Tensor IOps.Expand(Tensor X, TensorShape shape)
{
var O = m_Ops.Expand(X, shape);
m_Mem += (long)X.length + (long)O.length;
return(O);
}
/// <inheritdoc/>
Tensor IOps.Expand(Tensor X, TensorShape shape)
{
var O = m_Ops.Expand(X, shape);
LogLayerSummary(X.shape + " $ " + O.shape);
return(O);
}
/// <inheritdoc/>
Tensor IOps.Expand(Tensor X, TensorShape shape)
{
var O = m_Ops.Expand(X, shape);
D.Log(X.shape + " $ " + O.shape);
return(O);
}
/// <inheritdoc/>
Tensor IOps.Expand(Tensor X, TensorShape shape)
{
var Y = m_Ops1.Expand(X, shape);
var Z = m_Ops2.Expand(X, shape);
CheckSame(Y, Z, Layer.Type.Expand);
return(Y);
}
/// <inheritdoc/>
Tensor IOps.Expand(Tensor X, TensorShape shape)
{
var O = m_Ops.Expand(X, shape);
m_Alu += 0;
m_Mem += (long)X.length + (long)O.length;
RegisterLayerStats();
return(O);
}
public virtual IEnumerator StartManualSchedule()
{
Profiler.BeginSample("Barracuda.Execute");
ResetAllocatorIfRequested();
m_Vars.PrepareStorage(m_Model, m_Ops, m_InputShapes);
if (m_ModelCompiler != null)
{
m_ModelCompiler.PrepareModel(m_Model, m_InputShapes);
}
int idx = 0;
foreach (var l in m_Model.layers)
{
idx++;
m_Progress = idx / (float)m_Model.layers.Count;
Profiler.BeginSample(l.name);
var inputs = m_Vars.GatherInputs(l);
Tensor X = inputs.Length > 0 ? inputs[0] : new Tensor();
if (m_Verbose)
{
D.Log("Layer: " + l.type + ((l.type == Layer.Type.Activation) ? ("." + l.activation) : "") + " " + l.name);
}
m_Vars.PrepareStorage(l);
if (m_ModelCompiler != null)
{
m_ModelCompiler.PreExecuteLayer(l, inputs);
}
// No operation, identity
if (l.type == Layer.Type.Nop)
{
Profiler.BeginSample("Barracuda.Nop");
X = m_Ops.Copy(X);
}
// Load const
else if (l.type == Layer.Type.Load)
{
Profiler.BeginSample("Barracuda.Load");
}
// GEMM
else if (l.type == Layer.Type.Dense)
{
Assert.AreEqual(inputs.Length, 3);
Profiler.BeginSample("Barracuda.Dense");
X = m_Ops.Dense(X, inputs[1], inputs[2], GetAndVerifyFusedActivation(l));
}
// 2D
else if (l.type == Layer.Type.Conv2D)
{
Assert.AreEqual(inputs.Length, 3);
Profiler.BeginSample("Barracuda.Conv2D");
var pad = X.AdjustPadToKernel(inputs[1], l.stride, l.pad);
X = m_Ops.Conv2D(X, inputs[1], inputs[2], l.stride, pad, GetAndVerifyFusedActivation(l));
}
else if (l.type == Layer.Type.DepthwiseConv2D)
{
Assert.AreEqual(inputs.Length, 3);
Profiler.BeginSample("Barracuda.DepthwiseConv2D");
var pad = X.AdjustPadToKernel(inputs[1], l.stride, l.pad);
X = m_Ops.DepthwiseConv2D(X, inputs[1], inputs[2], l.stride, pad, GetAndVerifyFusedActivation(l));
}
else if (l.type == Layer.Type.Conv2DTrans)
{
Assert.AreEqual(inputs.Length, 3);
Profiler.BeginSample("Barracuda.Conv2DTrans");
// pool size is treated as output_adjustment aka output_padding here
var outputAdjustment = l.pool;
var pad = X.AdjustPadToKernel(inputs[1], l.stride, l.pad);
X = m_Ops.Conv2DTrans(X, inputs[1], inputs[2], l.stride, pad, outputAdjustment, GetAndVerifyFusedActivation(l));
}
else if (l.type == Layer.Type.Upsample2D)
{
Profiler.BeginSample("Barracuda.Upsample2D");
// pool size is treated as upsample scale coefficient here
var scale = l.pool;
