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);
}
public IWeightTensor Decode(IWeightTensor input, AttentionPreProcessResult attenPreProcessResult, int batchSize, IComputeGraph g)
{
IWeightTensor V = input;
IWeightTensor lastStatus = m_decoders.LastOrDefault().Cell;
IWeightTensor context = m_attentionLayer.Perform(lastStatus, attenPreProcessResult, batchSize, g);
foreach (LSTMAttentionDecoderCell decoder in m_decoders)
{
IWeightTensor e = decoder.Step(context, V, g);
V = e;
}
IWeightTensor eOutput = g.Dropout(V, batchSize, m_dropoutRatio, false);
eOutput = m_decoderFFLayer.Process(eOutput, batchSize, g);
return(eOutput);
}
/// <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="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);
}
}
/// <summary>
/// Transformer encoder
/// </summary>
/// <param name="rawInputs"></param>
/// <param name="g"></param>
/// <returns></returns>
///
public IWeightTensor Decode(IWeightTensor tgtInputs, IWeightTensor encOutputBatchFirst, IWeightTensor tgtSelfMask, IWeightTensor srcTgtMask, int batchSize, IComputeGraph g)
{
using (IComputeGraph subg = g.CreateSubGraph($"{m_name}_Decoder"))
{
for (int k = 0; k < m_selfAttns.Count; k++)
{
tgtInputs = m_selfAttns[k].Perform(tgtInputs, tgtInputs, tgtInputs, tgtSelfMask, batchSize, subg);
tgtInputs = m_encAttns[k].Perform(tgtInputs, encOutputBatchFirst, encOutputBatchFirst, srcTgtMask, batchSize, subg);
tgtInputs = m_posFFNs[k].Perform(tgtInputs, batchSize, subg);
}
tgtInputs = layerNorm.Norm(tgtInputs, subg);
tgtInputs.UnbindFromComputeGraph();
}
tgtInputs = m_decoderFFLayer.Process(tgtInputs, batchSize, g);
return(tgtInputs);
}
/// <summary>
/// Decode output sentences in training
/// </summary>
/// <param name="outputSentences">In training mode, they are golden target sentences, otherwise, they are target sentences generated by the decoder</param>
/// <param name="g"></param>
/// <param name="encodedOutputs"></param>
/// <param name="decoder"></param>
/// <param name="decoderFFLayer"></param>
/// <param name="embedding"></param>
/// <returns></returns>
private float Decode(List <List <string> > outputSentences, IComputeGraph g, IWeightTensor encodedOutputs, AttentionDecoder decoder, FeedForwardLayer decoderFFLayer, IWeightTensor embedding,
int batchSize, bool isTraining = true)
{
float cost = 0.0f;
int[] ix_inputs = new int[batchSize];
for (int i = 0; i < ix_inputs.Length; i++)
{
ix_inputs[i] = (int)SENTTAGS.START;
}
// Initialize variables accoridng to current mode
var originalOutputLengths = isTraining ? ParallelCorpus.PadSentences(outputSentences) : null;
int seqLen = isTraining ? outputSentences[0].Count : 64;
var dropoutRatio = isTraining ? m_dropoutRatio : 0.0f;
HashSet <int> setEndSentId = isTraining ? null : new HashSet <int>();
if (!isTraining)
{
if (outputSentences.Count != 0)
{
throw new ArgumentException($"The list for output sentences must be empty if current is not in training mode.");
}
for (int i = 0; i < batchSize; i++)
{
outputSentences.Add(new List <string>());
}
}
// Pre-process for attention model
var attPreProcessResult = decoder.PreProcess(encodedOutputs, batchSize, g);
for (int i = 0; i < seqLen; i++)
{
//Get embedding for all sentence in the batch at position i
List <IWeightTensor> inputs = new List <IWeightTensor>();
for (int j = 0; j < batchSize; j++)
{
inputs.Add(g.PeekRow(embedding, ix_inputs[j]));
}
var inputsM = g.ConcatRows(inputs);
//Decode output sentence at position i
var eOutput = decoder.Decode(inputsM, attPreProcessResult, batchSize, g);
eOutput = g.Dropout(eOutput, batchSize, dropoutRatio, true);
eOutput = decoderFFLayer.Process(eOutput, batchSize, g);
//Softmax for output
using (var probs = g.Softmax(eOutput, runGradients: false, inPlace: true))
{
if (isTraining)
{
//Calculate loss for each word in the batch
for (int k = 0; k < batchSize; k++)
{
using (var probs_k = g.PeekRow(probs, k, runGradients: false))
{
var ix_targets_k = m_modelMetaData.Vocab.GetTargetWordIndex(outputSentences[k][i]);
var score_k = probs_k.GetWeightAt(ix_targets_k);
if (i < originalOutputLengths[k])
{
cost += (float)-Math.Log(score_k);
}
probs_k.SetWeightAt(score_k - 1, ix_targets_k);
ix_inputs[k] = ix_targets_k;
}
}
eOutput.CopyWeightsToGradients(probs);
}
else
{
// Output "i"th target word
var targetIdx = g.Argmax(probs, 1);
var targetWords = m_modelMetaData.Vocab.ConvertTargetIdsToString(targetIdx.ToList());
for (int j = 0; j < targetWords.Count; j++)
{
if (setEndSentId.Contains(j) == false)
{
outputSentences[j].Add(targetWords[j]);
if (targetWords[j] == ParallelCorpus.EOS)
{
setEndSentId.Add(j);
}
}
}
ix_inputs = targetIdx;
}
}
if (isTraining)
{
////Hacky: Run backward for last feed forward layer and dropout layer in order to save memory usage, since it's not time sequence dependency
g.RunTopBackward();
if (m_dropoutRatio > 0.0f)
{
g.RunTopBackward();
}
}
else
{
if (setEndSentId.Count == batchSize)
{
// All target sentences in current batch are finished, so we exit.
