/// <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>
public override List <NetworkResult> RunForwardOnSingleDevice(IComputeGraph g, ISntPairBatch sntPairBatch, int deviceIdIdx, bool isTraining, DecodingOptions decodingOptions)
{
List <NetworkResult> nrs = new List <NetworkResult>();
var srcSnts = sntPairBatch.GetSrcTokens(0);
var tgtSnts = sntPairBatch.GetTgtTokens(0);
(IEncoder encoder, IWeightTensor srcEmbedding, IWeightTensor posEmbedding, FeedForwardLayer decoderFFLayer) = GetNetworksOnDeviceAt(deviceIdIdx);
// Reset networks
encoder.Reset(g.GetWeightFactory(), srcSnts.Count);
var originalSrcLengths = BuildInTokens.PadSentences(srcSnts);
var srcTokensList = m_modelMetaData.SrcVocab.GetWordIndex(srcSnts);
BuildInTokens.PadSentences(tgtSnts);
var tgtTokensLists = m_modelMetaData.ClsVocab.GetWordIndex(tgtSnts);
int seqLen = srcSnts[0].Count;
int batchSize = srcSnts.Count;
// Encoding input source sentences
IWeightTensor encOutput = Encoder.Run(g, sntPairBatch, encoder, m_modelMetaData, m_shuffleType, srcEmbedding, posEmbedding, null, srcTokensList, originalSrcLengths);
IWeightTensor ffLayer = decoderFFLayer.Process(encOutput, batchSize, g);
float cost = 0.0f;
IWeightTensor probs = g.Softmax(ffLayer, inPlace: true);
if (isTraining)
{
var tgtTokensTensor = g.CreateTokensTensor(tgtTokensLists);
cost = g.CrossEntropyLoss(probs, tgtTokensTensor);
}
else
{
// Output "i"th target word
using var targetIdxTensor = g.Argmax(probs, 1);
float[] targetIdx = targetIdxTensor.ToWeightArray();
List <string> targetWords = m_modelMetaData.ClsVocab.ConvertIdsToString(targetIdx.ToList());
for (int k = 0; k < batchSize; k++)
{
tgtSnts[k] = targetWords.GetRange(k * seqLen, seqLen);
}
}
NetworkResult nr = new NetworkResult
{
Cost = cost,
Output = new List <List <List <string> > >()
};
nr.Output.Add(tgtSnts);
nrs.Add(nr);
return(nrs);
}
/// <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="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>
/// 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);
}
/// <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>
public override List <NetworkResult> RunForwardOnSingleDevice(IComputeGraph computeGraph, ISntPairBatch sntPairBatch, int deviceIdIdx, bool isTraining, DecodingOptions decodingOptions)
{
int batchSize = sntPairBatch.BatchSize;
float cost = 0.0f;
var nrs = new List <NetworkResult>();
var nr = new NetworkResult {
Output = new List <List <List <string> > >()
};
(IEncoder encoder, IWeightTensor srcEmbedding, IFeedForwardLayer encoderFFLayer, IWeightTensor posEmbedding, IWeightTensor segmentEmbedding) = GetNetworksOnDeviceAt(deviceIdIdx);
IWeightTensor encOutput1;
IWeightTensor encOutput2;
if (!isTraining && (m_options.ProcessorType == ProcessorTypeEnums.CPU))
{
//We only check cache at inference time
string cacheKey1 = GenerateCacheKey(sntPairBatch.GetSrcTokens(0));
if (!m_memoryCache.TryGetValue(cacheKey1, out encOutput1))
{
encOutput1 = Encoder.BuildTensorForSourceTokenGroupAt(computeGraph, sntPairBatch, m_shuffleType, encoder, m_modelMetaData, srcEmbedding, posEmbedding, segmentEmbedding, 0); // output shape: [batch_size, dim]
var cacheEntryOptions = new MemoryCacheEntryOptions().SetSize(1);
m_memoryCache.Set(cacheKey1, encOutput1.CopyWeightsRef($"cache_{encOutput1.Name}", false), cacheEntryOptions);
}
string cacheKey2 = GenerateCacheKey(sntPairBatch.GetSrcTokens(1));
if (!m_memoryCache.TryGetValue(cacheKey2, out encOutput2))
{