// axis is treated as upsample point/bilinear flag
var bilinear = l.axis > 0;
if (scale.Length == 0 && inputs.Length > 1)
{
var scaleTensor = inputs[1];
Assert.AreEqual(scaleTensor.length, 4);
scale = new int[] { (int)scaleTensor[2], (int)scaleTensor[1] };
}
X = m_Ops.Upsample2D(X, scale, bilinear);
}
else if (l.type == Layer.Type.Resample2D)
{
Profiler.BeginSample("Barracuda.Resample2D");
// pool size is treated as resample size here
var size = l.pool;
// axis is treated as upsample point/bilinear flag
var bilinear = l.axis > 0;
if (inputs.Length > 1)
{
var sizeTensor = inputs[1];
Assert.AreEqual(sizeTensor.length, 4);
size = new int[] { (int)sizeTensor[2], (int)sizeTensor[1] };
}
X = m_Ops.Resample2D(X, size, bilinear);
}
else if (l.type == Layer.Type.DepthToSpace)
{
Profiler.BeginSample("Barracuda.DepthToSpace");
// pool size is treated as blocksize
var blocksize = l.pool;
// axis is treated as mode enum
var mode = (Layer.DepthToSpaceMode)l.axis;
X = m_Ops.DepthToSpace(X, blocksize, mode);
}
else if (l.type == Layer.Type.SpaceToDepth)
{
Profiler.BeginSample("Barracuda.SpaceToDepth");
// pool size is treated as blocksize
var blocksize = l.pool;
X = m_Ops.SpaceToDepth(X, blocksize);
}
else if (l.type == Layer.Type.MaxPool2D)
{
Profiler.BeginSample("Barracuda.MaxPool2D");
var pad = X.AdjustPadToPool(l.pool, l.stride, l.pad);
X = m_Ops.MaxPool2D(X, l.pool, l.stride, pad);
}
else if (l.type == Layer.Type.AvgPool2D)
{
Profiler.BeginSample("Barracuda.AvgPool2D");
var pad = X.AdjustPadToPool(l.pool, l.stride, l.pad);
X = m_Ops.AvgPool2D(X, l.pool, l.stride, pad);
}
else if (l.type == Layer.Type.GlobalMaxPool2D)
{
Profiler.BeginSample("Barracuda.GlobalMaxPool2D");
X = m_Ops.GlobalMaxPool2D(X);
}
else if (l.type == Layer.Type.GlobalAvgPool2D)
{
Profiler.BeginSample("Barracuda.GlobalAvgPool2D");
X = m_Ops.GlobalAvgPool2D(X);
}
else if (l.type == Layer.Type.Border2D)
{
Profiler.BeginSample("Barracuda.Border2D");
Assert.IsNotNull(l.pad);
// NOTE: beta is used to retrieve fillin value
// because beta is 0 by default (while alpha is 1 by default)
// 0 value is more inline with zero padding
float fillValue = l.beta;
X = m_Ops.Border2D(X, l.pad, fillValue);
}
else if (l.type == Layer.Type.Pad2DReflect)
{
Profiler.BeginSample("Barracuda.Pad2DReflect");
Assert.IsNotNull(l.pad);
X = m_Ops.Pad2DReflect(X, l.pad);
}
else if (l.type == Layer.Type.Pad2DSymmetric)
{
Profiler.BeginSample("Barracuda.Pad2DSymmetric");
Assert.IsNotNull(l.pad);
X = m_Ops.Pad2DSymmetric(X, l.pad);
}
else if (l.type == Layer.Type.Pad2DEdge)
{
Profiler.BeginSample("Barracuda.Pad2DEdge");
Assert.IsNotNull(l.pad);
X = m_Ops.Pad2DEdge(X, l.pad);
}
// 3D
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("3D operations are not implemented yet!");
}
else if (l.type == Layer.Type.ScaleBias)
{
Assert.AreEqual(inputs.Length, 3);
Profiler.BeginSample("Barracuda.ScaleBias");
X = m_Ops.ScaleBias(X, inputs[1], inputs[2]);
}
else if (l.type == Layer.Type.Normalization)
{
Assert.AreEqual(inputs.Length, 3);
Profiler.BeginSample("Barracuda.Normalization");
// @TODO: support other types of Normalization at test time.
