public IWeightMatrix Step(IWeightMatrix input, IComputeGraph innerGraph)
{
var hidden_prev = ht;
var cell_prev = ct;
var inputs = innerGraph.ConcatColumns(input, hidden_prev);
var bs = innerGraph.RepeatRows(b, input.Rows);
var hhSum = innerGraph.MulAdd(inputs, Wxh, bs);
(var gates_raw, var cell_write_raw) = innerGraph.SplitColumns(hhSum, hdim * 3, hdim);
var gates = innerGraph.Sigmoid(gates_raw);
var cell_write = innerGraph.Tanh(cell_write_raw);
(var input_gate, var forget_gate, var output_gate) = innerGraph.SplitColumns(gates, hdim, hdim, hdim);
// compute new cell activation
var retain_cell = innerGraph.EltMul(forget_gate, cell_prev); // what do we keep from cell
var write_cell = innerGraph.EltMul(input_gate, cell_write); // what do we write to cell
ct = innerGraph.Add(retain_cell, write_cell); // new cell contents
// compute hidden state as gated, saturated cell activations
ht = innerGraph.EltMul(output_gate, innerGraph.Tanh(ct));
return(ht);
}
/// <summary>
/// Scaled multi-heads attention component with skip connectioned feed forward layers
/// </summary>
/// <param name="input">The input tensor</param>
/// <param name="g">The instance of computing graph</param>
/// <returns></returns>
public IWeightTensor Perform(IWeightTensor input, IComputeGraph graph)
{
IComputeGraph g = graph.CreateSubGraph(m_name);
var seqLen = input.Rows / m_batchSize;
//Input projections
var allQ = g.View(Q.Process(input, g), m_batchSize, seqLen, m_multiHeadNum, m_d);
var allK = g.View(K.Process(input, g), m_batchSize, seqLen, m_multiHeadNum, m_d);
var allV = g.View(V.Process(input, g), m_batchSize, seqLen, m_multiHeadNum, m_d);
//Multi-head attentions
var Qs = g.View(g.Permute(allQ, 2, 0, 1, 3), m_multiHeadNum * m_batchSize, seqLen, m_d);
var Ks = g.View(g.Permute(allK, 2, 0, 3, 1), m_multiHeadNum * m_batchSize, m_d, seqLen);
var Vs = g.View(g.Permute(allV, 2, 0, 1, 3), m_multiHeadNum * m_batchSize, seqLen, m_d);
// Scaled softmax
float scale = 1.0f / (float)Math.Sqrt(m_d);
var attn = g.MulBatch(Qs, Ks, m_multiHeadNum * m_batchSize, scale);
var attn2 = g.View(attn, m_multiHeadNum * m_batchSize * seqLen, seqLen);
var softmax = g.Softmax(attn2);
var softmax2 = g.View(softmax, m_multiHeadNum * m_batchSize, seqLen, seqLen);
var o = g.View(g.MulBatch(softmax2, Vs, m_multiHeadNum * m_batchSize), m_multiHeadNum, m_batchSize, seqLen, m_d);
var W = g.View(g.Permute(o, 1, 2, 0, 3), m_batchSize * seqLen, m_multiHeadNum * m_d);
// Output projection
var finalAttResults = g.Affine(W, W0, b0);
//Skip connection and layer normaliztion
var addedAttResult = g.Add(finalAttResults, input);
var normAddedAttResult = layerNorm1.Process(addedAttResult, g);
//Feed forward
var ffnResult = feedForwardLayer1.Process(normAddedAttResult, g);
var reluFFNResult = g.Relu(ffnResult);
var ffn2Result = feedForwardLayer2.Process(reluFFNResult, g);
//Skip connection and layer normaliztion
var addFFNResult = g.Add(ffn2Result, normAddedAttResult);
var normAddFFNResult = layerNorm2.Process(addFFNResult, g);
return(normAddFFNResult);
}
/// <summary>
/// Scaled multi-heads attention component with skip connectioned feed forward layers
/// </summary>
/// <param name="inputQ">The input Q tensor</param>
/// <param name="keyMask">The mask for softmax</param>
/// <param name="batchSize">Batch size of input data set</param>
/// <param name="graph">The instance of computing graph</param>
/// <returns>Transformered output tensor</returns>
public (IWeightTensor, IWeightTensor) Perform(IWeightTensor inputQ, IWeightTensor keyMask, int batchSize, IComputeGraph graph, bool outputAttenWeights = false)
{
using IComputeGraph g = graph.CreateSubGraph($"{m_name}_MultiHeadAttention");
int seqLenQ = inputQ.Rows / batchSize;
IWeightTensor inputQNorm = layerNormQ.Norm(inputQ, g);
//Input projections