encOutput2 = Encoder.BuildTensorForSourceTokenGroupAt(computeGraph, sntPairBatch, m_shuffleType, encoder, m_modelMetaData, srcEmbedding, posEmbedding, segmentEmbedding, 1); // output_shape: [batch_size, dim]
var cacheEntryOptions = new MemoryCacheEntryOptions().SetSize(1);
m_memoryCache.Set(cacheKey2, encOutput2.CopyWeightsRef($"cache_{encOutput2.Name}", false), cacheEntryOptions);
}
}
else
{
//We always run encoder network during training time or using GPUs
encOutput1 = Encoder.BuildTensorForSourceTokenGroupAt(computeGraph, sntPairBatch, m_shuffleType, encoder, m_modelMetaData, srcEmbedding, posEmbedding, segmentEmbedding, 0); // output shape: [batch_size, dim]
encOutput2 = Encoder.BuildTensorForSourceTokenGroupAt(computeGraph, sntPairBatch, m_shuffleType, encoder, m_modelMetaData, srcEmbedding, posEmbedding, segmentEmbedding, 1); // output_shape: [batch_size, dim]
}
if (m_modelMetaData.SimilarityType.Equals("Continuous", StringComparison.InvariantCultureIgnoreCase))
{
// Cosine similairy
var w12 = computeGraph.EltMul(encOutput1, encOutput2);
w12 = computeGraph.Sum(w12, 1);
var w1 = computeGraph.EltMul(encOutput1, encOutput1);
w1 = computeGraph.Sum(w1, 1);
var w2 = computeGraph.EltMul(encOutput2, encOutput2);
w2 = computeGraph.Sum(w2, 1);
var n12 = computeGraph.EltMul(w1, w2);
n12 = computeGraph.Rsqrt(n12);
var probs = computeGraph.EltMul(w12, n12);
if (isTraining)
{
var tgtSnts = sntPairBatch.GetTgtTokens(0);
for (int k = 0; k < batchSize; k++)
{
float golden_score_k = float.Parse(tgtSnts[k][0]); // Get golden similiary score from target side
float score_k = probs.GetWeightAt(new long[] { k, 0 });
probs.SetWeightAt(score_k - golden_score_k, new long[] { k, 0 });
cost += (float)Math.Abs(score_k - golden_score_k);
}
probs.CopyWeightsToGradients(probs);
nr.Cost = cost / batchSize;
}
else
{
nr.Output.Add(new List <List <string> >());
for (int k = 0; k < batchSize; k++)
{
float score_k = probs.GetWeightAt(new long[] { k, 0 });
nr.Output[0].Add(new List <string>());
nr.Output[0][k].Add(score_k.ToString());
}
}
}
else
{
IWeightTensor encOutput = computeGraph.EltMul(encOutput1, encOutput2);
IWeightTensor ffLayer = encoderFFLayer.Process(encOutput, batchSize, computeGraph);
using (IWeightTensor probs = computeGraph.Softmax(ffLayer, runGradients: false, inPlace: true))
{
if (isTraining)
{
var tgtSnts = sntPairBatch.GetTgtTokens(0);
for (int k = 0; k < batchSize; k++)
{
int ix_targets_k_j = m_modelMetaData.ClsVocab.GetWordIndex(tgtSnts[k][0]);
float score_k = probs.GetWeightAt(new long[] { k, ix_targets_k_j });
cost += (float)-Math.Log(score_k);
probs.SetWeightAt(score_k - 1, new long[] { k, ix_targets_k_j });
}
ffLayer.CopyWeightsToGradients(probs);
nr.Cost = cost / batchSize;
}
else
{
// Output "i"th target word
using var targetIdxTensor = computeGraph.Argmax(probs, 1);
float[] targetIdx = targetIdxTensor.ToWeightArray();
List <string> targetWords = m_modelMetaData.ClsVocab.ConvertIdsToString(targetIdx.ToList());
nr.Output.Add(new List <List <string> >());
for (int k = 0; k < batchSize; k++)
{
nr.Output[0].Add(new List <string>());
nr.Output[0][k].Add(targetWords[k]);
}
}
}
}
nrs.Add(nr);
return(nrs);
}
/// <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>
public override List <NetworkResult> RunForwardOnSingleDevice(IComputeGraph computeGraph, ISntPairBatch sntPairBatch, int deviceIdIdx, bool isTraining, DecodingOptions decodingOptions)
{
List <NetworkResult> nrs = new List <NetworkResult>();
(IEncoder encoder, IWeightTensor srcEmbedding, List <IFeedForwardLayer> encoderFFLayer, IWeightTensor posEmbedding, IWeightTensor segmentEmbedding) = GetNetworksOnDeviceAt(deviceIdIdx);
