output all beam search results in layer group.

paddle/gserver/gradientmachines/RecurrentGradientMachine.cpp

@@ -1012,11 +1012,6 @@ void RecurrentGradientMachine::generateSequence() {
                            /* width */ resultNum,
                            false,
                            /* useGpu */ false);
-    Matrix::resizeOrCreate(generator_.outArg.value,
-                           /* height */ maxGenWordCount,
-                           /* width */ 1,
-                           false,
-                           /* useGpu */ false);
   }
   ICpuGpuVector::resizeOrCreate(generator_.outArg.sequenceStartPositions,
                                 numSequences + 1,
@@ -1026,7 +1021,7 @@ void RecurrentGradientMachine::generateSequence() {
   } else {
     oneWaySearch(numSequences);
   }
-  if (dataArgsSize_) createDataOutlink(batchMachineIdVec_);
+  if (dataArgsSize_) createDataOutlink();
 
   size_t size = generator_.ids.size();
   generator_.outArg.ids->resize(size);
@@ -1106,6 +1101,7 @@ void RecurrentGradientMachine::oneWaySearch(size_t batchSize) {
   }
 
   batchMachineIdVec_.clear();
+  batchMachineStartPos_.clear();
   int* starts = generator_.outArg.sequenceStartPositions->getMutableData(false);
   starts[0] = 0;
   generator_.ids.clear();
@@ -1312,13 +1308,20 @@ void RecurrentGradientMachine::fillGenOutputs() {
     finalPaths_[i].resize(minFinalPathsSize);
   }
 
-  batchMachineIdVec_.clear();
   generator_.ids.clear();
   int* starts = generator_.outArg.sequenceStartPositions->getMutableData(false);
   starts[0] = 0;
   if (numResults > 1) {
-    real* probs = generator_.outArg.in->getData();
+    int idsProbSaveSize = 0;
+    for (auto inSeq : finalPaths_) {
+      for (auto path : inSeq) idsProbSaveSize += path.ids.size();
+      idsProbSaveSize += inSeq.size();
+    }
+    Matrix::resizeOrCreate(
+        generator_.outArg.value, idsProbSaveSize, 1, false, false);
     real* idsProb = generator_.outArg.value->getData();
+
+    real* probs = generator_.outArg.in->getData();
     size_t curPos = 0;
     for (size_t i = 0; i < finalPaths_.size(); ++i) {
       for (size_t j = 0; j < finalPaths_[i].size(); ++j) {
@@ -1333,24 +1336,16 @@ void RecurrentGradientMachine::fillGenOutputs() {
         curPos += genLen;
         idsProb[curPos++] = -1.0;
         probs[i * numResults + j] = path.logProb;
-
-        if (!j && dataArgsSize_) {
-          // in beam search, here only reserved the top 1 generated result
-          // for out_links that are not the generated word indices.
-          batchMachineIdVec_.insert(batchMachineIdVec_.end(),
-                                    path.machineIdVec.begin(),
-                                    path.machineIdVec.end());
-        }
       }
       starts[i + 1] = generator_.ids.size();
     }
   } else {
     for (size_t i = 0; i < finalPaths_.size(); ++i) {
       CHECK(!finalPaths_[i].empty());
-      generator_.ids.insert(generator_.ids.begin(),
-                            finalPaths_[i][0].ids.begin(),
-                            finalPaths_[i][0].ids.end());
-      starts[i + 1] = starts[i] + finalPaths_[i][0].ids.size();
+      Path& path = finalPaths_[i][0];
+      generator_.ids.insert(
+          generator_.ids.begin(), path.ids.begin(), path.ids.end());
+      starts[i + 1] = starts[i] + path.ids.size();
     }
   }
 }
@@ -1364,25 +1359,70 @@ void RecurrentGradientMachine::copyDataOutlinkFrame(size_t machineCur) {
   }
 }
 