var weightedQKV = g.View(g.Affine(inputQNorm, QKV, QKVb), dims: new long[] { batchSize, seqLenQ, 3, m_multiHeadNum, m_d });
var allQ = g.Select(weightedQKV, 2, 0);
var allK = g.Select(weightedQKV, 2, 1);
var allV = g.Select(weightedQKV, 2, 2);
//Multi-head attentions
IWeightTensor Qs = g.View(g.AsContiguous(g.Transpose(allQ, 1, 2)), dims: new long[] { batchSize *m_multiHeadNum, seqLenQ, m_d });
IWeightTensor Ks = g.View(g.AsContiguous(g.Transpose(g.Transpose(allK, 1, 2), 2, 3)), dims: new long[] { batchSize *m_multiHeadNum, m_d, seqLenQ });
IWeightTensor Vs = g.View(g.AsContiguous(g.Transpose(allV, 1, 2)), dims: new long[] { batchSize *m_multiHeadNum, seqLenQ, m_d });
// Scaled softmax
float scale = 1.0f / (float)(Math.Sqrt(m_d));
var attn = g.MulBatch(Qs, Ks, scale);
attn = g.View(attn, dims: new long[] { batchSize, m_multiHeadNum, seqLenQ, seqLenQ });
if (keyMask != null)
{
attn = g.Add(attn, keyMask, inPlace: true);
}
var attnProbs = g.Softmax(attn, inPlace: true);
IWeightTensor sumAttnWeights = null;
if (outputAttenWeights)
{
//Merge all attention probs over multi-heads
sumAttnWeights = graph.Sum(attnProbs, 1);
sumAttnWeights = graph.Div(sumAttnWeights, (float)m_multiHeadNum);
sumAttnWeights = graph.View(sumAttnWeights, new long[] { batchSize *seqLenQ, seqLenQ });
}
attnProbs = g.View(attnProbs, dims: new long[] { batchSize *m_multiHeadNum, seqLenQ, seqLenQ });
IWeightTensor o = g.View(g.MulBatch(attnProbs, Vs), dims: new long[] { batchSize, m_multiHeadNum, seqLenQ, m_d });
IWeightTensor W = g.View(g.AsContiguous(g.Transpose(o, 1, 2)), dims: new long[] { batchSize *seqLenQ, m_multiHeadNum *m_d });
// Output projection
IWeightTensor finalAttResults = g.Dropout(g.Affine(W, W0, b0), batchSize, m_dropoutRatio, inPlace: true);
IWeightTensor result = graph.Add(finalAttResults, inputQ, inPlace: true);
return(result, sumAttnWeights);
}
public IWeightTensor Perform(IWeightTensor inputQ, IWeightTensor keyMask, int batchSize, IComputeGraph graph)
{
if (m_sharedQKV == false)
{
throw new ArgumentException($"Layer '{m_name}' is not in shared QKV mode, please call another Perform function with three separated input tensors.");
}
using (IComputeGraph g = graph.CreateSubGraph($"{m_name}_MultiHeadAttention_SharedQKV"))
{
int seqLenQ = inputQ.Rows / batchSize;
IWeightTensor inputQNorm = layerNormQ.Norm(inputQ, g);
//Input projections
float scale = 1.0f / (float)(m_inputDim);
IWeightTensor mulQ, mulK, mulV;
using (IWeightTensor inputQNormView = g.View(inputQNorm, dims: new long[] { 1, inputQ.Rows, inputQ.Columns }))
{
using (IWeightTensor inputQNormViewExp = g.Expand(inputQNormView, dims: new long[] { 3, inputQ.Rows, inputQ.Columns }))
{
using (IWeightTensor mulQKV = g.MulBatch(inputQNormViewExp, QKV, 3, scale))
{
mulQ = g.Select(mulQKV, 0, 0);
mulK = g.Select(mulQKV, 0, 1);
mulV = g.Select(mulQKV, 0, 2);
}
}
}
IWeightTensor allQ = g.View(mulQ, dims: new long[] { batchSize, seqLenQ, m_multiHeadNum, m_d });
IWeightTensor allK = g.View(mulK, dims: new long[] { batchSize, seqLenQ, m_multiHeadNum, m_d });
IWeightTensor allV = g.View(mulV, dims: new long[] { batchSize, seqLenQ, m_multiHeadNum, m_d });
//Multi-head attentions
IWeightTensor Qs = g.View(g.Permute(allQ, 2, 0, 1, 3), dims: new long[] { m_multiHeadNum *batchSize, seqLenQ, m_d });
IWeightTensor Ks = g.View(g.Permute(allK, 2, 0, 3, 1), dims: new long[] { m_multiHeadNum *batchSize, m_d, seqLenQ });
IWeightTensor Vs = g.View(g.Permute(allV, 2, 0, 1, 3), dims: new long[] { m_multiHeadNum *batchSize, seqLenQ, m_d });
// Scaled softmax
scale = 1.0f / (float)(m_d);
IWeightTensor attn = g.MulBatch(Qs, Ks, m_multiHeadNum * batchSize, scale);
IWeightTensor softmax = g.Softmax(attn, keyMask, inPlace: true);
IWeightTensor o = g.View(g.MulBatch(softmax, Vs, m_multiHeadNum * batchSize), dims: new long[] { m_multiHeadNum, batchSize, seqLenQ, m_d });