var srcSnts = sntPairBatch.GetSrcTokens(0);
var originalSrcLengths = BuildInTokens.PadSentences(srcSnts);
var srcTokensList = m_modelMetaData.SrcVocab.GetWordIndex(srcSnts);
IWeightTensor encOutput = Encoder.Run(computeGraph, sntPairBatch, encoder, m_modelMetaData, m_shuffleType, srcEmbedding, posEmbedding, segmentEmbedding, srcTokensList, originalSrcLengths);
int srcSeqPaddedLen = srcSnts[0].Count;
int batchSize = srcSnts.Count;
float[] clsIdxs = new float[batchSize];
for (int i = 0; i < batchSize; i++)
{
for (int j = 0; j < srcSnts[i].Count; j++)
{
if (srcSnts[i][j] == BuildInTokens.CLS)
{
clsIdxs[i] = i * srcSeqPaddedLen + j;
break;
}
}
}
IWeightTensor clsWeightTensor = computeGraph.IndexSelect(encOutput, clsIdxs);
for (int i = 0; i < m_encoderFFLayer.Length; i++)
{
float cost = 0.0f;
NetworkResult nr = new NetworkResult
{
Output = new List <List <List <string> > >()
};
IWeightTensor ffLayer = encoderFFLayer[i].Process(clsWeightTensor, batchSize, computeGraph);
using (IWeightTensor probs = computeGraph.Softmax(ffLayer, runGradients: false, inPlace: true))
{
if (isTraining)
{
var tgtSnts = sntPairBatch.GetTgtTokens(i);
for (int k = 0; k < batchSize; k++)
{
int ix_targets_k_j = m_modelMetaData.ClsVocabs[i].GetWordIndex(tgtSnts[k][0]);
float score_k = probs.GetWeightAt(new long[] { k, ix_targets_k_j });
cost += (float)-Math.Log(score_k);
probs.SetWeightAt(score_k - 1, new long[] { k, ix_targets_k_j });
}
ffLayer.CopyWeightsToGradients(probs);
nr.Cost = cost / batchSize;
}
else
{
// Output "i"th target word
using var targetIdxTensor = computeGraph.Argmax(probs, 1);
float[] targetIdx = targetIdxTensor.ToWeightArray();
List <string> targetWords = m_modelMetaData.ClsVocabs[i].ConvertIdsToString(targetIdx.ToList());
nr.Output.Add(new List <List <string> >());
for (int k = 0; k < batchSize; k++)
{
nr.Output[0].Add(new List <string>());
nr.Output[0][k].Add(targetWords[k]);
}
}
}
nrs.Add(nr);
}
return(nrs);
}
/// <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>
public override List <NetworkResult> RunForwardOnSingleDevice(IComputeGraph computeGraph, ISntPairBatch sntPairBatch, int deviceIdIdx, bool isTraining, DecodingOptions decodingOptions)
{
(IEncoder encoder, IDecoder decoder, IFeedForwardLayer encoderFFLayer, IFeedForwardLayer decoderFFLayer, IWeightTensor srcEmbedding, IWeightTensor tgtEmbedding, IWeightTensor posEmbedding, IWeightTensor segmentEmbedding) = GetNetworksOnDeviceAt(deviceIdIdx);
var srcSnts = sntPairBatch.GetSrcTokens(0);
var originalSrcLengths = BuildInTokens.PadSentences(srcSnts);
var srcTokensList = m_modelMetaData.SrcVocab.GetWordIndex(srcSnts);
IWeightTensor encOutput = Encoder.Run(computeGraph, sntPairBatch, encoder, m_modelMetaData, m_shuffleType, srcEmbedding, posEmbedding, segmentEmbedding, srcTokensList, originalSrcLengths);
List <NetworkResult> nrs = new List <NetworkResult>();
int srcSeqPaddedLen = srcSnts[0].Count;
int batchSize = srcSnts.Count;
float[] clsIdxs = new float[batchSize];
for (int i = 0; i < batchSize; i++)
{
for (int j = 0; j < srcSnts[i].Count; j++)
{
if (srcSnts[i][j] == BuildInTokens.CLS)
{
clsIdxs[i] = i * srcSeqPaddedLen + j;
break;
}
}
}
IWeightTensor clsWeightTensor = computeGraph.IndexSelect(encOutput, clsIdxs);
float cost = 0.0f;
NetworkResult nrCLS = new NetworkResult
{
Output = new List <List <List <string> > >()
};
IWeightTensor ffLayer = encoderFFLayer.Process(clsWeightTensor, batchSize, computeGraph);