-void RecurrentGradientMachine::createDataOutlink(
-    std::vector<int>& machineIdVec) {
-  size_t seqNum =
-      getBeamSize() > 1UL ? finalPaths_.size() : finalPaths_[0].size();
-  std::vector<int> starts(seqNum + 1, 0);
-  for (size_t i = 0; i < seqNum; ++i) {
-    size_t seqLen = getBeamSize() > 1UL ? finalPaths_[i][0].ids.size()
-                                        : finalPaths_[0][i].ids.size();
-    starts[i + 1] = starts[i] + seqLen;
+void RecurrentGradientMachine::createDataOutlinkSelRowsInfo(
+    bool isSeq, std::vector<Argument>& outArgs) {
+  batchMachineIdVec_.clear();
+
+  size_t seqIdx = 0;
+  for (size_t i = 0; i < finalPaths_.size(); ++i) {
+    for (size_t j = 0; j < finalPaths_[i].size(); ++j) {
+      std::vector<int>& machineIdVec = finalPaths_[i][j].machineIdVec;
+      if (isSeq) {
+        for (size_t i = 0; i < machineIdVec.size(); ++i) {
+          size_t rowId = machineIdVec[i];
+          int* seqPos =
+              outArgs[i].sequenceStartPositions->getMutableData(false);
+          batchMachineIdVec_.push_back(seqPos[rowId]);
+        }
+      } else {
+        batchMachineIdVec_.insert(
+            batchMachineIdVec_.end(), machineIdVec.begin(), machineIdVec.end());
+      }
+      seqIdx++;
+    }
+  }
+}
+
+void RecurrentGradientMachine::createDataOutlinkCopySizeInfo(
+    bool isSeq, std::vector<Argument>& outArgs, std::vector<int>& copySize) {
+  size_t totalSeqNum = std::accumulate(
+      finalPaths_.begin(),
+      finalPaths_.end(),
+      0UL,
+      [](size_t a, const std::vector<Path>& b) { return a + b.size(); });
+  copySize.resize(totalSeqNum, 1);
+
+  batchMachineStartPos_.resize(totalSeqNum + 1, 0);
+  if (isSeq) {
+    ICpuGpuVectorPtr inputSeqStartPos = outArgs[0].sequenceStartPositions;
+    CHECK_EQ(inputSeqStartPos->getSize() - 1, finalPaths_.size());
+    int* starts = inputSeqStartPos->getMutableData(false);
+    int seqId = 0;
+    for (int i = 0; i < finalPaths_.size(); ++i) {
+      for (int j = 0; j < finalPaths_[i].size(); ++j) {
+        copySize[seqId] = starts[i + 1] - starts[i];
+        batchMachineStartPos_[seqId + 1] =
+            batchMachineStartPos_[seqId] + finalPaths_[i][j].ids.size();
+        seqId++;
+      }
+    }
   }
+}
 
+void RecurrentGradientMachine::createDataOutlink() {
   for (size_t i = 0; i < dataArgsSize_; i++) {
+    bool isSeq = dataArgsFrame_[i][0].hasSeq();
+    std::vector<int> copySize;
+    createDataOutlinkCopySizeInfo(isSeq, dataArgsFrame_[i], copySize);
+    createDataOutlinkSelRowsInfo(isSeq, dataArgsFrame_[i]);
+
     dataArgs_[i].concat(dataArgsFrame_[i],
-                        machineIdVec,
-                        starts,
+                        batchMachineIdVec_,
+                        batchMachineStartPos_,
+                        copySize,
                         useGpu_,
                         HPPL_STREAM_1,
                         PASS_TEST);
-
     auto dataAgent =
         dynamic_cast<DataLayer*>(outFrameLines_[i + 1].agentLayer.get());
     CHECK_NOTNULL(dataAgent);

paddle/gserver/gradientmachines/RecurrentGradientMachine.h

@@ -480,7 +480,11 @@ private:
    * @param machineIdVec : select a row of output matrix in each frame
    * that the generation process expanded.
    */
-  void createDataOutlink(std::vector<int>& machineIdVec);
+  void createDataOutlink();
+  void createDataOutlinkCopySizeInfo(bool isSeq,
+                                     std::vector<Argument>& outArgs,
+                                     std::vector<int>& copySize);
+  void createDataOutlinkSelRowsInfo(bool isSeq, std::vector<Argument>& outArgs);
 
   /*
    * @brief used in beam search, connect previous frame to form recurrent link
@@ -543,6 +547,7 @@ private:
   std::vector<int> topIds_;
   std::vector<int> seqIds_;
   std::vector<int> batchMachineIdVec_;
+  std::vector<int> batchMachineStartPos_;
   std::vector<std::vector<Path>> finalPaths_;
   std::vector<real> minFinalPathLogProb_;
   BeamSearchControlCallbacks* beamSearchCtrlCallbacks_;

paddle/parameter/Argument.cpp

@@ -276,17 +276,21 @@ int32_t Argument::resizeAndCopyFrom(const Argument& src,
 void Argument::concat(const std::vector<Argument>& args,
                       const std::vector<int>& selectRows,
                       const std::vector<int>& seqStartPos,
+                      const std::vector<int>& copySize,
                       bool useGpu,
                       hl_stream_t stream,
                       PassType passType) {
   CHECK(!subSequenceStartPositions)
       << "undefined behavior for subsequence positions";
 
