/// <summary>
/// Get output as `Tensor`
/// </summary>
/// <param name="idx">output index</param>
/// <param name="batchCount">max batch count</param>
/// <param name="fromBatch">start from batch</param>
/// <returns>`Tensor`</returns>
/// <exception cref="Exception">thrown if called on raw test set (only JSON test set is supported)</exception>
public Tensor GetOutputAsTensor(int idx = 0, int batchCount = -1, int fromBatch = 0)
{
if (rawTestSet != null)
{
throw new Exception("GetOutputAsTensor is not supported for RAW test suites");
}
TensorShape shape = GetOutputShape(idx);
Assert.IsTrue(shape.sequenceLength == 1 && shape.numberOfDirections == 1);
var barracudaArray = new BarracudaArrayFromManagedArray(GetOutputData(idx));
var maxBatchCount = barracudaArray.Length / shape.flatWidth;
fromBatch = Math.Min(fromBatch, maxBatchCount - 1);
if (batchCount < 0)
{
batchCount = maxBatchCount - fromBatch;
}
batchCount = Math.Min(batchCount, maxBatchCount - fromBatch);
var shapeArray = shape.ToArray();
shapeArray[TensorShape.DataBatch] = batchCount;
var tensorShape = new TensorShape(shapeArray);
var offset = fromBatch * tensorShape.flatWidth;
var res = new Tensor(tensorShape, new SharedArrayTensorData(barracudaArray, tensorShape, offset));
res.name = GetOutputName(idx);
res.name = res.name.EndsWith(":0") ? res.name.Remove(res.name.Length - 2) : res.name;
return(res);
}
/// <summary>
/// Add an input to the model
/// </summary>
public Model.Input Input(string name, TensorShape shape)
{
m_Model.inputs.Add(new Model.Input {
name = name, shape = shape.ToArray()
});
return(m_Model.inputs.Last());
}
static public TensorShape Gather(TensorShape[] shapes, int axis)
{
TensorShape X = shapes[0];
TensorShape indices = shapes[1];
int[] shape = X.ToArray();
shape[axis] = indices.length;
return(new TensorShape(shape));
}
/// <summary>
/// Generates a Tensor with random values drawn from a uniform distribution.
/// The shape of the tensor is specified by shape
/// The uniform distribution scale is specified by min and max range
/// </summary>
public Layer RandomUniform(string name, TensorShape shape, float min, float max, float seed)
{
Layer layer = new Layer(name, Layer.Type.RandomUniform);
layer.alpha = (max - min);
layer.beta = min;
layer.pad = new int[1] {
(int)seed
};
layer.pool = shape.ToArray();
m_Model.layers.Add(layer);
return(layer);
}
/// <summary>
/// Generates a Tensor with random values drawn from a normal distribution.
/// The shape of the tensor is specified by scale
/// The normal distribution is specified by mean and scale
/// </summary>
public Layer RandomNormal(string name, TensorShape shape, float mean, float scale, float seed)
{
Layer layer = new Layer(name, Layer.Type.RandomNormal);
layer.alpha = scale;
layer.beta = mean;
layer.pad = new int[1] {
(int)seed
};
layer.pool = shape.ToArray();
m_Model.layers.Add(layer);
return(layer);
}
/// <summary>
/// Get input as `Tensor`
/// </summary>
/// <param name="idx">input index</param>
/// <param name="batchCount">max batch count</param>
/// <param name="fromBatch">start from batch</param>
/// <returns>`Tensor`</returns>
/// <exception cref="Exception">thrown if called on raw test set (only JSON test set is supported)</exception>
public Tensor GetInputAsTensor(int idx = 0, int batchCount = -1, int fromBatch = 0)
{
if (rawTestSet != null)
{
throw new Exception("GetInputAsTensor is not supported for RAW test suites");
}
TensorShape shape = GetInputShape(idx);
Assert.IsTrue(shape.sequenceLength == 1 && shape.numberOfDirections == 1);
var array = GetInputData(idx);
var maxBatchCount = array.Length / shape.flatWidth;
fromBatch = Math.Min(fromBatch, maxBatchCount - 1);
if (batchCount < 0)
{
batchCount = maxBatchCount - fromBatch;
}
// pad data with 0s, if test-set doesn't have enough batches
var shapeArray = shape.ToArray();
shapeArray[TensorShape.DataBatch] = batchCount;
var tensorShape = new TensorShape(shapeArray);
var managedBufferStartIndex = fromBatch * tensorShape.flatWidth;
var count = Math.Min(batchCount, maxBatchCount - fromBatch) * tensorShape.flatWidth;
float[] dataToUpload = new float[tensorShape.length];
Array.Copy(array, managedBufferStartIndex, dataToUpload, 0, count);
var data = new ArrayTensorData(tensorShape.length);
data.Upload(dataToUpload, tensorShape, 0);
var res = new Tensor(tensorShape, data);
res.name = GetInputName(idx);
res.name = res.name.EndsWith(":0") ? res.name.Remove(res.name.Length - 2) : res.name;
return(res);
}
static public TensorShape Reshape(this TensorShape shape, int[] size)
{
Assert.AreEqual(size.Length, 4);
var newShapeArray = shape.ToArray();
// From: https://github.com/onnx/onnx/blob/master/docs/Operators.md#Reshape
//
// At most one dimension of the new shape can be -1.
// In this case, the value is inferred from the size of the tensor and the remaining dimensions.
//
// A dimension could also be 0,
// in which case the actual dimension value is unchanged (i.e. taken from the input tensor).
var multipleOf = 1;
var unknownIndex = -1;
for (int q = 0; q < size.Length; ++q)
{
if (size[q] > 0)
{
multipleOf *= size[q];
newShapeArray[q] = size[q];
}
else if (size[q] == 0)
{
multipleOf *= newShapeArray[q];
}
else if (unknownIndex == -1)
{
unknownIndex = q;
}
else
{
throw new ArgumentException("Can only specify one unknown dimension");
}
}
if (unknownIndex == -1)
{
// all dimensions are given
var newShape = new TensorShape(newShapeArray);
if (shape.length != newShape.length)
{
throw new ArgumentException("Cannot reshape array of size " + shape.length +
" into shape " + newShape);
}
return(newShape);
}
var solveForIndex = shape.length / multipleOf;
bool remainderLeft = shape.length % multipleOf != 0;
if (remainderLeft)
{
throw new ArgumentException("Cannot reshape array of size " + shape.length +
" into shape with multiple of " + multipleOf + " elements");
}
newShapeArray[unknownIndex] = solveForIndex;
return(new TensorShape(newShapeArray));
}
/// <summary>
/// Apply shape to the input tensor. Number of elements in the shape must match number of elements in input tensor.
/// </summary>
public Layer Reshape(string name, object input, TensorShape shape)
{
return(Reshape(name, input, shape.ToArray()));
}
