public override Tensor ElementwiseWithBroadcast(string kernelName, Tensor[] tensors)
{
Assert.IsTrue(tensors.Length > 0);
Tensor outputTensor1 = NewTensor(TensorExtensions.MaxShape(tensors));
Tensor outputTensor2 = null;
if (tensors.Length > 2)
{
outputTensor2 = NewTensor(TensorExtensions.MaxShape(tensors));
}
var X = tensors[0];
var fn = BestKernel(ComputeKernelLibrary.Broadcast(X.shape, outputTensor1.shape, kernelName));
Tensor O = null;
for (int t = 1; t < tensors.Length; ++t)
{
var B = tensors[t];
O = (t % 2 == 1)?outputTensor1:outputTensor2;
fn.SetTensor("X", X.shape, Pin(X).buffer);
fn.SetTensor("O", O.shape, Pin(O).buffer);
fn.SetTensor("B", B.shape, Pin(B).buffer, Pin(B).offset);
fn.Dispatch();
X = O;
}
return(O);
}
public override Tensor Concat(Tensor[] tensors, int axis)
{
var O = NewTensor(TensorExtensions.Concat(tensors.Select(t => t.shape).ToArray(), axis));
var offsets = new int[] { 0, 0, 0, 0 };
axis = O.shape.Axis(axis);
foreach (var X in tensors)
{
var fn = BestKernel(ComputeKernelLibrary.Copy(X.shape, O.shape));
fn.SetTensor("X", X.shape, Pin(X).buffer);
fn.SetTensor("O", O.shape, Pin(O).buffer);
fn.shader.SetInts("_Pad", offsets);
fn.Dispatch();
offsets[axis] += X.shape[axis];
}
return(O);
}
static public TensorShape[] ListTemporaryTensorShapes(Model model, IDictionary <string, TensorShape> inputShapes,
out IDictionary <string, TensorShape> shapesByName)
{
Profiler.BeginSample("Barracuda.ListTemporaryTensorShapes");
var shapes = new List <TensorShape>();
shapesByName = new Dictionary <string, TensorShape>();
foreach (var entry in inputShapes)
{
shapesByName.Add(entry.Key, entry.Value);
}
TensorShape X;
shapesByName.TryGetValue(GetDefaultInputName(model), out X); // default input
var O = X;
foreach (var l in model.layers)
{
if (l.inputs.Length > 0 && shapesByName.ContainsKey(l.inputs[0]))
{
X = shapesByName[l.inputs[0]];
}
else
{
X = O; // previous output is used, if-and-only-if layer has no explicit inputs
}
if (l.type == Layer.Type.Dense)
{
Assert.IsNotNull(l.datasets);
var W = l.datasets[0].shape;
O = new TensorShape(X.flatHeight, W.flatWidth);
}
else if (
l.type == Layer.Type.Conv2D ||
l.type == Layer.Type.DepthwiseConv2D)
{
var K = l.datasets[0].shape;
Assert.IsNotNull(l.stride);
Assert.IsNotNull(l.pad);
var pad = X.AdjustPadToKernel(K, l.stride, l.pad);
O = X.ApplyKernel(K, l.stride, pad);
}
else if (
l.type == Layer.Type.Conv2DTrans)
{
var K = l.datasets[0].shape;
Assert.IsNotNull(l.stride);
Assert.IsNotNull(l.pad);
// pool size is treated as output_adjustment aka output_padding here
var outputAdjustment = l.pool;
var pad = X.AdjustPadToKernel(K, l.stride, l.pad);
O = X.ApplyKernelInverse(K, l.stride, pad, outputAdjustment);
}
else if (
l.type == Layer.Type.Upsample2D)
{
// pool size is treated as upsample coefficient here
Assert.IsNotNull(l.pool);
Assert.AreEqual(l.pool.Length, 2);
O = new TensorShape(X.batch, X.height * l.pool[1], X.width * l.pool[0], X.channels);
}
else if (
l.type == Layer.Type.MaxPool2D ||
l.type == Layer.Type.AvgPool2D)
{
Assert.IsNotNull(l.pool);
Assert.IsNotNull(l.stride);
Assert.IsNotNull(l.pad);
var pad = X.AdjustPadToPool(l.pool, l.stride, l.pad);
O = X.ApplyPool(l.pool, l.stride, pad);
}
else if (
l.type == Layer.Type.GlobalMaxPool2D ||
l.type == Layer.Type.GlobalAvgPool2D)
{
O = new TensorShape(X.batch, 1, 1, X.channels);
}
else if (
l.type == Layer.Type.Border2D ||
l.type == Layer.Type.Pad2DReflect ||
l.type == Layer.Type.Pad2DSymmetric ||
l.type == Layer.Type.Pad2DEdge)
{
Assert.IsNotNull(l.pad);
O = X.ApplyBorder(l.pad);
}
else if (
l.type == Layer.Type.Conv3D ||
l.type == Layer.Type.Conv3DTrans ||
l.type == Layer.Type.Upsample3D ||
l.type == Layer.Type.MaxPool3D ||
l.type == Layer.Type.AvgPool3D ||
l.type == Layer.Type.GlobalMaxPool3D ||