-  size_t batchSize = selectRows.size();
+  size_t batchSize = 0;
+  for (size_t i = 0; i < copySize.size(); ++i)
+    batchSize += copySize[i] * (seqStartPos[i + 1] - seqStartPos[i]);
+
   auto copyArg = [batchSize, stream](MatrixPtr& dst,
                                      MatrixPtr src,
-                                     int startRow,
-                                     int pos,
+                                     int desStartRow,
+                                     int srcStartRow,
                                      int size,
                                      bool useGpu) {
     if (!src) {
@@ -300,8 +304,8 @@ void Argument::concat(const std::vector<Argument>& args,
       dst->resize(batchSize, width);
     }
 
-    MatrixPtr tmpMatrix = dst->subMatrix(startRow, size);
-    tmpMatrix->copyFrom(*src->subMatrix(pos, size), stream);
+    MatrixPtr tmpMatrix = dst->subMatrix(desStartRow, size);
+    tmpMatrix->copyFrom(*src->subMatrix(srcStartRow, size), stream);
   };
 
   auto copyIds = [batchSize, stream](IVectorPtr& dst,
@@ -339,24 +343,24 @@ void Argument::concat(const std::vector<Argument>& args,
 
   dataId = args[0].dataId;
   CHECK_NE(seqStartPos.size(), 0UL);
-  size_t sampleNum = seqStartPos.size() - 1;
-  for (size_t i = 0; i < sampleNum; ++i) {
+  int desStartRow = 0;
+  for (size_t i = 0; i < copySize.size(); ++i) {
     int startPos = seqStartPos[i];
     int endPos = seqStartPos[i + 1];
     CHECK_GE(args.size(), static_cast<size_t>(endPos - startPos));
     for (int j = startPos; j < endPos; ++j) {
       const Argument& arg = args[j - startPos];
-      CHECK_EQ(arg.dataId, dataId) << "Arguments in concat should have"
-                                   << " same dataId";
-      const int copySize = 1;
-      const int rowIdx = selectRows[j];
-      copyArg(in, arg.in, j, rowIdx, copySize, useGpu);
-      copyArg(value, arg.value, j, rowIdx, copySize, useGpu);
+      CHECK_EQ(arg.dataId, dataId) << "Arguments in concat should have the "
+                                   << "same dataId";
+      const int srcStartRow = selectRows[j];
+      copyArg(in, arg.in, desStartRow, srcStartRow, copySize[i], useGpu);
+      copyArg(value, arg.value, desStartRow, srcStartRow, copySize[i], useGpu);
       if (passType != PASS_TEST) {
-        copyArg(grad, arg.grad, j, rowIdx, copySize, useGpu);
+        copyArg(grad, arg.grad, desStartRow, srcStartRow, copySize[i], useGpu);
       }
-      copyIds(ids, arg.ids, j, rowIdx, copySize, useGpu);
-      copyStrs(strs, arg.strs, j, rowIdx, copySize, useGpu);
+      copyIds(ids, arg.ids, desStartRow, srcStartRow, copySize[i], useGpu);
+      copyStrs(strs, arg.strs, desStartRow, srcStartRow, copySize[i], useGpu);
+      desStartRow += copySize[i];
     }
   }
   ICpuGpuVector::resizeOrCreate(

paddle/parameter/Argument.h

@@ -240,6 +240,7 @@ struct Argument {
   void concat(const std::vector<Argument>& args,
               const std::vector<int>& selectRows,
               const std::vector<int>& seqStartPos,
+              const std::vector<int>& copySize,
               bool useGpu,
               hl_stream_t stream,
               PassType passType);

python/paddle/trainer_config_helpers/networks.py

@@ -1370,14 +1370,7 @@ def simple_attention(encoded_sequence,
         param_attr=softmax_param_attr,
         name="%s_softmax" % name,
         bias_attr=False)
-
-    scaled = scaling_layer(
-        weight=attention_weight,
-        input=encoded_sequence,
-        name='%s_scaling' % name)
-
-    return pooling_layer(
-        input=scaled, pooling_type=SumPooling(), name="%s_pooling" % name)
+    return attention_weight
 
 
 def inputs(layers, *args):
@@ -1395,7 +1388,7 @@ def inputs(layers, *args):
     if len(args) != 0:
         layers.extend(args)
 
-    Inputs(* [l.name for l in layers])
+    Inputs(*[l.name for l in layers])
 
 
 def outputs(layers, *args):
@@ -1438,7 +1431,7 @@ def outputs(layers, *args):
     assert len(layers) > 0
 
     if HasInputsSet():  # input already set
-        Outputs(* [l.name for l in layers])
+        Outputs(*[l.name for l in layers])
         return  # just return outputs.
 
     if len(layers) != 1: