save log probabilty for every generated words.

paddle/gserver/gradientmachines/RecurrentGradientMachine.cpp

@@ -967,8 +967,9 @@ void RecurrentGradientMachine::generateSequence() {
   size_t numSequences = getGenBatchSize();
 
   resizeBootFrame(numSequences);
-  // We create only two sub-network in generation for alternate use.
+  // We create only two sub-network in generation, one stores states of all
-  // Thus, we can reduce total memory of output_ in layer forward.
+  // layers in previous time step and the other storing the states at current
+  // time step.
   resizeOrCreateFrames(2);
 
   // outFrameLines_.size() > 1UL
@@ -1001,10 +1002,9 @@ void RecurrentGradientMachine::generateSequence() {
 
   // init outArg
   size_t resultNum = generator_.config.num_results_per_sample();
-  IVector::resizeOrCreate(
+  size_t maxGenWordCount =
-      generator_.outArg.ids,
+      generator_.config.max_num_frames() * numSequences * resultNum;
-      generator_.config.max_num_frames() * numSequences * resultNum,
+  IVector::resizeOrCreate(generator_.outArg.ids, maxGenWordCount, false);
-      false);
   if (resultNum > 1) {
     CHECK_LE(resultNum, static_cast<size_t>(generator_.config.beam_size()));
     Matrix::resizeOrCreate(generator_.outArg.in,
@@ -1012,6 +1012,11 @@ 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,
@@ -1313,13 +1318,20 @@ void RecurrentGradientMachine::fillGenOutputs() {
   starts[0] = 0;
   if (numResults > 1) {
     real* probs = generator_.outArg.in->getData();
+    real* idsProb = generator_.outArg.value->getData();
+    size_t curPos = 0;
     for (size_t i = 0; i < finalPaths_.size(); ++i) {
       for (size_t j = 0; j < finalPaths_[i].size(); ++j) {
         Path& path = finalPaths_[i][j];
-        generator_.ids.push_back(path.ids.size());  // sequence size
+        size_t genLen = path.ids.size();
+        generator_.ids.push_back(genLen);  // sequence size
         generator_.ids.insert(
             generator_.ids.end(), path.ids.begin(), path.ids.end());
         generator_.ids.push_back(-1);  // end of sequence
+
+        memcpy(idsProb + curPos, path.idsProb.data(), sizeof(real) * genLen);
+        curPos += genLen;
+        idsProb[curPos++] = -1.0;
         probs[i * numResults + j] = path.logProb;
 
         if (!j && dataArgsSize_) {

paddle/gserver/gradientmachines/RecurrentGradientMachine.h

@@ -189,6 +189,11 @@ public:
      */
     std::vector<int> ids;
 
+    /**
+     * @brief idsProb, log probability of each generated words.
+     */
+    std::vector<real> idsProb;
+
     /**
      * @brief logProb, current probability of path.
      */
@@ -228,11 +233,13 @@ public:
      */
     Path(Path& old, int newId, real logProb, int machineId, int topIndex)
         : ids(old.ids),
+          idsProb(old.idsProb),
           logProb(old.logProb + logProb),
           machineId(machineId),
           topIndex(topIndex),
           seqId(old.seqId) {
       ids.push_back(newId);
+      idsProb.push_back(logProb);
       if (!old.probHistory.empty()) {
         this->probHistory = old.probHistory;
         // probHistory store current prob, not sum
@@ -411,8 +418,9 @@ protected:
 
   struct Generator {
     GeneratorConfig config;
-    std::vector<int> ids;  // store generated sequences
+    std::vector<int> ids;       // store generated sequences
-    Argument outArg;       // final output argument
+    std::vector<real> idsProb;  // log probability of each generated word
+    Argument outArg;            // final output argument
   };
   bool generating_;
   Generator generator_;