l.type == Layer.Type.GlobalAvgPool3D ||
l.type == Layer.Type.Border3D)
{
throw new NotImplementedException();
}
else if (
l.type == Layer.Type.RandomNormal ||
l.type == Layer.Type.RandomUniform)
{
Assert.IsNotNull(l.pool);
// pool size is treated as shape constant, if not empty
// otherwise shape of the previous tensor is used
if (l.pool.Length > 0)
{
O = new TensorShape(l.pool);
}
else
{
O = X;
}
}
else if (
l.type == Layer.Type.Multinomial)
{
Assert.IsNotNull(l.pool);
Assert.AreEqual(l.pool.Length, 1);
O = new TensorShape(X.batch, l.pool[0]);
}
else if (
l.type == Layer.Type.OneHot)
{
Assert.IsNotNull(l.pool);
Assert.AreEqual(l.pool.Length, 1);
int features = X.flatWidth;
int depth = l.pool[0];
O = new TensorShape(X.batch, 1, features, depth);
}
else if (
l.type == Layer.Type.Add ||
l.type == Layer.Type.Sub ||
l.type == Layer.Type.Mul ||
l.type == Layer.Type.Div ||
l.type == Layer.Type.Pow ||
l.type == Layer.Type.Min ||
l.type == Layer.Type.Max ||
l.type == Layer.Type.Mean ||
l.type == Layer.Type.Greater ||
l.type == Layer.Type.GreaterEqual ||
l.type == Layer.Type.Less ||
l.type == Layer.Type.LessEqual ||
l.type == Layer.Type.Equal ||
l.type == Layer.Type.LogicalOr ||
l.type == Layer.Type.LogicalAnd ||
l.type == Layer.Type.LogicalXor)
{
// gather shapes by names
var list = new List <TensorShape>(l.inputs.Length);
foreach (var i in l.inputs)
{
if (shapesByName.ContainsKey(i))
{
list.Add(shapesByName[i]);
}
}
O = TensorExtensions.Max(list.ToArray());
}
else if (
l.type == Layer.Type.ReduceL1 ||
l.type == Layer.Type.ReduceL2 ||
l.type == Layer.Type.ReduceLogSum ||
l.type == Layer.Type.ReduceLogSumExp ||
l.type == Layer.Type.ReduceMax ||
l.type == Layer.Type.ReduceMean ||
l.type == Layer.Type.ReduceMin ||
l.type == Layer.Type.ReduceProd ||
l.type == Layer.Type.ReduceSum ||
l.type == Layer.Type.ReduceSumSquare)
{
O = X.Reduce(l.axis);
}
else if (
l.type == Layer.Type.Flatten)
{
O = X.Flatten();
}
else if (
l.type == Layer.Type.Reshape)
{
// pool size is treated as reshape coefficient, if not empty
// otherwise shape of the 2nd input tensor is used
var size = l.pool;
Assert.IsNotNull(size);
if (size.Length == 0 && l.inputs.Length > 1)
{
size = shapesByName[l.inputs[1]].ToArray();
}
Assert.AreEqual(size.Length, 4);
// pool size is treated as reshape coefficient here
O = X.Reshape(size);
}
else if (
l.type == Layer.Type.Transpose)
{
O = new TensorShape(X.flatWidth, X.flatHeight);
}
else if (
l.type == Layer.Type.Squeeze ||
l.type == Layer.Type.Unsqueeze)
{
throw new NotImplementedException();
}
else if (
l.type == Layer.Type.Concat)
{
// gather shapes by names
var list = new List <TensorShape>(l.inputs.Length);
foreach (var i in l.inputs)
{
if (shapesByName.ContainsKey(i))
{
list.Add(shapesByName[i]);
}
}
O = TensorExtensions.Concat(list.ToArray(), l.axis);
}
else if (
l.type == Layer.Type.StridedSlice)
{
Assert.IsNotNull(l.pad);
Assert.IsNotNull(l.pool);
Assert.IsNotNull(l.stride);
O = X.ApplyStridedSlice(l.pad, l.pool, l.stride);
}
else if (
l.type == Layer.Type.Tile)
{
// pool size is treated as tiling coefficient here
Assert.IsNotNull(l.pool);
Assert.AreEqual(l.pool.Length, 4);
var scale = l.pool;
O = X.Scale(scale);
}
else if (
l.type == Layer.Type.Load)
{
O = l.datasets[0].shape;
}
else if (// elementwise operations
l.type == Layer.Type.Nop ||
l.type == Layer.Type.Activation ||
l.type == Layer.Type.ScaleBias ||
l.type == Layer.Type.Normalization ||
l.type == Layer.Type.LRN ||
l.type == Layer.Type.Dropout ||
l.type == Layer.Type.LogicalNot ||
l.activation == Layer.Activation.PRelu)
{
// works in place, keeps the same shape size
O = X;
}
else
{
throw new NotImplementedException();
}
shapes.Add(O);
shapesByName.Add(l.name, O);
}
Profiler.EndSample();
return(shapes.ToArray());
}