/// <summary>
/// Encode source sentences and output encoded weights
/// </summary>
/// <param name="g"></param>
/// <param name="inputSentences"></param>
/// <param name="encoder"></param>
/// <param name="reversEncoder"></param>
/// <param name="Embedding"></param>
/// <returns></returns>
private IWeightTensor Encode(IComputeGraph g, List <List <string> > inputSentences, IEncoder encoder, IWeightTensor Embedding)
{
ParallelCorpus.PadSentences(inputSentences);
int seqLen = inputSentences[0].Count;
int batchSize = inputSentences.Count;
List <IWeightTensor> forwardInput = new List <IWeightTensor>();
for (int i = 0; i < seqLen; i++)
{
for (int j = 0; j < inputSentences.Count; j++)
{
int ix_source = m_modelMetaData.Vocab.GetSourceWordIndex(inputSentences[j][i], logUnk: true);
forwardInput.Add(g.PeekRow(Embedding, ix_source));
}
}
return(encoder.Encode(g.ConcatRows(forwardInput), batchSize, g));
}
/// <summary>
/// Run forward part on given single device
/// </summary>
/// <param name="computeGraph">The computing graph for current device. It gets created and passed by the framework</param>
/// <param name="srcSnts">A batch of input tokenized sentences in source side</param>
/// <param name="tgtSnts">A batch of output tokenized sentences in target side</param>
/// <param name="deviceIdIdx">The index of current device</param>
/// <returns>The cost of forward part</returns>
private float RunForwardOnSingleDevice(IComputeGraph computeGraph, List <List <string> > srcSnts, List <List <string> > tgtSnts, int deviceIdIdx, bool isTraining)
{
var(encoder, decoder, srcEmbedding, tgtEmbedding, posEmbedding) = this.GetNetworksOnDeviceAt(deviceIdIdx);
// Reset networks
encoder.Reset(computeGraph.GetWeightFactory(), srcSnts.Count);
decoder.Reset(computeGraph.GetWeightFactory(), srcSnts.Count);
var originalSrcLengths = ParallelCorpus.PadSentences(srcSnts);
var srcSeqPaddedLen = srcSnts[0].Count;
var batchSize = srcSnts.Count;
var srcSelfMask = this.m_shuffleType == ShuffleEnums.NoPaddingInSrc ? null : MaskUtils.BuildPadSelfMask(computeGraph, srcSeqPaddedLen, originalSrcLengths, this.DeviceIds[deviceIdIdx]); // The length of source sentences are same in a single mini-batch, so we don't have source mask.
// Encoding input source sentences
var encOutput = this.Encode(computeGraph, srcSnts, encoder, srcEmbedding, srcSelfMask, posEmbedding, originalSrcLengths);
srcSelfMask?.Dispose();
// Generate output decoder sentences
if (decoder is AttentionDecoder)
{
return(this.DecodeAttentionLSTM(tgtSnts, computeGraph, encOutput, decoder as AttentionDecoder, tgtEmbedding, srcSnts.Count, isTraining));
}
else
{
if (isTraining)
{
return(this.DecodeTransformer(tgtSnts, computeGraph, encOutput, decoder as TransformerDecoder, tgtEmbedding, posEmbedding, batchSize, this.DeviceIds[deviceIdIdx], originalSrcLengths, isTraining));
}
else
{
for (var i = 0; i < this.m_maxTgtSntSize; i++)
{
using (var g = computeGraph.CreateSubGraph($"TransformerDecoder_Step_{i}"))
{
this.DecodeTransformer(tgtSnts, g, encOutput, decoder as TransformerDecoder, tgtEmbedding, posEmbedding, batchSize, this.DeviceIds[deviceIdIdx], originalSrcLengths, isTraining);
var allSntsEnd = true;
foreach (var t in tgtSnts)
{
if (t[^ 1] != ParallelCorpus.EOS)
/// <summary>
/// Run forward part on given single device
/// </summary>
