
imageTensorBuilder.TensorShape = imageFeatureShape;
var px = 0;
for (int i = 0; i < width; ++i)
{
for (int j = 0; j < height; ++j)
{
var color = pixels[i + (j * width)];
imageTensorBuilder.IntVal.Add((byte)(0xFF & (color >> 16)));
imageTensorBuilder.IntVal.Add((byte)(0xFF & (color >> 8)));
imageTensorBuilder.IntVal.Add((byte)(0xFF & (color >> 0)));
}
}
return(imageTensorBuilder);
}
public static TensorProto CreateTensorFromImage(int[][][] dimArray, float revertsBits, int height, int width, int channels)
{
var imageFeatureShape = new TensorShapeProto();
imageFeatureShape.Dim.Add(new TensorShapeProto.Types.Dim()
{
Size = 1
});
imageFeatureShape.Dim.Add(new TensorShapeProto.Types.Dim()
{
Size = height
});
imageFeatureShape.Dim.Add(new TensorShapeProto.Types.Dim()
{
Size = width
});
imageFeatureShape.Dim.Add(new TensorShapeProto.Types.Dim()
{
Size = channels
});
var imageTensorBuilder = new TensorProto();
imageTensorBuilder.Dtype = DataType.DtUint8;
imageTensorBuilder.TensorShape = imageFeatureShape;
for (int i = 0; i < height; ++i)
{
for (int j = 0; j < width; ++j)
{
for (int c = 0; c < channels; c++)
{
//imageTensorBuilder.FloatVal.Add(dimArray[i][j][c] / revertsBits);
imageTensorBuilder.IntVal.Add(dimArray[i][j][c]);
}
}
}
return(imageTensorBuilder);
}
public static int[] ConvertSymbolicShapeToBarracuda(TensorShapeProto shape, string onnxLayout)
{
// TODO: use dimension denotation from TensorShapeProto to figure, if this particular tensor has specific data layout
// https://github.com/onnx/onnx/blob/master/docs/DimensionDenotation.md
return(ConvertSymbolicShapeToBarracuda(shape.Dim.Select(d => d.DimValue).ToArray(), onnxLayout));
}
public static PredictResponse ImageDetectionRequest(string image_url, out int height, out int width)
{
//Create gRPC Channel
var channel = new Channel(scoringServer, ChannelCredentials.Insecure,
new List <Grpc.Core.ChannelOption> {
new ChannelOption(ChannelOptions.MaxReceiveMessageLength, int.MaxValue),
new ChannelOption(ChannelOptions.MaxSendMessageLength, int.MaxValue)
});
var client = new PredictionService.PredictionServiceClient(channel);
//Create prediction request
var request = new PredictRequest()
{
ModelSpec = new ModelSpec()
{
Name = "ssd", SignatureName = "serving_default"
}
};
//Add image tensor
WebClient wc = new WebClient();
byte[] data = wc.DownloadData(image_url);
Stream stream = new MemoryStream(wc.DownloadData(image_url));
var dimArray = ImageUtils.ConvertImageStreamToDimArrays(stream);
height = dimArray.Length;
width = dimArray[0].Length;
var channels = dimArray[0][0].Length;
var imageTensorBuilder = new TensorProto();
var imageFeatureShape = new TensorShapeProto();
imageFeatureShape.Dim.Add(new TensorShapeProto.Types.Dim()
{
Size = 1
});
imageFeatureShape.Dim.Add(new TensorShapeProto.Types.Dim()
{
Size = height
});
imageFeatureShape.Dim.Add(new TensorShapeProto.Types.Dim()
{
Size = width
});
imageFeatureShape.Dim.Add(new TensorShapeProto.Types.Dim()
{
Size = channels
});
imageTensorBuilder.Dtype = DataType.DtUint8;
imageTensorBuilder.TensorShape = imageFeatureShape;
for (int i = 0; i < height; ++i)
{
for (int j = 0; j < width; ++j)
{
for (int c = 0; c < channels; c++)
{
//imageTensorBuilder.FloatVal.Add(dimArray[i][j][c] / revertsBits);
imageTensorBuilder.IntVal.Add(dimArray[i][j][c]);
}
}
}
request.Inputs.Add("inputs", imageTensorBuilder);
//using (Stream stream = new MemoryStream(wc.DownloadData(image_url)))
//{
// request.Inputs.Add("inputs", TensorBuilder.CreateTensorFromImage(stream, 1.0f));
//}
// Run the prediction
var predictResponse = client.Predict(request);
//Console.WriteLine(predictResponse.Outputs["detection_classes"]);
//Console.WriteLine(predictResponse.Outputs["detection_boxes"]);
//Console.WriteLine(predictResponse.Outputs["detection_scores"]);
return(predictResponse);
}
public static int[] AsInts(this TensorShapeProto shape)
{
return(shape.Dim.Select(v => v.DimValue <int.MinValue?int.MinValue : v.DimValue> int.MaxValue ? int.MaxValue : (int)v.DimValue).ToArray());
}
