internal static Tensor Permute(Tensor inTensor, int[] permutations) // TODO: unify Permute() arguments
{
var padPermutationsToBarracudaRank = TensorShape.MaxRank - permutations.Length;
if (padPermutationsToBarracudaRank > 0)
{
permutations = permutations.Concat(Enumerable.Range(permutations.Length, padPermutationsToBarracudaRank)).ToArray();
}
Debug.Assert(permutations.Length == TensorShape.MaxRank);
// See: https://stackoverflow.com/a/32034565
Profiler.BeginSample("ONNXTensor.Permute");
var outTensor = new Tensor(ONNXLayout.Permute(inTensor.shape.ToArray(), permutations));
Debug.Assert(outTensor.length == inTensor.length);
// {0, 2, 3, 1} => {0, 3, 1, 2}
// {2, 3, 1, 0} => {3, 2, 0, 1}
// => {find_index(0), find_index(1), find_index(2), find_index(3)}
var reversePermute = new int[permutations.Length];
for (var i = 0; i < permutations.Length; ++i)
{
reversePermute[i] = Array.IndexOf(permutations, i);
}
// outTensor strides
var tempOutStrides = new int[TensorShape.MaxRank + 1];
tempOutStrides[8] = 1;
for (int i = 7; i >= 0; --i)
{
tempOutStrides[i] = tempOutStrides[i + 1] * outTensor.shape[i];
}
var outStride = new int[reversePermute.Length];
for (var i = 0; i < reversePermute.Length; ++i)
{
outStride[i] = tempOutStrides[reversePermute[i] + 1];
}
// inTensor strides
var inStrides = new int[TensorShape.MaxRank];
inStrides[7] = 1;
for (int i = 6; i >= 0; --i)
{
inStrides[i] = inStrides[i + 1] * inTensor.shape[i + 1];
}
for (var it = new TensorIterator(inTensor.shape); it.IsValid(); ++it)
{
outTensor[it.d0 * outStride[0] + it.d1 * outStride[1] + it.d2 * outStride[2] + it.d3 * outStride[3] + it.d4 * outStride[4] + it.d5 * outStride[5] + it.d6 * outStride[6] + it.d7 * outStride[7]] = inTensor[
it.d0 * inStrides[0] + it.d1 * inStrides[1] + it.d2 * inStrides[2] + it.d3 * inStrides[3] + it.d4 * inStrides[4] + it.d5 * inStrides[5] + it.d6 * inStrides[6] + it.d7 * inStrides[7]];
}
Profiler.EndSample();
return(outTensor);
}
public void AddVariable(string nodeId, long[] onnxShape, string onnxLayout)
{
var onnxRank = onnxShape.Length;
var permuatations = ONNXLayout.AxisPermutationsForMappingONNXLayoutToBarracuda(onnxRank, onnxLayout);
var barracudaChannelIndex = permuatations.Length - 1;
var onnxChannelIndex = permuatations[barracudaChannelIndex];
var channels = (onnxLayout != "?" && onnxChannelIndex >= 0) ? (int)onnxShape[onnxChannelIndex]: -1;
var layout = VariableTensor.Layout.Unknown;
if (onnxLayout == "NCHW")
{
layout = VariableTensor.Layout.NCHW;
}
else if (onnxLayout == "NHWC")
{
layout = VariableTensor.Layout.NHWC;
}
variables[nodeId] = new VariableTensor {
features = channels,
rank = onnxRank,
layout = layout,
productOfShape = null
};
}
public ONNXTensor Permute(int[] permutations)
{
// transpose both data & shape
var transposedData = Permute(m_Data, permutations);
var transposedShape = ONNXLayout.Permute(m_Shape, permutations);
return(new ONNXTensor(transposedData, transposedShape));
}
// slow version - kept just for performance comparison and validation
internal static Tensor PermuteSlow(Tensor readTensor, int[] permutations) // TODO: unify Permute() arguments
{
var padPermutationsToBarracudaRank = 8 - permutations.Length;
if (padPermutationsToBarracudaRank > 0)
{
permutations = permutations.Concat(Enumerable.Range(permutations.Length, padPermutationsToBarracudaRank)).ToArray();
}
Assert.IsTrue(permutations.Length == 8);
var outputTensor = new Tensor(ONNXLayout.Permute(readTensor.shape.ToArray(), permutations));
Assert.IsTrue(outputTensor.length == readTensor.length);
var inShape = readTensor.shape.ToArray();
for (var s = 0; s < inShape[0]; ++s)
{