/// <param name="g">The computing graph for current device. It gets created and passed by the framework</param>
/// <param name="srcSnts">A batch of input tokenized sentences in source side</param>
/// <param name="tgtSnts">A batch of output tokenized sentences in target side. In training mode, it inputs target tokens, otherwise, it outputs target tokens generated by decoder</param>
/// <param name="deviceIdIdx">The index of current device</param>
/// <returns>The cost of forward part</returns>
private float RunForwardOnSingleDevice(IComputeGraph g, List <List <string> > srcSnts, List <List <string> > tgtSnts, int deviceIdIdx, bool isTraining)
{
(IEncoder encoder, IWeightTensor srcEmbedding, FeedForwardLayer decoderFFLayer) = GetNetworksOnDeviceAt(deviceIdIdx);
int batchSize = srcSnts.Count;
// Reset networks
encoder.Reset(g.GetWeightFactory(), batchSize);
// Encoding input source sentences
ParallelCorpus.PadSentences(srcSnts);
if (isTraining)
{
ParallelCorpus.PadSentences(tgtSnts);
if (srcSnts[0].Count != tgtSnts[0].Count)
{
throw new ArgumentException($"The length of source side and target side must be equal. source length = '{srcSnts[0].Count}', target length = '{tgtSnts[0].Count}'");
}
}
int seqLen = srcSnts[0].Count;
IWeightTensor encodedWeightMatrix = Encode(g.CreateSubGraph("Encoder"), srcSnts, encoder, srcEmbedding);
IWeightTensor ffLayer = decoderFFLayer.Process(encodedWeightMatrix, batchSize, g);
IWeightTensor ffLayerBatch = g.TransposeBatch(ffLayer, batchSize);
// Logger.WriteLine("1");
float cost = 0.0f;
using (var probs = g.Softmax(ffLayerBatch, runGradients: false, inPlace: true))
{
if (isTraining)
{
//Calculate loss for each word in the batch
for (int k = 0; k < batchSize; k++)
{
for (int j = 0; j < seqLen; j++)
{
using (var probs_k_j = g.PeekRow(probs, k * seqLen + j, runGradients: false))
{
var ix_targets_k_j = m_modelMetaData.Vocab.GetTargetWordIndex(tgtSnts[k][j]);
var score_k = probs_k_j.GetWeightAt(ix_targets_k_j);
cost += (float)-Math.Log(score_k);
probs_k_j.SetWeightAt(score_k - 1, ix_targets_k_j);
}
}
////CRF part
//using (var probs_k = g.PeekRow(probs, k * seqLen, seqLen, runGradients: false))
//{
// var weights_k = probs_k.ToWeightArray();
// var crfOutput_k = m_crfDecoder.ForwardBackward(seqLen, weights_k);
// int[] trueTags = new int[seqLen];
// for (int j = 0; j < seqLen; j++)
// {
// trueTags[j] = m_modelMetaData.Vocab.GetTargetWordIndex(tgtSnts[k][j]);
// }
// m_crfDecoder.UpdateBigramTransition(seqLen, crfOutput_k, trueTags);
//}
}
ffLayerBatch.CopyWeightsToGradients(probs);
}
else
{
// CRF decoder
//for (int k = 0; k < batchSize; k++)
//{
// //CRF part
// using (var probs_k = g.PeekRow(probs, k * seqLen, seqLen, runGradients: false))
// {
// var weights_k = probs_k.ToWeightArray();
// var crfOutput_k = m_crfDecoder.DecodeNBestCRF(weights_k, seqLen, 1);
// var targetWords = m_modelMetaData.Vocab.ConvertTargetIdsToString(crfOutput_k[0].ToList());
// tgtSnts.Add(targetWords);
// }
//}
// Output "i"th target word
var targetIdx = g.Argmax(probs, 1);
var targetWords = m_modelMetaData.Vocab.ConvertTargetIdsToString(targetIdx.ToList());
for (int k = 0; k < batchSize; k++)
{
tgtSnts.Add(targetWords.GetRange(k * seqLen, seqLen));
}
}
}
return(cost);
}
/// <summary>
/// Decode output sentences in training
/// </summary>
/// <param name="outputSnts">In training mode, they are golden target sentences, otherwise, they are target sentences generated by the decoder</param>
/// <param name="g"></param>
/// <param name="encOutputs"></param>
/// <param name="decoder"></param>
/// <param name="decoderFFLayer"></param>
/// <param name="tgtEmbedding"></param>
/// <returns></returns>
private float DecodeAttentionLSTM(List <List <string> > outputSnts, IComputeGraph g, IWeightTensor encOutputs, AttentionDecoder decoder, IWeightTensor tgtEmbedding, 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] = m_modelMetaData.Vocab.GetTargetWordIndex(outputSnts[i][0]);
}
// Initialize variables accoridng to current mode
List <int> originalOutputLengths = isTraining ? ParallelCorpus.PadSentences(outputSnts) : null;
int seqLen = isTraining ? outputSnts[0].Count : 64;
float dropoutRatio = isTraining ? m_dropoutRatio : 0.0f;
HashSet <int> setEndSentId = isTraining ? null : new HashSet <int>();
// Pre-process for attention model
AttentionPreProcessResult attPreProcessResult = decoder.PreProcess(encOutputs, batchSize, g);
for (int i = 1; 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(tgtEmbedding, ix_inputs[j]));
}
IWeightTensor inputsM = g.ConcatRows(inputs);
//Decode output sentence at position i
IWeightTensor eOutput = decoder.Decode(inputsM, attPreProcessResult, batchSize, g);
//Softmax for output
using (IWeightTensor 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 (IWeightTensor probs_k = g.PeekRow(probs, k, runGradients: false))
{
int ix_targets_k = m_modelMetaData.Vocab.GetTargetWordIndex(outputSnts[k][i]);
float 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
int[] targetIdx = g.Argmax(probs, 1);
List <string> targetWords = m_modelMetaData.Vocab.ConvertTargetIdsToString(targetIdx.ToList());
for (int j = 0; j < targetWords.Count; j++)
{
if (setEndSentId.Contains(j) == false)
{
outputSnts[j].Add(targetWords[j]);
if (targetWords[j] == ParallelCorpus.EOS)
{
setEndSentId.Add(j);
}
}
}
if (setEndSentId.Count == batchSize)
{
// All target sentences in current batch are finished, so we exit.
break;
}
ix_inputs = targetIdx;
}
}
}
return(cost);
}
private float DecodeTransformer(List <List <string> > outInputSeqs, IComputeGraph g, IWeightTensor encOutputs, IWeightTensor encMask, TransformerDecoder decoder,
IWeightTensor tgtEmbedding, int batchSize, int deviceId, bool isTraining = true)
{
float cost = 0.0f;
var originalInputLengths = ParallelCorpus.PadSentences(outInputSeqs);
int tgtSeqLen = outInputSeqs[0].Count;
IWeightTensor tgtDimMask = MaskUtils.BuildPadDimMask(g, tgtSeqLen, originalInputLengths, m_modelMetaData.HiddenDim, deviceId);
using (IWeightTensor tgtSelfTriMask = MaskUtils.BuildPadSelfTriMask(g, tgtSeqLen, originalInputLengths, deviceId))
{
List <IWeightTensor> inputs = new List <IWeightTensor>();
for (int i = 0; i < batchSize; i++)
{
for (int j = 0; j < tgtSeqLen; j++)
{
int ix_targets_k = m_modelMetaData.Vocab.GetTargetWordIndex(outInputSeqs[i][j], logUnk: true);
inputs.Add(g.PeekRow(tgtEmbedding, ix_targets_k));
}
}
IWeightTensor tgtInputEmbeddings = inputs.Count > 1 ? g.ConcatRows(inputs) : inputs[0];
IWeightTensor decOutput = decoder.Decode(tgtInputEmbeddings, encOutputs, tgtSelfTriMask, encMask, tgtDimMask, batchSize, g);
decOutput = g.Mul(decOutput, g.Transpose(tgtEmbedding));
using (IWeightTensor probs = g.Softmax(decOutput, runGradients: false, inPlace: true))