IWeightTensor W = g.View(g.Permute(o, 1, 2, 0, 3), dims: new long[] { batchSize *seqLenQ, m_multiHeadNum *m_d });
// Output projection
IWeightTensor finalAttResults = g.Dropout(g.Affine(W, W0, b0), batchSize, m_dropoutRatio, inPlace: true);
return(graph.Add(finalAttResults, inputQ));
}
}
public IWeightTensor Perform(IWeightTensor input, int batchSize, IComputeGraph graph)
{
using IComputeGraph g = graph.CreateSubGraph($"{m_name}_PositionwiseFeedForward");
var inputNorm = layerNorm2.Norm(input, g);
//Feed forward
IWeightTensor ffnResult = feedForwardLayer1.Process(inputNorm, batchSize, g);
IWeightTensor reluFFNResult = g.Relu(ffnResult, inPlace: true);
IWeightTensor ffn2Result = feedForwardLayer2.Process(reluFFNResult, batchSize, g);
//Skip connection and layer normaliztion
IWeightTensor addFFNResult = graph.Add(ffn2Result, input, inPlace: true);
return(addFFNResult);
}
/// <summary>
/// Scaled multi-heads attention component with skip connectioned feed forward layers
/// </summary>
/// <param name="inputQ">The input Q tensor</param>
/// <param name="inputK">The input K tensor</param>
/// <param name="inputV">The input V tensor</param>
/// <param name="batchSize">Batch size of input data set</param>
/// <param name="graph">The instance of computing graph</param>
/// <returns>Transformered output tensor</returns>
public IWeightTensor Perform(IWeightTensor inputQ, IWeightTensor inputK, IWeightTensor inputV, IWeightTensor keyMask, int batchSize, IComputeGraph graph)
{
using (IComputeGraph g = graph.CreateSubGraph($"{m_name}_MultiHeadAttention"))
{
int seqLenQ = inputQ.Rows / batchSize;
// SeqLenK must be euqal to SeqLenV
int seqLenK = inputK.Rows / batchSize;
int seqLenV = inputV.Rows / batchSize;
IWeightTensor inputQNorm = layerNorm1.Norm(inputQ, g);
IWeightTensor inputKNorm = (inputK == inputQ) ? inputQNorm : inputK; // layerNorm1.Norm(inputK, g);
IWeightTensor inputVNorm = (inputK == inputV) ? inputKNorm : inputV; // layerNorm1.Norm(inputV, g);
//Input projections
IWeightTensor allQ = g.View(g.Affine(inputQNorm, Q, Qb), dims: new long[] { batchSize, seqLenQ, m_multiHeadNum, m_d });
IWeightTensor allK = g.View(g.Affine(inputKNorm, K, Kb), dims: new long[] { batchSize, seqLenK, m_multiHeadNum, m_d });
IWeightTensor allV = g.View(g.Affine(inputVNorm, V, Vb), dims: new long[] { batchSize, seqLenV, m_multiHeadNum, m_d });
//Multi-head attentions
IWeightTensor Qs = g.View(g.Permute(allQ, 2, 0, 1, 3), dims: new long[] { m_multiHeadNum *batchSize, seqLenQ, m_d });
IWeightTensor Ks = g.View(g.Permute(allK, 2, 0, 3, 1), dims: new long[] { m_multiHeadNum *batchSize, m_d, seqLenK });
IWeightTensor Vs = g.View(g.Permute(allV, 2, 0, 1, 3), dims: new long[] { m_multiHeadNum *batchSize, seqLenV, m_d });
// Scaled softmax
float scale = 1.0f / (float)Math.Sqrt(m_d);
IWeightTensor attn = g.MulBatch(Qs, Ks, m_multiHeadNum * batchSize, scale);
IWeightTensor attn2 = g.View(attn, dims: new long[] { m_multiHeadNum *batchSize *seqLenQ, seqLenK });
if (keyMask != null)
{
// attn2 = g.Add(attn2, mask, runGradient2: false);
attn2 = g.MaskFill(attn2, keyMask, -1e9f);
}
IWeightTensor softmax = g.Softmax(attn2, inPlace: true);
IWeightTensor softmax2 = g.View(softmax, dims: new long[] { m_multiHeadNum *batchSize, seqLenQ, seqLenK });
IWeightTensor o = g.View(g.MulBatch(softmax2, Vs, m_multiHeadNum * batchSize), dims: new long[] { m_multiHeadNum, batchSize, seqLenQ, m_d });
IWeightTensor W = g.View(g.Permute(o, 1, 2, 0, 3), dims: new long[] { batchSize *seqLenQ, m_multiHeadNum *m_d });
// Output projection
IWeightTensor finalAttResults = g.Dropout(g.Affine(W, W0, b0), batchSize, m_dropoutRatio, inPlace: true);
return(graph.Add(finalAttResults, inputQ));
}
}
/// <summary>
/// Scaled multi-heads attention component with skip connectioned feed forward layers
/// </summary>
/// <param name="input">The input tensor</param>
/// <param name="g">The instance of computing graph</param>