for (var n = 0; n < inShape[1]; ++n)
{
for (var i0 = 0; i0 < inShape[2]; ++i0)
{
for (var i1 = 0; i1 < inShape[3]; ++i1)
{
for (var i2 = 0; i2 < inShape[4]; ++i2)
{
for (var h = 0; h < inShape[5]; ++h)
{
for (var w = 0; w < inShape[6]; ++w)
{
for (var c = 0; c < inShape[7]; ++c)
{
var it = new int[] { 0, s, n, i0, i1, i2, h, w, c }; // prepend with 0 to handle "new axis" -1 value in permutations
var oS = it[permutations[0] + 1];
var oN = it[permutations[1] + 1];
var oI0 = it[permutations[2] + 1];
var oI1 = it[permutations[3] + 1];
var oI2 = it[permutations[4] + 1];
var oH = it[permutations[5] + 1];
var oW = it[permutations[6] + 1];
var oC = it[permutations[7] + 1];
outputTensor[oS, oN, oI0, oI1, oI2, oH, oW, oC] = readTensor[s, n, i0, i1, i2, h, w, c];
}
}
}
}
}
}
}
}
return(outputTensor);
}
public ONNXTensor Reshape(int[] onnxShape)
{
var symbolicShape = ONNXLayout.ConvertSymbolicShapeToBarracuda(onnxShape, "?");
var reshapedData = m_Data.Reshape(symbolicShape);
for (var i = 0; i < onnxShape.Length; ++i)
{
if (onnxShape[i] < 0)
{
onnxShape[i] = reshapedData.shape[i];
}
Assert.IsTrue(onnxShape[i] == reshapedData.shape[i]);
}
return(new ONNXTensor(reshapedData, onnxShape));
}
public Tensor ToBarracuda(string onnxLayout)
{
Profiler.BeginSample("ONNXTensor.ToBarracuda");
if (onnxLayout == "?")
{
throw new OnnxLayerImportException("Unknown ONNX layout in not supported when converting constant tensor to Barracuda");
}
Assert.IsTrue(m_Shape.All(v => v > 0));
var permutations = ONNXLayout.AxisPermutationsForMappingONNXLayoutToBarracuda(rank, onnxLayout);
Assert.IsTrue(rank <= permutations.Length);
var outTensor = Permute(m_Data, permutations);
Profiler.EndSample();
return(outTensor);
}
public ONNXTensor(TensorProto onnxTensor)
{
// read shape
var onnxShape = onnxTensor.Dims.Select(v => v <int.MinValue?int.MinValue : v> int.MaxValue ? int.MaxValue : (int)v).ToArray();
if (onnxShape.Any(s => s == 0))
{
// empty tensor, not data
m_Shape = onnxShape;
m_Data = null;
}
else
{
// read data
var shape = ONNXLayout.ConvertShapeToBarracuda(onnxShape, onnxLayout: "?");
float[] data;
if ((onnxTensor.RawData != null) && (!onnxTensor.RawData.IsEmpty))
{
var byteArray = new byte[onnxTensor.RawData.Length];
onnxTensor.RawData.CopyTo(byteArray, 0);
// Double
if (onnxTensor.DataType == (int)TensorProto.Types.DataType.Double)
{
var typedData = new double[shape.length];
Assert.IsTrue((sizeof(double) * shape.length) == onnxTensor.RawData.Length);
Buffer.BlockCopy(byteArray, 0, typedData, 0, byteArray.Length);
data = typedData.Select(v => v <int.MinValue?(float)int.MinValue : v> int.MaxValue ? (float)int.MaxValue : (float)v).ToArray();
}
// Float32
else if (onnxTensor.DataType == (int)TensorProto.Types.DataType.Float)
{
data = new float[shape.length];
Assert.IsTrue((sizeof(float) * shape.length) == onnxTensor.RawData.Length);
Buffer.BlockCopy(byteArray, 0, data, 0, byteArray.Length);
}
// Float16
else if (onnxTensor.DataType == (int)TensorProto.Types.DataType.Float16)
{
var typedData = new UInt16[shape.length];
Assert.IsTrue((sizeof(UInt16) * shape.length) == onnxTensor.RawData.Length);
Buffer.BlockCopy(byteArray, 0, typedData, 0, byteArray.Length);
data = typedData.Select(v => HalfHelper.HalfToSingle(v)).ToArray();
}
// Int8
else if (onnxTensor.DataType == (int)TensorProto.Types.DataType.Int8)
{
var typedData = new sbyte[shape.length];
Assert.IsTrue((sizeof(sbyte) * shape.length) == onnxTensor.RawData.Length);
Buffer.BlockCopy(byteArray, 0, typedData, 0, byteArray.Length);
data = typedData.Select(v => (float)v).ToArray();
}
// Int16
else if (onnxTensor.DataType == (int)TensorProto.Types.DataType.Int16)