{
if (isTraining)
{
var leftShiftInputSeqs = ParallelCorpus.LeftShiftSnts(outInputSeqs, ParallelCorpus.EOS);
var originalOutputLengths = ParallelCorpus.PadSentences(leftShiftInputSeqs, tgtSeqLen);
for (int i = 0; i < batchSize; i++)
{
for (int j = 0; j < tgtSeqLen; j++)
{
using (IWeightTensor probs_i_j = g.PeekRow(probs, i * tgtSeqLen + j, runGradients: false))
{
if (j < originalOutputLengths[i])
{
int ix_targets_i_j = m_modelMetaData.Vocab.GetTargetWordIndex(leftShiftInputSeqs[i][j], logUnk: true);
float score_i_j = probs_i_j.GetWeightAt(ix_targets_i_j);
if (j < originalOutputLengths[i])
{
cost += (float)-Math.Log(score_i_j);
}
probs_i_j.SetWeightAt(score_i_j - 1, ix_targets_i_j);
}
else
{
probs_i_j.CleanWeight();
}
}
}
}
decOutput.CopyWeightsToGradients(probs);
}
else
{
// Output "i"th target word
int[] targetIdx = g.Argmax(probs, 1);
List <string> targetWords = m_modelMetaData.Vocab.ConvertTargetIdsToString(targetIdx.ToList());
for (int i = 0; i < batchSize; i++)
{
outInputSeqs[i].Add(targetWords[i * tgtSeqLen + tgtSeqLen - 1]);
}
}
}
}
return(cost);
}
/// <summary>
/// Run forward part on given single device
/// </summary>
/// <param name="computeGraph">The computing graph for current device. It gets created and passed by the framework</param>
/// <param name="srcSnts">A batch of input tokenized sentences in source side</param>
/// <param name="tgtSnts">A batch of output tokenized sentences in target side</param>
/// <param name="deviceIdIdx">The index of current device</param>
/// <returns>The cost of forward part</returns>
private float RunForwardOnSingleDevice(IComputeGraph computeGraph, List <List <string> > srcSnts, List <List <string> > tgtSnts, int deviceIdIdx, bool isTraining)
{
(IEncoder encoder, IDecoder decoder, IWeightTensor srcEmbedding, IWeightTensor tgtEmbedding) = GetNetworksOnDeviceAt(deviceIdIdx);
// Reset networks
encoder.Reset(computeGraph.GetWeightFactory(), srcSnts.Count);
decoder.Reset(computeGraph.GetWeightFactory(), srcSnts.Count);
List <int> originalSrcLengths = ParallelCorpus.PadSentences(srcSnts);
int srcSeqPaddedLen = srcSnts[0].Count;
int batchSize = srcSnts.Count;
IWeightTensor encSelfMask = MaskUtils.BuildPadSelfMask(computeGraph, srcSeqPaddedLen, originalSrcLengths, deviceIdIdx);
IWeightTensor encDimMask = MaskUtils.BuildPadDimMask(computeGraph, srcSeqPaddedLen, originalSrcLengths, m_modelMetaData.HiddenDim, deviceIdIdx);
// Encoding input source sentences
IWeightTensor encOutput = Encode(computeGraph, srcSnts, encoder, srcEmbedding, encSelfMask, encDimMask);
// Generate output decoder sentences
if (decoder is AttentionDecoder)
{
return(DecodeAttentionLSTM(tgtSnts, computeGraph, encOutput, decoder as AttentionDecoder, tgtEmbedding, srcSnts.Count, isTraining));
}
else
{
if (isTraining)
{
List <int> originalTgtLengths = ParallelCorpus.PadSentences(tgtSnts);
int tgtSeqPaddedLen = tgtSnts[0].Count;
IWeightTensor encDecMask = MaskUtils.BuildSrcTgtMask(computeGraph, srcSeqPaddedLen, tgtSeqPaddedLen, originalSrcLengths, originalTgtLengths, deviceIdIdx);
return(DecodeTransformer(tgtSnts, computeGraph, encOutput, encDecMask, decoder as TransformerDecoder, tgtEmbedding, batchSize, deviceIdIdx, isTraining));
}
else
{
for (int i = 0; i < m_maxTgtSntSize; i++)
{
using (var g = computeGraph.CreateSubGraph($"TransformerDecoder_Step_{i}"))