/// <returns></returns>
public IWeightTensor Perform(IWeightTensor input, int batchSize, IComputeGraph graph)
{
using (IComputeGraph g = graph.CreateSubGraph(m_name))
{
int seqLen = input.Rows / batchSize;
IWeightTensor nInput = layerNorm1.Norm(input, g);
//Input projections
IWeightTensor allQ = g.View(g.Affine(nInput, Q, Qb), batchSize, seqLen, m_multiHeadNum, m_d);
IWeightTensor allK = g.View(g.Affine(nInput, K, Kb), batchSize, seqLen, m_multiHeadNum, m_d);
IWeightTensor allV = g.View(g.Affine(nInput, V, Vb), batchSize, seqLen, m_multiHeadNum, m_d);
//Multi-head attentions
IWeightTensor Qs = g.View(g.Permute(allQ, 2, 0, 1, 3), m_multiHeadNum * batchSize, seqLen, m_d);
IWeightTensor Ks = g.View(g.Permute(allK, 2, 0, 3, 1), m_multiHeadNum * batchSize, m_d, seqLen);
IWeightTensor Vs = g.View(g.Permute(allV, 2, 0, 1, 3), m_multiHeadNum * batchSize, seqLen, m_d);
// Scaled softmax
float scale = 1.0f / (float)Math.Sqrt(m_d);
IWeightTensor attn = g.MulBatch(Qs, Ks, m_multiHeadNum * batchSize, scale);
IWeightTensor attn2 = g.View(attn, m_multiHeadNum * batchSize * seqLen, seqLen);
IWeightTensor softmax = g.Softmax(attn2, inPlace: true);
IWeightTensor softmax2 = g.View(softmax, m_multiHeadNum * batchSize, seqLen, seqLen);
IWeightTensor o = g.View(g.MulBatch(softmax2, Vs, m_multiHeadNum * batchSize), m_multiHeadNum, batchSize, seqLen, m_d);
IWeightTensor W = g.View(g.Permute(o, 1, 2, 0, 3), batchSize * seqLen, m_multiHeadNum * m_d);
// Output projection
IWeightTensor finalAttResults = g.Dropout(g.Affine(W, W0, b0), batchSize, m_dropoutRatio, inPlace: true);
//Skip connection and layer normaliztion
IWeightTensor normAddedAttResult = layerNorm2.AddNorm(finalAttResults, input, g);
//Feed forward
IWeightTensor ffnResult = feedForwardLayer1.Process(normAddedAttResult, batchSize, g);
IWeightTensor reluFFNResult = g.Relu(ffnResult);
IWeightTensor ffn2Result = feedForwardLayer2.Process(reluFFNResult, batchSize, g);
//Skip connection and layer normaliztion
IWeightTensor addFFNResult = graph.Add(ffn2Result, normAddedAttResult);
return(addFFNResult);
}
}
private IWeightTensor AddPositionEmbedding(IComputeGraph g, IWeightTensor posEmbedding, int batchSize, int seqLen, IWeightTensor inputEmbs)
{
using (var posEmbeddingPeek = g.PeekRow(posEmbedding, 0, seqLen, false))
{
using (var posEmbeddingPeekView = g.View(posEmbeddingPeek, false, new long[] { 1, seqLen, this.m_modelMetaData.EmbeddingDim }))
{
using (var posEmbeddingPeekViewExp = g.Expand(posEmbeddingPeekView, false, new long[] { batchSize, seqLen, this.m_modelMetaData.EmbeddingDim }))
{
inputEmbs = g.View(inputEmbs, dims: new long[] { batchSize, seqLen, this.m_modelMetaData.EmbeddingDim });
inputEmbs = g.Add(inputEmbs, posEmbeddingPeekViewExp, true, false);
inputEmbs = g.View(inputEmbs, dims: new long[] { batchSize *seqLen, this.m_modelMetaData.EmbeddingDim });
}
}
}
inputEmbs = g.Dropout(inputEmbs, batchSize, this.m_dropoutRatio, true);
return(inputEmbs);
}
/// <summary>
/// Scaled multi-heads attention component with skip connectioned feed forward layers
/// </summary>
/// <param name="inputQ">The input Q tensor</param>
/// <param name="inputK">The input K tensor</param>
/// <param name="inputV">The input V tensor</param>
/// <param name="batchSize">Batch size of input data set</param>
/// <param name="graph">The instance of computing graph</param>
/// <returns>Transformered output tensor</returns>
public IWeightTensor Perform(IWeightTensor inputQ, IWeightTensor inputK, IWeightTensor inputV, IWeightTensor keyMask, int batchSize, IComputeGraph graph)
{
if (m_sharedQKV)
{
throw new ArgumentException($"Layer '{m_name}' is in shared QKV mode, please call antoher Perform function with single input tensor.");
}
using (IComputeGraph g = graph.CreateSubGraph($"{m_name}_MultiHeadAttention"))
{
int seqLenQ = inputQ.Rows / batchSize;
// SeqLenK must be euqal to SeqLenV
int seqLenK = inputK.Rows / batchSize;
int seqLenV = inputV.Rows / batchSize;
IWeightTensor inputQNorm = layerNormQ.Norm(inputQ, g);
//Input projections
float scale = 1.0f / (float)(m_inputDim);
IWeightTensor allQ = g.View(g.Affine(inputQNorm, Q, Qb, scale), dims: new long[] { batchSize, seqLenQ, m_multiHeadNum, m_d });
IWeightTensor allK = g.View(g.Affine(inputK, K, Kb, scale), dims: new long[] { batchSize, seqLenK, m_multiHeadNum, m_d });
IWeightTensor allV = g.View(g.Affine(inputV, V, Vb, scale), dims: new long[] { batchSize, seqLenV, m_multiHeadNum, m_d });
//Multi-head attentions
IWeightTensor Qs = g.View(g.Permute(allQ, 2, 0, 1, 3), dims: new long[] { m_multiHeadNum *batchSize, seqLenQ, m_d });
IWeightTensor Ks = g.View(g.Permute(allK, 2, 0, 3, 1), dims: new long[] { m_multiHeadNum *batchSize, m_d, seqLenK });
IWeightTensor Vs = g.View(g.Permute(allV, 2, 0, 1, 3), dims: new long[] { m_multiHeadNum *batchSize, seqLenV, m_d });
// Scaled softmax
scale = 1.0f / (float)(m_d);
IWeightTensor attn = g.MulBatch(Qs, Ks, m_multiHeadNum * batchSize, scale);
IWeightTensor softmax = g.Softmax(attn, keyMask, inPlace: true);
IWeightTensor o = g.View(g.MulBatch(softmax, Vs, m_multiHeadNum * batchSize), dims: new long[] { m_multiHeadNum, batchSize, seqLenQ, m_d });
IWeightTensor W = g.View(g.Permute(o, 1, 2, 0, 3), dims: new long[] { batchSize *seqLenQ, m_multiHeadNum *m_d });
// Output projection
IWeightTensor finalAttResults = g.Dropout(g.Affine(W, W0, b0), batchSize, m_dropoutRatio, inPlace: true);
return(graph.Add(finalAttResults, inputQ));
}
}
/// <summary>
/// Transformer encoder
/// </summary>
/// <param name="rawInputs"></param>
/// <param name="g"></param>
/// <returns></returns>
public IWeightTensor Encode(IWeightTensor rawInput, IComputeGraph g)
{
int seqLen = rawInput.Rows / m_batchSize;
var posEmbedding = g.BuildPositionMatrix(seqLen, m_inputDim);
var posEmbeddingRepeat = g.RepeatRows(posEmbedding, m_batchSize);
// Transpose to batch-first based sequence
var inputs = g.TransposeBatch(rawInput, m_batchSize);
inputs = g.Mul(inputs, (float)Math.Sqrt(m_inputDim));
inputs = g.Add(inputs, posEmbeddingRepeat);
for (int k = 0; k < m_encoders.Count; k++)
{
inputs = m_encoders[k].Perform(inputs, g);
}
// Transpose back to time-first based sequence
rawInput = g.TransposeBatch(inputs, seqLen);
return(rawInput);
}
public static IWeightTensor AddPositionEmbedding(IComputeGraph g, IWeightTensor posEmbedding, int batchSize, IWeightTensor inputEmbs, float dropoutRatio)
{
var Column = posEmbedding.Columns;
int seqLen = inputEmbs.Rows / batchSize;
using (var posEmbeddingPeek = g.Peek(posEmbedding, 0, 0, seqLen))
{
using (var posEmbeddingPeekView = g.View(posEmbeddingPeek, dims: new long[] { 1, seqLen, Column }))
{
using (var posEmbeddingPeekViewExp = g.Expand(posEmbeddingPeekView, dims: new long[] { batchSize, seqLen, Column }))
{
inputEmbs = g.View(inputEmbs, dims: new long[] { batchSize, seqLen, Column });
inputEmbs = g.Add(inputEmbs, posEmbeddingPeekViewExp, inPlace: true);
inputEmbs = g.View(inputEmbs, dims: new long[] { batchSize *seqLen, Column });
}
}
}
inputEmbs = g.Dropout(inputEmbs, batchSize, dropoutRatio, inPlace: true);
return(inputEmbs);
}
private IWeightTensor AddPositionEmbedding(IComputeGraph g, IWeightTensor posEmbedding, int batchSize, int seqLen, IWeightTensor inputEmbs)
{
var Column = posEmbedding.Columns;
inputEmbs = g.Mul(inputEmbs, (float)Math.Sqrt(m_modelMetaData.HiddenDim));
using (var posEmbeddingPeek = g.PeekRow(posEmbedding, 0, seqLen, false))
{
using (var posEmbeddingPeekView = g.View(posEmbeddingPeek, runGradient: false, dims: new long[] { 1, seqLen, Column }))
{
using (var posEmbeddingPeekViewExp = g.Expand(posEmbeddingPeekView, runGradient: false, dims: new long[] { batchSize, seqLen, Column }))
{