{
var typedData = new short[shape.length];
Assert.IsTrue((sizeof(short) * shape.length) == onnxTensor.RawData.Length);
Buffer.BlockCopy(byteArray, 0, typedData, 0, byteArray.Length);
data = typedData.Select(v => (float)v).ToArray();
}
// Int32
else if (onnxTensor.DataType == (int)TensorProto.Types.DataType.Int32)
{
var typedData = new int[shape.length];
Assert.IsTrue((sizeof(int) * shape.length) == onnxTensor.RawData.Length);
Buffer.BlockCopy(byteArray, 0, typedData, 0, byteArray.Length);
data = typedData.Select(v => (float)v).ToArray();
}
// Int64
else if (onnxTensor.DataType == (int)TensorProto.Types.DataType.Int64)
{
var typedData = new long[shape.length];
Assert.IsTrue((sizeof(long) * shape.length) == onnxTensor.RawData.Length);
Buffer.BlockCopy(byteArray, 0, typedData, 0, byteArray.Length);
data = typedData.Select(v => v <int.MinValue?(float)int.MinValue : v> int.MaxValue ? (float)int.MaxValue : (float)v).ToArray();
}
// UInt8
else if (onnxTensor.DataType == (int)TensorProto.Types.DataType.Uint8)
{
var typedData = new byte[shape.length];
Assert.IsTrue((sizeof(byte) * shape.length) == onnxTensor.RawData.Length);
Buffer.BlockCopy(byteArray, 0, typedData, 0, byteArray.Length);
data = typedData.Select(v => (float)v).ToArray();
}
// UInt16
else if (onnxTensor.DataType == (int)TensorProto.Types.DataType.Uint16)
{
var typedData = new ushort[shape.length];
Assert.IsTrue((sizeof(ushort) * shape.length) == onnxTensor.RawData.Length);
Buffer.BlockCopy(byteArray, 0, typedData, 0, byteArray.Length);
data = typedData.Select(v => (float)v).ToArray();
}
// UInt32
else if (onnxTensor.DataType == (int)TensorProto.Types.DataType.Uint32)
{
var typedData = new uint[shape.length];
Assert.IsTrue((sizeof(uint) * shape.length) == onnxTensor.RawData.Length);
Buffer.BlockCopy(byteArray, 0, typedData, 0, byteArray.Length);
data = typedData.Select(v => (float)v).ToArray();
}
// UInt64
else if (onnxTensor.DataType == (int)TensorProto.Types.DataType.Uint64)
{
var typedData = new ulong[shape.length];
Assert.IsTrue((sizeof(ulong) * shape.length) == onnxTensor.RawData.Length);
Buffer.BlockCopy(byteArray, 0, typedData, 0, byteArray.Length);
data = typedData.Select(v => v > uint.MaxValue ? (float)uint.MaxValue : (float)v).ToArray();
}
// Bool
else if (onnxTensor.DataType == (int)TensorProto.Types.DataType.Bool)
{
var typedData = new bool[shape.length];
Assert.IsTrue((sizeof(bool) * shape.length) == onnxTensor.RawData.Length);
Buffer.BlockCopy(byteArray, 0, typedData, 0, byteArray.Length);
data = typedData.Select(v => v ? 1.0f : 0.0f).ToArray();
}
else
{
throw new OnnxLayerImportException($"Tensor data type {(TensorProto.Types.DataType)onnxTensor.DataType} is not supported.");
}
}
// Float32
else if ((onnxTensor.FloatData != null) && (onnxTensor.FloatData.Count != 0))
{
Assert.IsTrue(shape.length == onnxTensor.FloatData.Count);
data = new float[shape.length];
onnxTensor.FloatData.CopyTo(data, 0);
}
// Int32
else if ((onnxTensor.Int32Data != null) && (onnxTensor.Int32Data.Count != 0))
{
Assert.IsTrue(shape.length == onnxTensor.Int32Data.Count);
data = onnxTensor.Int32Data.Select(v => (float)v).ToArray();
}
// Int64
else if ((onnxTensor.Int64Data != null) && (onnxTensor.Int64Data.Count != 0))
{
Assert.IsTrue(shape.length == onnxTensor.Int64Data.Count);
data = onnxTensor.Int64Data.Select(v => v <int.MinValue?(float)int.MinValue : v> int.MaxValue ? (float)int.MaxValue : (float)v).ToArray();
}
else
{
throw new OnnxLayerImportException("Could not read tensor data for constant tensor.");
}
m_Data = new Tensor(shape, new SharedArrayTensorData(data));
m_Shape = onnxShape;
}
}