{
List <int> originalTgtLengths = ParallelCorpus.PadSentences(tgtSnts);
int tgtSeqPaddedLen = tgtSnts[0].Count;
IWeightTensor encDecMask = MaskUtils.BuildSrcTgtMask(g, srcSeqPaddedLen, tgtSeqPaddedLen, originalSrcLengths, originalTgtLengths, deviceIdIdx);
DecodeTransformer(tgtSnts, g, encOutput, encDecMask, decoder as TransformerDecoder, tgtEmbedding, batchSize, deviceIdIdx, isTraining);
bool allSntsEnd = true;
for (int j = 0; j < tgtSnts.Count; j++)
{
if (tgtSnts[j][tgtSnts[j].Count - 1] != ParallelCorpus.EOS)
{
allSntsEnd = false;
break;
}
}
if (allSntsEnd)
{
break;
}
}
}
return(0.0f);
}
}
}
/// <summary>
/// Run forward part on given single device
/// </summary>
/// <param name="g">The computing graph for current device. It gets created and passed by the framework</param>
/// <param name="srcSnts">A batch of input tokenized sentences in source side</param>
/// <param name="tgtSnts">A batch of output tokenized sentences in target side. In training mode, it inputs target tokens, otherwise, it outputs target tokens generated by decoder</param>
/// <param name="deviceIdIdx">The index of current device</param>
/// <returns>The cost of forward part</returns>
private float RunForwardOnSingleDevice(IComputeGraph g, List <List <string> > srcSnts, List <List <string> > tgtSnts, int deviceIdIdx, bool isTraining)
{
var(encoder, srcEmbedding, posEmbedding, decoderFFLayer) = this.GetNetworksOnDeviceAt(deviceIdIdx);
// Reset networks
encoder.Reset(g.GetWeightFactory(), srcSnts.Count);
var originalSrcLengths = ParallelCorpus.PadSentences(srcSnts);
var seqLen = srcSnts[0].Count;
var batchSize = srcSnts.Count;
// Encoding input source sentences
var encOutput = this.Encode(g, srcSnts, encoder, srcEmbedding, null, posEmbedding, originalSrcLengths);
var ffLayer = decoderFFLayer.Process(encOutput, batchSize, g);
var ffLayerBatch = g.TransposeBatch(ffLayer, batchSize);
var cost = 0.0f;
using (var probs = g.Softmax(ffLayerBatch, runGradients: false, inPlace: true))
{
if (isTraining)
{
//Calculate loss for each word in the batch
for (var k = 0; k < batchSize; k++)
{
for (var j = 0; j < seqLen; j++)
{
using (var probs_k_j = g.PeekRow(probs, k * seqLen + j, runGradients: false))
{
var ix_targets_k_j = this.m_modelMetaData.Vocab.GetTargetWordIndex(tgtSnts[k][j]);
var score_k = probs_k_j.GetWeightAt(ix_targets_k_j);
cost += (float)-Math.Log(score_k);
probs_k_j.SetWeightAt(score_k - 1, ix_targets_k_j);
}
}
////CRF part
//using (var probs_k = g.PeekRow(probs, k * seqLen, seqLen, runGradients: false))
//{
// var weights_k = probs_k.ToWeightArray();
// var crfOutput_k = m_crfDecoder.ForwardBackward(seqLen, weights_k);
// int[] trueTags = new int[seqLen];
// for (int j = 0; j < seqLen; j++)
// {
// trueTags[j] = m_modelMetaData.Vocab.GetTargetWordIndex(tgtSnts[k][j]);
// }
// m_crfDecoder.UpdateBigramTransition(seqLen, crfOutput_k, trueTags);
//}
}
ffLayerBatch.CopyWeightsToGradients(probs);
}
else
{
// CRF decoder
//for (int k = 0; k < batchSize; k++)
//{
// //CRF part
// using (var probs_k = g.PeekRow(probs, k * seqLen, seqLen, runGradients: false))
// {
// var weights_k = probs_k.ToWeightArray();
// var crfOutput_k = m_crfDecoder.DecodeNBestCRF(weights_k, seqLen, 1);
// var targetWords = m_modelMetaData.Vocab.ConvertTargetIdsToString(crfOutput_k[0].ToList());
// tgtSnts.Add(targetWords);
// }
//}
// Output "i"th target word
var targetIdx = g.Argmax(probs, 1);
var targetWords = this.m_modelMetaData.Vocab.ConvertTargetIdsToString(targetIdx.ToList());
for (var k = 0; k < batchSize; k++)
{
tgtSnts[k] = targetWords.GetRange(k * seqLen, seqLen);
}
}
}
return(cost);
}
private float DecodeTransformer(List <List <string> > tgtSeqs, IComputeGraph g, IWeightTensor encOutputs, TransformerDecoder decoder,
IWeightTensor tgtEmbedding, IWeightTensor posEmbedding, int batchSize, int deviceId, List <int> srcOriginalLenghts, bool isTraining = true)
{
float cost = 0.0f;
var tgtOriginalLengths = ParallelCorpus.PadSentences(tgtSeqs);
int tgtSeqLen = tgtSeqs[0].Count;
int srcSeqLen = encOutputs.Rows / batchSize;
using (IWeightTensor srcTgtMask = MaskUtils.BuildSrcTgtMask(g, srcSeqLen, tgtSeqLen, tgtOriginalLengths, srcOriginalLenghts, deviceId))
{
using (IWeightTensor tgtSelfTriMask = MaskUtils.BuildPadSelfTriMask(g, tgtSeqLen, tgtOriginalLengths, deviceId))
{
List <IWeightTensor> inputs = new List <IWeightTensor>();
for (int i = 0; i < batchSize; i++)
{
for (int j = 0; j < tgtSeqLen; j++)
{
int ix_targets_k = m_modelMetaData.Vocab.GetTargetWordIndex(tgtSeqs[i][j], logUnk: true);
var emb = g.PeekRow(tgtEmbedding, ix_targets_k, runGradients: j < tgtOriginalLengths[i] ? true : false);
inputs.Add(emb);
}
}
IWeightTensor inputEmbs = inputs.Count > 1 ? g.ConcatRows(inputs) : inputs[0];
inputEmbs = AddPositionEmbedding(g, posEmbedding, batchSize, tgtSeqLen, inputEmbs);
IWeightTensor decOutput = decoder.Decode(inputEmbs, encOutputs, tgtSelfTriMask, srcTgtMask, batchSize, g);
using (IWeightTensor probs = g.Softmax(decOutput, runGradients: false, inPlace: true))
{
if (isTraining)
{
var leftShiftInputSeqs = ParallelCorpus.LeftShiftSnts(tgtSeqs, ParallelCorpus.EOS);
for (int i = 0; i < batchSize; i++)
{
for (int j = 0; j < tgtSeqLen; j++)
{
using (IWeightTensor probs_i_j = g.PeekRow(probs, i * tgtSeqLen + j, runGradients: false))
{
if (j < tgtOriginalLengths[i])
{
int ix_targets_i_j = m_modelMetaData.Vocab.GetTargetWordIndex(leftShiftInputSeqs[i][j], logUnk: true);
float score_i_j = probs_i_j.GetWeightAt(ix_targets_i_j);
cost += (float)-Math.Log(score_i_j);
probs_i_j.SetWeightAt(score_i_j - 1, ix_targets_i_j);
}
else
{
probs_i_j.CleanWeight();
}
}
}
}
decOutput.CopyWeightsToGradients(probs);
}
//if (isTraining)
//{
// var leftShiftInputSeqs = ParallelCorpus.LeftShiftSnts(tgtSeqs, ParallelCorpus.EOS);
// int[] targetIds = new int[batchSize * tgtSeqLen];
// int ids = 0;
// for (int i = 0; i < batchSize; i++)
// {
// for (int j = 0; j < tgtSeqLen; j++)
// {
// targetIds[ids] = j < tgtOriginalLengths[i] ? m_modelMetaData.Vocab.GetTargetWordIndex(leftShiftInputSeqs[i][j], logUnk: true) : -1;
// ids++;
// }
// }
// cost += g.UpdateCost(probs, targetIds);
// decOutput.CopyWeightsToGradients(probs);
//}
else
{
// Output "i"th target word
int[] targetIdx = g.Argmax(probs, 1);
List <string> targetWords = m_modelMetaData.Vocab.ConvertTargetIdsToString(targetIdx.ToList());
for (int i = 0; i < batchSize; i++)
{
tgtSeqs[i].Add(targetWords[i * tgtSeqLen + tgtSeqLen - 1]);
}
}
}
}
}
return(cost);
}
/// <summary>
/// Run forward part on given single device
/// </summary>
/// <param name="computeGraph">The computing graph for current device. It gets created and passed by the framework</param>