break;
}
}
}
return(cost);
}
/// <summary>
/// Decode output sentences in training
/// </summary>
/// <param name="outputSentences"></param>
/// <param name="g"></param>
/// <param name="encodedOutputs"></param>
/// <param name="decoder"></param>
/// <param name="Whd"></param>
/// <param name="bd"></param>
/// <param name="Embedding"></param>
/// <param name="predictSentence"></param>
/// <returns></returns>
private float Decode(List <List <string> > outputSentences, IComputeGraph g, IWeightMatrix encodedOutputs, AttentionDecoder decoder, FeedForwardLayer decoderFFLayer, IWeightMatrix Embedding, out List <List <string> > predictSentence)
{
predictSentence = null;
float cost = 0.0f;
var attPreProcessResult = decoder.PreProcess(encodedOutputs, g);
var originalOutputLengths = PadSentences(outputSentences);
int seqLen = outputSentences[0].Count;
int[] ix_inputs = new int[m_batchSize];
int[] ix_targets = new int[m_batchSize];
for (int i = 0; i < ix_inputs.Length; i++)
{
ix_inputs[i] = (int)SENTTAGS.START;
}
for (int i = 0; i < seqLen + 1; i++)
{
//Get embedding for all sentence in the batch at position i
List <IWeightMatrix> inputs = new List <IWeightMatrix>();
for (int j = 0; j < m_batchSize; j++)
{
List <string> OutputSentence = outputSentences[j];
ix_targets[j] = (int)SENTTAGS.UNK;
if (i >= seqLen)
{
ix_targets[j] = (int)SENTTAGS.END;
}
else
{
if (m_tgtWordToIndex.ContainsKey(OutputSentence[i]))
{
ix_targets[j] = m_tgtWordToIndex[OutputSentence[i]];
}
}
var x = g.PeekRow(Embedding, ix_inputs[j]);
inputs.Add(x);
}
var inputsM = g.ConcatRows(inputs);
//Decode output sentence at position i
var eOutput = decoder.Decode(inputsM, attPreProcessResult, g);
if (m_dropoutRatio > 0.0f)
{
eOutput = g.Dropout(eOutput, m_dropoutRatio);
}
var o = decoderFFLayer.Process(eOutput, g);
//Softmax for output
// var o = g.MulAdd(eOutput, Whd, bds);
var probs = g.Softmax(o, false);
o.ReleaseWeight();
//Calculate loss for each word in the batch
List <IWeightMatrix> probs_g = g.UnFolderRow(probs, m_batchSize, false);
for (int k = 0; k < m_batchSize; k++)
{
var probs_k = probs_g[k];
var score_k = probs_k.GetWeightAt(ix_targets[k]);
if (i < originalOutputLengths[k] + 1)
{
cost += (float)-Math.Log(score_k);
}
probs_k.SetWeightAt(score_k - 1, ix_targets[k]);
ix_inputs[k] = ix_targets[k];
probs_k.Dispose();
}
o.SetGradientByWeight(probs);
//Hacky: Run backward for last feed forward layer and dropout layer in order to save memory usage, since it's not time sequence dependency
g.RunTopBackward();
g.RunTopBackward();
if (m_dropoutRatio > 0.0f)
{
g.RunTopBackward();
}
}
return(cost);
}
