/// <summary>
/// Given input sentence and generate output sentence by seq2seq model with beam search
/// </summary>
/// <param name="input"></param>
/// <param name="beamSearchSize"></param>
/// <param name="maxOutputLength"></param>
/// <returns></returns>
public List <List <string> > Predict(List <string> input, int beamSearchSize = 1, int maxOutputLength = 100)
{
(IEncoder encoder, IDecoder decoder, IWeightTensor srcEmbedding, IWeightTensor tgtEmbedding, IWeightTensor posEmbedding) = GetNetworksOnDeviceAt(-1);
List <List <string> > inputSeqs = ParallelCorpus.ConstructInputTokens(input);
int batchSize = 1; // For predict with beam search, we currently only supports one sentence per call
IComputeGraph g = CreateComputGraph(m_defaultDeviceId, needBack: false);
AttentionDecoder rnnDecoder = decoder as AttentionDecoder;
encoder.Reset(g.GetWeightFactory(), batchSize);
rnnDecoder.Reset(g.GetWeightFactory(), batchSize);
// Construct beam search status list
List <BeamSearchStatus> bssList = new List <BeamSearchStatus>();
BeamSearchStatus bss = new BeamSearchStatus();
bss.OutputIds.Add((int)SENTTAGS.START);
bss.CTs = rnnDecoder.GetCTs();
bss.HTs = rnnDecoder.GetHTs();
bssList.Add(bss);
IWeightTensor encodedWeightMatrix = Encode(g, inputSeqs, encoder, srcEmbedding, null, posEmbedding, null);
AttentionPreProcessResult attPreProcessResult = rnnDecoder.PreProcess(encodedWeightMatrix, batchSize, g);
List <BeamSearchStatus> newBSSList = new List <BeamSearchStatus>();
bool finished = false;
int outputLength = 0;
while (finished == false && outputLength < maxOutputLength)
{
finished = true;
for (int i = 0; i < bssList.Count; i++)
{
bss = bssList[i];
if (bss.OutputIds[bss.OutputIds.Count - 1] == (int)SENTTAGS.END)
{
newBSSList.Add(bss);
}
else if (bss.OutputIds.Count > maxOutputLength)
{
newBSSList.Add(bss);
}
else
{
finished = false;
int ix_input = bss.OutputIds[bss.OutputIds.Count - 1];
rnnDecoder.SetCTs(bss.CTs);
rnnDecoder.SetHTs(bss.HTs);
IWeightTensor x = g.PeekRow(tgtEmbedding, ix_input);
IWeightTensor eOutput = rnnDecoder.Decode(x, attPreProcessResult, batchSize, g);
using (IWeightTensor probs = g.Softmax(eOutput))
{
List <int> preds = probs.GetTopNMaxWeightIdx(beamSearchSize);
for (int j = 0; j < preds.Count; j++)
{
BeamSearchStatus newBSS = new BeamSearchStatus();
newBSS.OutputIds.AddRange(bss.OutputIds);
newBSS.OutputIds.Add(preds[j]);
newBSS.CTs = rnnDecoder.GetCTs();
newBSS.HTs = rnnDecoder.GetHTs();
float score = probs.GetWeightAt(preds[j]);
newBSS.Score = bss.Score;
newBSS.Score += (float)(-Math.Log(score));
//var lengthPenalty = Math.Pow((5.0f + newBSS.OutputIds.Count) / 6, 0.6);
//newBSS.Score /= (float)lengthPenalty;
newBSSList.Add(newBSS);
}
}
}
}
bssList = BeamSearch.GetTopNBSS(newBSSList, beamSearchSize);
newBSSList.Clear();
outputLength++;
}
// Convert output target word ids to real string
List <List <string> > results = new List <List <string> >();
for (int i = 0; i < bssList.Count; i++)
{
results.Add(m_modelMetaData.Vocab.ConvertTargetIdsToString(bssList[i].OutputIds));
}
return(results);
}
/// <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)
{
(var encoder, var decoder, var decoderFFLayer, var srcEmbedding, var tgtEmbedding, var posEmbedding, var segmentEmbedding, var pointerGenerator) = GetNetworksOnDeviceAt(deviceIdIdx);
var srcSnts = sntPairBatch.GetSrcTokens(0);
var originalSrcLengths = BuildInTokens.PadSentences(srcSnts);
var srcTokensList = m_modelMetaData.SrcVocab.GetWordIndex(srcSnts);
if (isTraining && srcSnts[0].Count > m_options.MaxTrainSrcSentLength + 2)
{
throw new InvalidDataException($"The source sentence is too long. Its length = '{srcSnts[0].Count}', but MaxTrainSrcSentLength is '{m_options.MaxTrainSrcSentLength}'. The sentence is '{string.Join(" ", srcSnts[0])}'");
}
IWeightTensor encOutput;
if (!isTraining && (m_options.ProcessorType == ProcessorTypeEnums.CPU))
{
// Try to get src tensor from cache
string cacheKey = GenerateCacheKey(srcSnts);
if (!m_memoryCache.TryGetValue(cacheKey, out encOutput))
{
encOutput = Encoder.Run(computeGraph, sntPairBatch, encoder, m_modelMetaData, m_shuffleType, srcEmbedding, posEmbedding, segmentEmbedding, srcTokensList, originalSrcLengths);
var cacheEntryOptions = new MemoryCacheEntryOptions().SetSize(1);
m_memoryCache.Set(cacheKey, encOutput.CopyWeightsRef($"cache_{encOutput.Name}", false), cacheEntryOptions);
}
}
else
{
// Compute src tensor
encOutput = Encoder.Run(computeGraph, sntPairBatch, encoder, m_modelMetaData, m_shuffleType, srcEmbedding, posEmbedding, segmentEmbedding, srcTokensList, originalSrcLengths);
}
List <NetworkResult> nrs = new List <NetworkResult>();
// Generate output decoder sentences
int batchSize = srcSnts.Count;
var tgtSnts = sntPairBatch.GetTgtTokens(0);
var tgtTokensList = m_modelMetaData.TgtVocab.GetWordIndex(tgtSnts);
NetworkResult nr = new NetworkResult();
decoder.Reset(computeGraph.GetWeightFactory(), srcSnts.Count);
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, pointerGenerator: pointerGenerator, srcSeqs: srcTokensList);
nr.Cost = c;
nr.Output = null;
}
else
{
Dictionary <string, IWeightTensor> cachedTensors = new Dictionary <string, IWeightTensor>();
List <List <BeamSearchStatus> > beam2batchStatus = Decoder.InitBeamSearchStatusListList(batchSize, tgtTokensList);
for (int i = tgtTokensList[0].Count; 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,
pointerGenerator: pointerGenerator, srcSeqs: srcTokensList,
cachedTensors: cachedTensors);
bssSeqList = Decoder.SwapBeamAndBatch(bssSeqList); // Swap shape: (beam_search_size, batch_size) -> (batch_size, beam_search_size)
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);
if (cachedTensors != null)
{
foreach (var pair in cachedTensors)
{
pair.Value.Dispose();
}
}
}
}
nr.RemoveDuplicatedEOS();
nrs.Add(nr);
return(nrs);
}