using (IWeightTensor probs = computeGraph.Softmax(ffLayer, runGradients: false, inPlace: true))
{
if (isTraining)
{
var clsSnts = sntPairBatch.GetTgtTokens(0);
for (int k = 0; k < batchSize; k++)
{
int ix_targets_k_j = m_modelMetaData.ClsVocab.GetWordIndex(clsSnts[k][0]);
float score_k = probs.GetWeightAt(new long[] { k, ix_targets_k_j });
cost += (float)-Math.Log(score_k);
probs.SetWeightAt(score_k - 1, new long[] { k, ix_targets_k_j });
}
ffLayer.CopyWeightsToGradients(probs);
nrCLS.Cost = cost / batchSize;
}
else
{
// Output "i"th target word
using var targetIdxTensor = computeGraph.Argmax(probs, 1);
float[] targetIdx = targetIdxTensor.ToWeightArray();
List <string> targetWords = m_modelMetaData.ClsVocab.ConvertIdsToString(targetIdx.ToList());
nrCLS.Output.Add(new List <List <string> >());
for (int k = 0; k < batchSize; k++)
{
nrCLS.Output[0].Add(new List <string>());
nrCLS.Output[0][k].Add(targetWords[k]);
}
}
}
// Reset networks
decoder.Reset(computeGraph.GetWeightFactory(), srcSnts.Count);
// Generate output decoder sentences
var tgtSnts = sntPairBatch.GetTgtTokens(1);
var tgtTokensList = m_modelMetaData.TgtVocab.GetWordIndex(tgtSnts);
NetworkResult nr = new NetworkResult();
if (decoder is AttentionDecoder)
{
nr.Cost = Decoder.DecodeAttentionLSTM(tgtTokensList, computeGraph, encOutput, decoder as AttentionDecoder, decoderFFLayer, tgtEmbedding, m_modelMetaData.TgtVocab, srcSnts.Count, isTraining);
nr.Output = new List <List <List <string> > >
{
m_modelMetaData.TgtVocab.ConvertIdsToString(tgtTokensList)
};
}
else
{
if (isTraining)
{
(var c, _) = Decoder.DecodeTransformer(tgtTokensList, computeGraph, encOutput, decoder as TransformerDecoder, decoderFFLayer, tgtEmbedding, posEmbedding, originalSrcLengths, m_modelMetaData.TgtVocab, m_shuffleType, m_options.DropoutRatio, null, isTraining);
nr.Cost = c;
nr.Output = null;
}
else
{
List <List <BeamSearchStatus> > beam2batchStatus = Decoder.InitBeamSearchStatusListList(batchSize, tgtTokensList);
for (int i = 0; i < decodingOptions.MaxTgtSentLength; i++)
{
List <List <BeamSearchStatus> > batch2beam2seq = null; //(batch_size, beam_search_size)
try
{
foreach (var batchStatus in beam2batchStatus)
{
var batch2tgtTokens = Decoder.ExtractBatchTokens(batchStatus);
using var g = computeGraph.CreateSubGraph($"TransformerDecoder_Step_{i}");
(var cost2, var bssSeqList) = Decoder.DecodeTransformer(batch2tgtTokens, g, encOutput, decoder as TransformerDecoder, decoderFFLayer, tgtEmbedding, posEmbedding,
originalSrcLengths, m_modelMetaData.TgtVocab, m_shuffleType, 0.0f, decodingOptions, isTraining,
outputSentScore: decodingOptions.BeamSearchSize > 1, previousBeamSearchResults: batchStatus);
bssSeqList = Decoder.SwapBeamAndBatch(bssSeqList);
batch2beam2seq = Decoder.CombineBeamSearchResults(batch2beam2seq, bssSeqList);
}
}
catch (OutOfMemoryException)
{
GC.Collect();
Logger.WriteLine(Logger.Level.warn, $"We have out of memory while generating '{i}th' tokens, so terminate decoding for current sequences.");
break;
}
if (decodingOptions.BeamSearchSize > 1)
{
// Keep top N result and drop all others
for (int k = 0; k < batchSize; k++)
{
batch2beam2seq[k] = BeamSearch.GetTopNBSS(batch2beam2seq[k], decodingOptions.BeamSearchSize);
}
}
beam2batchStatus = Decoder.SwapBeamAndBatch(batch2beam2seq);
if (Decoder.AreAllSentsCompleted(beam2batchStatus))
{
break;
}
}
nr.Cost = 0.0f;
nr.Output = m_modelMetaData.TgtVocab.ExtractTokens(beam2batchStatus);
}
}
nr.RemoveDuplicatedEOS();
nrs.Add(nrCLS);
nrs.Add(nr);
return(nrs);
}