inputEmbs = g.View(inputEmbs, dims: new long[] { batchSize, seqLen, Column });
inputEmbs = g.Add(inputEmbs, posEmbeddingPeekViewExp, true, false);
inputEmbs = g.View(inputEmbs, dims: new long[] { batchSize *seqLen, Column });
}
}
}
inputEmbs = g.Dropout(inputEmbs, batchSize, m_dropoutRatio, inPlace: true);
return(inputEmbs);
}
/// <summary>
/// Update LSTM-Attention cells according to given weights
/// </summary>
/// <param name="context">The context weights for attention</param>
/// <param name="input">The input weights</param>
/// <param name="computeGraph">The compute graph to build workflow</param>
/// <returns>Update hidden weights</returns>
public IWeightMatrix Step(IWeightMatrix context, IWeightMatrix input, IComputeGraph computeGraph)
{
var cell_prev = ct;
var hidden_prev = ht;
var hxhc = computeGraph.ConcatColumns(input, hidden_prev, context);
var bs = computeGraph.RepeatRows(b, input.Rows);
var hhSum = computeGraph.MulAdd(hxhc, Wxhc, bs);
(var gates_raw, var cell_write_raw) = computeGraph.SplitColumns(hhSum, hdim * 3, hdim);
var gates = computeGraph.Sigmoid(gates_raw);
var cell_write = computeGraph.Tanh(cell_write_raw);
(var input_gate, var forget_gate, var output_gate) = computeGraph.SplitColumns(gates, hdim, hdim, hdim);
// compute new cell activation
var retain_cell = computeGraph.EltMul(forget_gate, cell_prev);
var write_cell = computeGraph.EltMul(input_gate, cell_write);
ct = computeGraph.Add(retain_cell, write_cell);
ht = computeGraph.EltMul(output_gate, computeGraph.Tanh(ct));
return(ht);
}
/// <summary>
/// Scaled multi-heads attention component with skip connectioned feed forward layers
/// </summary>
/// <param name="inputQ">The input Q tensor</param>
/// <param name="inputK">The input K tensor</param>
/// <param name="inputV">The input V tensor</param>
/// <param name="keyMask">The mask for softmax</param>
/// <param name="batchSize">Batch size of input data set</param>
/// <param name="graph">The instance of computing graph</param>
/// <returns>Transformered output tensor</returns>
public (IWeightTensor, IWeightTensor) Perform(IWeightTensor inputQ, IWeightTensor inputK, IWeightTensor inputV, IWeightTensor keyMask, int batchSize, IComputeGraph graph, bool outputAttenWeights = false, Dictionary <string, IWeightTensor> cachedTensors = null)
{
string keyName = $"{m_name}_MultiHeadAttention";
using IComputeGraph g = graph.CreateSubGraph(keyName);
int seqLenQ = inputQ.Rows / batchSize;
// SeqLenK must be euqal to SeqLenV
int seqLenK = inputK.Rows / batchSize;
int seqLenV = inputV.Rows / batchSize;
IWeightTensor inputQNorm = layerNormQ.Norm(inputQ, g);
//Input projections
IWeightTensor allQ = g.View(g.Affine(inputQNorm, Q, Qb), dims: new long[] { batchSize, seqLenQ, m_multiHeadNum, m_d });
//Multi-head attentions
IWeightTensor Qs = g.View(g.AsContiguous(g.Transpose(allQ, 1, 2)), dims: new long[] { batchSize *m_multiHeadNum, seqLenQ, m_d });
IWeightTensor Ks = null;
IWeightTensor Vs = null;
if (cachedTensors == null) // We don't use any cached tensors
{
IWeightTensor allK = g.View(g.Affine(inputK, K, Kb), dims: new long[] { batchSize, seqLenK, m_multiHeadNum, m_d });
IWeightTensor allV = g.View(g.Affine(inputV, V, Vb), dims: new long[] { batchSize, seqLenV, m_multiHeadNum, m_d });
Ks = g.View(g.AsContiguous(g.Transpose(g.Transpose(allK, 1, 2), 2, 3)), dims: new long[] { batchSize *m_multiHeadNum, m_d, seqLenK });
Vs = g.View(g.AsContiguous(g.Transpose(allV, 1, 2)), dims: new long[] { batchSize *m_multiHeadNum, seqLenV, m_d });
}
else
{
string KsCacheName = keyName + "_" + nameof(Ks);
string VsCacheName = keyName + "_" + nameof(Vs);
if (cachedTensors.ContainsKey(KsCacheName) == false)
{
IWeightTensor allK = g.View(g.Affine(inputK, K, Kb), dims: new long[] { batchSize, seqLenK, m_multiHeadNum, m_d });
Ks = g.View(g.AsContiguous(g.Transpose(g.Transpose(allK, 1, 2), 2, 3)), dims: new long[] { batchSize *m_multiHeadNum, m_d, seqLenK });
cachedTensors.Add(KsCacheName, Ks.CopyWeightsRef(KsCacheName, Ks.NeedGradient));
}
else
{
Ks = cachedTensors[KsCacheName];
}
if (cachedTensors.ContainsKey(VsCacheName) == false)
{
IWeightTensor allV = g.View(g.Affine(inputV, V, Vb), dims: new long[] { batchSize, seqLenV, m_multiHeadNum, m_d });
Vs = g.View(g.AsContiguous(g.Transpose(allV, 1, 2)), dims: new long[] { batchSize *m_multiHeadNum, seqLenV, m_d });
cachedTensors.Add(VsCacheName, Vs.CopyWeightsRef(VsCacheName, Vs.NeedGradient));
}
else
{
Vs = cachedTensors[VsCacheName];
}
}
// Scaled softmax
float scale = 1.0f / (float)(Math.Sqrt(m_d));
var attn = g.MulBatch(Qs, Ks, scale);
attn = g.View(attn, dims: new long[] { batchSize, m_multiHeadNum, seqLenQ, seqLenK });
if (keyMask != null)
{
attn = g.Add(attn, keyMask, inPlace: true);
}
var attnProbs = g.Softmax(attn, inPlace: true);
IWeightTensor sumAttnWeights = null;
if (outputAttenWeights)
{
sumAttnWeights = g.Select(attnProbs, 1, 0);
for (int i = 1; i < m_multiHeadNum; i++)
{
var tmp = g.Select(attnProbs, 1, i);
sumAttnWeights = g.Add(sumAttnWeights, tmp);
}
sumAttnWeights = graph.Div(sumAttnWeights, (float)m_multiHeadNum);
sumAttnWeights = graph.View(sumAttnWeights, new long[] { batchSize *seqLenQ, seqLenK });
}
attnProbs = g.View(attnProbs, dims: new long[] { batchSize *m_multiHeadNum, seqLenQ, seqLenK });
IWeightTensor o = g.View(g.MulBatch(attnProbs, Vs), dims: new long[] { batchSize, m_multiHeadNum, seqLenQ, m_d });
IWeightTensor W = g.View(g.AsContiguous(g.Transpose(o, 1, 2)), dims: new long[] { batchSize *seqLenQ, m_multiHeadNum *m_d });
// Output projection
IWeightTensor finalAttResults = g.Dropout(g.Affine(W, W0, b0), batchSize, m_dropoutRatio, inPlace: true);
IWeightTensor result = graph.Add(finalAttResults, inputQ, inPlace: true);
return(result, sumAttnWeights);
}
/// <summary>
/// Scaled multi-heads attention component with skip connectioned feed forward layers
/// </summary>
/// <param name="inputQ">The input Q tensor</param>
/// <param name="inputK">The input K tensor</param>
/// <param name="inputV">The input V tensor</param>
/// <param name="keyMask">The mask for softmax</param>
/// <param name="batchSize">Batch size of input data set</param>
/// <param name="graph">The instance of computing graph</param>
/// <returns>Transformered output tensor</returns>
public IWeightTensor Perform(IWeightTensor inputQ, IWeightTensor inputK, IWeightTensor inputV, IWeightTensor keyMask, int batchSize, IComputeGraph graph)
{
using (IComputeGraph g = graph.CreateSubGraph($"{m_name}_MultiHeadAttention"))
{
int seqLenQ = inputQ.Rows / batchSize;
// SeqLenK must be euqal to SeqLenV
int seqLenK = inputK.Rows / batchSize;
int seqLenV = inputV.Rows / batchSize;
IWeightTensor inputQNorm = layerNormQ.Norm(inputQ, g);
if (inputK == inputQ)
{
inputK = inputQNorm;
}
if (inputV == inputQ)
{
inputV = inputQNorm;
}
//Input projections
float scale = 1.0f;
IWeightTensor allQ = g.View(g.Affine(inputQNorm, Q, Qb, scale), dims: new long[] { batchSize, seqLenQ, m_multiHeadNum, m_d });
IWeightTensor allK = g.View(g.Affine(inputK, K, Kb, scale), dims: new long[] { batchSize, seqLenK, m_multiHeadNum, m_d });
IWeightTensor allV = g.View(g.Affine(inputV, V, Vb, scale), dims: new long[] { batchSize, seqLenV, m_multiHeadNum, m_d });
//Multi-head attentions
IWeightTensor Qs = g.View(g.AsContiguous(g.Transpose(allQ, 1, 2)), dims: new long[] { batchSize *m_multiHeadNum, seqLenQ, m_d });
IWeightTensor Ks = g.View(g.AsContiguous(g.Transpose(g.Transpose(allK, 1, 2), 2, 3)), dims: new long[] { batchSize *m_multiHeadNum, m_d, seqLenK });
IWeightTensor Vs = g.View(g.AsContiguous(g.Transpose(allV, 1, 2)), dims: new long[] { batchSize *m_multiHeadNum, seqLenV, m_d });
// Scaled softmax
scale = 1.0f / (float)(Math.Sqrt(m_d));