// Currently supported only pool=1 (InstanceNormalization)
// NOTE: beta is used to retrieve epsilon value
// because beta is 0 by default (while alpha is 1 by default)
// 0 value is more inline with very small epsilon
var epsilon = l.beta;
if (epsilon == 0)
{
epsilon = Mathf.Epsilon; // safety check to prevent division by zero
}
X = m_Ops.Normalization(X, inputs[1], inputs[2], 1, l.axis, epsilon, GetAndVerifyFusedActivation(l));
}
else if (l.type == Layer.Type.LRN)
{
Profiler.BeginSample("Barracuda.LRN");
Assert.IsNotNull(l.pool);
Assert.AreEqual(l.pool.Length, 1);
int count = l.pool[0];
float bias = (l.weights.Length > 0) ? l.weights[0] : 1.0f;
X = m_Ops.LRN(X, l.alpha, l.beta, bias, count);
}
// Stochastic layers
else if (l.type == Layer.Type.Dropout)
{
Profiler.BeginSample("Barracuda.Dropout");
X = m_Ops.Dropout(X, l.alpha);
}
else if (l.type == Layer.Type.RandomNormal)
{
Profiler.BeginSample("Barracuda.RandomNormal");
Assert.IsNotNull(l.pool);
// pool size is treated as shape constant, if not empty
// otherwise shape of the previous tensor is used
var shape = X.shape;
if (l.pool.Length > 0)
{
shape = new TensorShape(l.pool);
}
int seed = (l.pad.Length > 0) ? l.pad[0] : 1337;
float scale = l.alpha, mean = l.beta;
X = m_Ops.RandomNormal(shape, mean, scale, seed);
}
else if (l.type == Layer.Type.RandomUniform)
{
Profiler.BeginSample("Barracuda.RandomUniform");
Assert.IsNotNull(l.pool);
// pool size is treated as shape constant, if not empty
// otherwise shape of the previous tensor is used
var shape = X.shape;
if (l.pool.Length > 0)
{
shape = new TensorShape(l.pool);
}
int seed = (l.pad.Length > 0) ? l.pad[0] : 1337;
float scale = l.alpha, mean = l.beta;
X = m_Ops.RandomUniform(shape, mean, scale, seed);
}
else if (l.type == Layer.Type.Multinomial)
{
Profiler.BeginSample("Barracuda.Multinomial");
Assert.IsNotNull(l.pool);
Assert.AreEqual(l.pool.Length, 1);
int count = l.pool[0];
int seed = (l.pad.Length > 0) ? l.pad[0] : 1337;
X = m_Ops.Multinomial(X, count, seed);
}
else if (l.type == Layer.Type.OneHot)
{
Profiler.BeginSample("Barracuda.OneHot");
Assert.IsNotNull(l.pool);
Assert.AreEqual(l.pool.Length, 1);
int depth = l.pool[0];
float on = l.alpha, off = l.beta;
X = m_Ops.OneHot(X, depth, on, off);
}
// Broadcast layers
else if (l.type == Layer.Type.Add)
{
Profiler.BeginSample("Barracuda.Add");
X = m_Ops.Add(inputs);
}
else if (l.type == Layer.Type.Sub)
{
Profiler.BeginSample("Barracuda.Sub");
X = m_Ops.Sub(inputs);
}
else if (l.type == Layer.Type.Mul)
{
Profiler.BeginSample("Barracuda.Mul");
X = m_Ops.Mul(inputs);
}
else if (l.type == Layer.Type.Div)
{
Profiler.BeginSample("Barracuda.Div");
X = m_Ops.Div(inputs);
}
else if (l.type == Layer.Type.Pow)
{
Profiler.BeginSample("Barracuda.Pow");
X = m_Ops.Pow(inputs);
}
else if (l.type == Layer.Type.Min)
{
Profiler.BeginSample("Barracuda.Min");
X = m_Ops.Min(inputs);
}
else if (l.type == Layer.Type.Max)
{
Profiler.BeginSample("Barracuda.Max");
X = m_Ops.Max(inputs);
}
else if (l.type == Layer.Type.Mean)
{
Profiler.BeginSample("Barracuda.Mean");
X = m_Ops.Mean(inputs);
}
// Reduction layers
else if (l.type == Layer.Type.ReduceMax)
{
Profiler.BeginSample("Barracuda.ReduceMax");
X = m_Ops.ReduceMax(X, l.axis);
}
else if (l.type == Layer.Type.ReduceMean)
{
Profiler.BeginSample("Barracuda.ReduceMean");
X = m_Ops.ReduceMean(X, l.axis);
}
else if (l.type == Layer.Type.ReduceMin)
{
Profiler.BeginSample("Barracuda.ReduceMin");
X = m_Ops.ReduceMin(X, l.axis);
}
else if (l.type == Layer.Type.ReduceProd)
{
Profiler.BeginSample("Barracuda.ReduceProd");
X = m_Ops.ReduceProd(X, l.axis);
}
else if (l.type == Layer.Type.ReduceSum)
{
Profiler.BeginSample("Barracuda.ReduceSum");
X = m_Ops.ReduceSum(X, l.axis);
}
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.ReduceSumSquare)
{
throw new NotImplementedException("This reduction operation is not implemented yet!");
}
// Logical operators with broadcast
else if (l.type == Layer.Type.Greater)
{
Assert.AreEqual(inputs.Length, 2);
Profiler.BeginSample("Barracuda.Greater");
X = m_Ops.Greater(X, inputs[1]);
}
else if (l.type == Layer.Type.GreaterEqual)
{
Assert.AreEqual(inputs.Length, 2);
Profiler.BeginSample("Barracuda.GreaterEqual");
X = m_Ops.GreaterEqual(X, inputs[1]);
}
else if (l.type == Layer.Type.Less)
{
Assert.AreEqual(inputs.Length, 2);
Profiler.BeginSample("Barracuda.Less");
X = m_Ops.Less(X, inputs[1]);
}
else if (l.type == Layer.Type.LessEqual)
{
Assert.AreEqual(inputs.Length, 2);
Profiler.BeginSample("Barracuda.LessEqual");
X = m_Ops.LessEqual(X, inputs[1]);
}
else if (l.type == Layer.Type.Equal)
{
Assert.AreEqual(inputs.Length, 2);
Profiler.BeginSample("Barracuda.Equal");
X = m_Ops.Equal(X, inputs[1]);
}
else if (l.type == Layer.Type.LogicalOr)
{
Assert.AreEqual(inputs.Length, 2);
Profiler.BeginSample("Barracuda.LogicalOr");
X = m_Ops.LogicalOr(X, inputs[1]);
}
else if (l.type == Layer.Type.LogicalAnd)
{
Assert.AreEqual(inputs.Length, 2);
Profiler.BeginSample("Barracuda.LogicalAnd");
X = m_Ops.LogicalAnd(X, inputs[1]);
}
else if (l.type == Layer.Type.LogicalXor)
{
Assert.AreEqual(inputs.Length, 2);
Profiler.BeginSample("Barracuda.LogicalXor");
X = m_Ops.LogicalXor(X, inputs[1]);
}
else if (l.type == Layer.Type.LogicalNot)
{
Profiler.BeginSample("Barracuda.LogicalNot");
X = m_Ops.LogicalNot(X);
}
// Shape affecting layers
else if (l.type == Layer.Type.Flatten)
{
Profiler.BeginSample("Barracuda.Flatten");
X = m_Ops.Flatten(X);
}
else if (l.type == Layer.Type.Reshape)
{
Profiler.BeginSample("Barracuda.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 && inputs.Length > 1)
{
size = inputs[1].shape.ToArray();
}
var newShape = X.shape.Reshape(size);
X = m_Ops.Reshape(X, newShape);
}
else if (l.type == Layer.Type.Expand)
{
Profiler.BeginSample("Barracuda.Expand");
// pool size is treated as new shape
var newShape = l.pool;
Assert.IsNotNull(newShape);
Assert.AreEqual(newShape.Length, 4);
X = m_Ops.Expand(X, new TensorShape(newShape));
}
else if (l.type == Layer.Type.Transpose)
{
Profiler.BeginSample("Barracuda.Transpose");
X = m_Ops.Transpose(X);
}
else if (l.type == Layer.Type.Gather)
{
Profiler.BeginSample("Barracuda.Gather");
X = m_Ops.Gather(inputs, l.axis);
}
else if (l.type == Layer.Type.Squeeze ||
l.type == Layer.Type.Unsqueeze)
{
throw new NotImplementedException();
}
else if (l.type == Layer.Type.Concat)
{
Profiler.BeginSample("Barracuda.Concat");
X = m_Ops.Concat(inputs, l.axis);
}
else if (l.type == Layer.Type.StridedSlice)
{
Profiler.BeginSample("Barracuda.StridedSlice");
Assert.IsNotNull(l.pad);
Assert.IsNotNull(l.pool);
Assert.IsNotNull(l.stride);
X = m_Ops.StridedSlice(X, l.pad, l.pool, l.stride);
}
else if (l.type == Layer.Type.Tile)
{
throw new NotImplementedException();
}
// Activations
else if (l.type == Layer.Type.Activation)
{
Profiler.BeginSample("Barracuda.Activation");
if (l.activation == Layer.Activation.Relu)
{
X = m_Ops.Relu(X);
}
else if (l.activation == Layer.Activation.Softmax)
{
X = m_Ops.Softmax(X);
}
else if (l.activation == Layer.Activation.LogSoftmax)
{
X = m_Ops.LogSoftmax(X);
}
else if (l.activation == Layer.Activation.Tanh)
{
X = m_Ops.Tanh(X);
}
else if (l.activation == Layer.Activation.Sigmoid)
{
X = m_Ops.Sigmoid(X);
}
else if (l.activation == Layer.Activation.Relu6)
{
X = m_Ops.Relu6(X);
}
else if (l.activation == Layer.Activation.Elu)
{
X = m_Ops.Elu(X, l.alpha);
}
else if (l.activation == Layer.Activation.LeakyRelu)
{
X = m_Ops.LeakyRelu(X, l.alpha);
}
else if (l.activation == Layer.Activation.Selu)
{
X = m_Ops.Selu(X, l.alpha, l.beta);
}
else if (l.activation == Layer.Activation.Swish)
{
X = m_Ops.Swish(X);
}
else if (l.activation == Layer.Activation.PRelu)
{
Assert.AreEqual(inputs.Length, 2);
X = m_Ops.PRelu(X, inputs[1]);
}
else if (
l.activation == Layer.Activation.Softplus ||
l.activation == Layer.Activation.Softsign ||
l.activation == Layer.Activation.Hardmax ||
l.activation == Layer.Activation.HardSigmoid)
{
throw new NotImplementedException("This activation function is not implemented yet!");
}
else if (l.activation == Layer.Activation.Abs)
{
X = m_Ops.Abs(X);
}
else if (l.activation == Layer.Activation.Neg)
{
X = m_Ops.Neg(X);
}
else if (l.activation == Layer.Activation.Ceil)
{
X = m_Ops.Ceil(X);
}
else if (l.activation == Layer.Activation.Clip)
{
X = m_Ops.Clip(X, l.alpha, l.beta);
}
else if (l.activation == Layer.Activation.Floor)
{
X = m_Ops.Floor(X);
}
else if (l.activation == Layer.Activation.Reciprocal)
{
X = m_Ops.Reciprocal(X);
}
else if (l.activation == Layer.Activation.Pow)
{
X = m_Ops.Pow(X, l.alpha);
}
else if (l.activation == Layer.Activation.Exp)
{
X = m_Ops.Exp(X);
}
else if (l.activation == Layer.Activation.Log)
{
X = m_Ops.Log(X);
}
else if (l.activation == Layer.Activation.Sqrt)
{
X = m_Ops.Sqrt(X);
}
else if ((int)l.activation >= (int)Layer.Activation.Acos &&
(int)l.activation <= (int)Layer.Activation.Tan)
{
throw new NotImplementedException("Trig functions are not implemented yet!");
}
else
{
X = m_Ops.Copy(X);
}
}
else
{
Profiler.BeginSample("Barracuda.Dummy");
Assert.AreEqual(l.activation, Layer.Activation.None);
}
m_Vars.Store(l, X);
m_SyncTensor = X;
// optype
Profiler.EndSample();
// layer.name
Profiler.EndSample();
yield return(null);
}
// request ResetAllocator before next Execute() starts
m_RequestResetAllocator = true;
Profiler.EndSample();
if (m_Verbose)
{
D.Log(m_Vars.GetAllocator());
}
}