/// <param name="srcSnts">A batch of input tokenized sentences in source side</param>
/// <param name="tgtSnts">A batch of output tokenized sentences in target side</param>
/// <param name="deviceIdIdx">The index of current device</param>
/// <returns>The cost of forward part</returns>
private float RunForwardOnSingleDevice(IComputeGraph computeGraph, List <List <string> > srcSnts, List <List <string> > tgtSnts, int deviceIdIdx, bool isTraining)
{
(IEncoder encoder, IDecoder decoder, IWeightTensor srcEmbedding, IWeightTensor tgtEmbedding, IWeightTensor posEmbedding) = GetNetworksOnDeviceAt(deviceIdIdx);
// Reset networks
encoder.Reset(computeGraph.GetWeightFactory(), srcSnts.Count);
decoder.Reset(computeGraph.GetWeightFactory(), srcSnts.Count);
List <int> originalSrcLengths = ParallelCorpus.PadSentences(srcSnts);
int srcSeqPaddedLen = srcSnts[0].Count;
int batchSize = srcSnts.Count;
IWeightTensor srcSelfMask = m_shuffleType == ShuffleEnums.NoPaddingInSrc ? null : MaskUtils.BuildPadSelfMask(computeGraph, srcSeqPaddedLen, originalSrcLengths, DeviceIds[deviceIdIdx]); // The length of source sentences are same in a single mini-batch, so we don't have source mask.
// Encoding input source sentences
IWeightTensor encOutput = Encode(computeGraph, srcSnts, encoder, srcEmbedding, srcSelfMask, posEmbedding, originalSrcLengths);
if (srcSelfMask != null)
{
srcSelfMask.Dispose();
}
// Generate output decoder sentences
if (decoder is AttentionDecoder)
{
return(DecodeAttentionLSTM(tgtSnts, computeGraph, encOutput, decoder as AttentionDecoder, tgtEmbedding, srcSnts.Count, isTraining));
}
else
{
if (isTraining)
{
return(DecodeTransformer(tgtSnts, computeGraph, encOutput, decoder as TransformerDecoder, tgtEmbedding, posEmbedding, batchSize, DeviceIds[deviceIdIdx], originalSrcLengths, isTraining));
}
else
{
for (int i = 0; i < m_maxTgtSntSize; i++)
{
using (var g = computeGraph.CreateSubGraph($"TransformerDecoder_Step_{i}"))
{
DecodeTransformer(tgtSnts, g, encOutput, decoder as TransformerDecoder, tgtEmbedding, posEmbedding, batchSize, DeviceIds[deviceIdIdx], originalSrcLengths, isTraining);
bool allSntsEnd = true;
for (int j = 0; j < tgtSnts.Count; j++)
{
if (tgtSnts[j][tgtSnts[j].Count - 1] != ParallelCorpus.EOS)
{
allSntsEnd = false;
break;
}
}
if (allSntsEnd)
{
break;
}
}
}
RemoveDuplicatedEOS(tgtSnts);
return(0.0f);
}
}
}
/// <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, 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
List <int> originalOutputLengths = isTraining ? ParallelCorpus.PadSentences(outputSentences) : null;
int seqLen = isTraining ? outputSentences[0].Count : 64;
float 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
AttentionPreProcessResult 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]));
}
IWeightTensor inputsM = g.ConcatRows(inputs);
//Decode output sentence at position i
IWeightTensor eOutput = decoder.Decode(inputsM, attPreProcessResult, batchSize, g);
//Softmax for output
using (IWeightTensor 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 (IWeightTensor probs_k = g.PeekRow(probs, k, runGradients: false))
{
int ix_targets_k = m_modelMetaData.Vocab.GetTargetWordIndex(outputSentences[k][i]);
float 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
int[] targetIdx = g.Argmax(probs, 1);
List <string> 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);
}