IWeightTensor attn = g.MulBatch(Qs, Ks, batchSize * m_multiHeadNum, scale);
if (keyMask != null)
{
using (var keyMaskView = g.View(keyMask, runGradient: false, dims: new long[] { batchSize, 1, seqLenQ, seqLenK }))
{
using (var keyMaskViewExp = g.Expand(keyMaskView, runGradient: false, dims: new long[] { batchSize, m_multiHeadNum, seqLenQ, seqLenK }))
{
using (var keyMaskViewExpConti = g.AsContiguous(keyMaskViewExp, runGradient: false))
{
using (var keyMaskViewExpContiView = g.View(keyMaskViewExpConti, runGradient: false, dims: new long[] { batchSize *m_multiHeadNum, seqLenQ, seqLenK }))
{
attn = g.Add(attn, keyMaskViewExpContiView, runGradient1: true, runGradient2: false);
}
}
}
}
}
IWeightTensor softmax = g.Softmax(attn, inPlace: true);
IWeightTensor o = g.View(g.MulBatch(softmax, Vs, batchSize * m_multiHeadNum), dims: new long[] { batchSize, m_multiHeadNum, seqLenQ, m_d });
IWeightTensor W = g.View(g.AsContiguous(g.Transpose(o, 1, 2)), dims: new long[] { batchSize *seqLenQ, m_multiHeadNum *m_d });
// Output projection
IWeightTensor finalAttResults = g.Dropout(g.Affine(W, W0, b0), batchSize, m_dropoutRatio, inPlace: true);
return(graph.Add(finalAttResults, inputQ));
}
}
/// <summary>
/// Create input embedding from token embeddings, segment embeddings
/// </summary>
/// <param name="seqs"></param>
/// <param name="g"></param>
/// <param name="embeddingsTensor"></param>
/// <param name="seqOriginalLengths"></param>
/// <param name="segmentEmbedding"></param>
/// <param name="vocab"></param>
/// <returns>The embedding tensor. shape: (batchsize * seqLen, embedding_dim) </returns>
public static IWeightTensor CreateTokensEmbeddings(List <List <int> > seqs, IComputeGraph g, IWeightTensor embeddingsTensor,
IWeightTensor segmentEmbedding, Vocab vocab, float scaleFactor = 1.0f, bool enableTagEmbedding = false)
{
int batchSize = seqs.Count;
int seqLen = seqs[0].Count;
float[] idxs = new float[batchSize * seqLen];
float[] segIdxs = new float[batchSize * seqLen];
List <float[]> tagIdxsList = new List <float[]>();
//float[] tagIdxs = new float[batchSize * seqLen];
for (int i = 0; i < batchSize; i++)
{
int segIdx = 0;
List <int> currTagIdxs = new List <int>();
int currTagLevel = 0;
for (int j = 0; j < seqLen; j++)
{
idxs[i * seqLen + j] = seqs[i][j];
segIdxs[i * seqLen + j] = segIdx;
string token = vocab.GetString(seqs[i][j]);
if (token == BuildInTokens.SEP)
{
//A new segment
segIdx++;
}
if (enableTagEmbedding)
{
if (token.StartsWith("<") && token.EndsWith(">") && BuildInTokens.IsPreDefinedToken(token) == false)
{
if (token[1] == '/')
{
currTagLevel--;
currTagIdxs[currTagLevel] = -1;
}
else
{
//A new opening tag
while (tagIdxsList.Count <= currTagLevel)
{
float[] tagIdxs = new float[batchSize * seqLen];
Array.Fill(tagIdxs, -1.0f);
tagIdxsList.Add(tagIdxs);
}
while (currTagIdxs.Count <= currTagLevel)
{
currTagIdxs.Add(-1);
}
currTagIdxs[currTagLevel] = seqs[i][j];
currTagLevel++;
}
}
else
{
for (int k = 0; k < currTagLevel; k++)
{
tagIdxsList[k][i * seqLen + j] = currTagIdxs[k];
//Logger.WriteLine($"Add tag embeddings: '{currTagIdxs[k]}'");
}
}
}
}
}
IWeightTensor tagEmbeddings = null;
if (enableTagEmbedding)
{
for (int k = 0; k < tagIdxsList.Count; k++)
{
var tagEmbeddings_k = g.IndexSelect(embeddingsTensor, tagIdxsList[k], clearWeights: true);
if (tagEmbeddings == null)
{
tagEmbeddings = tagEmbeddings_k;
}
else
{
tagEmbeddings = g.Add(tagEmbeddings, tagEmbeddings_k);
}
}
}
IWeightTensor embeddingRst = g.IndexSelect(embeddingsTensor, idxs);
if (scaleFactor != 1.0f)
{
embeddingRst = g.Mul(embeddingRst, scaleFactor, inPlace: true);
}
// Apply segment embeddings to the input sequence embeddings
if (segmentEmbedding != null)
{
embeddingRst = g.Add(embeddingRst, g.IndexSelect(segmentEmbedding, segIdxs));
}
if (tagEmbeddings != null)
{
embeddingRst = g.Add(embeddingRst, tagEmbeddings);
}
return(embeddingRst);
}