length penalty

source/sample/transformer/T2TLengthPenalty.cpp

@@ -27,15 +27,20 @@ namespace transformer
 GNMT-like length penalty: pl = ((5 + n)/(5 + 1))^\alpha 
 where n = length of the sequence 
 >> length - length of the sequence (for each entry)
->> lp - length penaltyof the sequence (for each entry)
+>> alpha - the parameter controls the length preference
+<< return - length penaltyof the sequence (for each entry)
 */
-void T2TLengthPenalizer::GNMT(const XTensor & length, XTensor & lp, float alpha)
+XTensor T2TLengthPenalizer::GNMT(const XTensor & length, float alpha)
 {
     XTensor base;
+    XTensor lp;
 
-    base = ScaleAndShift(ScaleAndShift(length, 0, 5.0F), 1.0F/(5 + 1));
+    //base = ScaleAndShift(ScaleAndShift(length, 0, 5.0F), 1.0F/(5 + 1));
+    base = (length + 5)/(1 + 5);
 
     lp = Power(base, alpha);
+    
+    return lp;
 }
 
 }

source/sample/transformer/T2TLengthPenalty.h

@@ -40,7 +40,7 @@ public:
     /* GNMT-like length penalty: pl = ((5 + n)/(5 + 1))^\alpha 
        where n = length of the sequence */
     static
-    void GNMT(const XTensor & length, XTensor & lp, float alpha);
+    XTensor GNMT(const XTensor & length, float alpha);
 };
 
 }

source/sample/transformer/T2TSearch.cpp

@@ -111,7 +111,7 @@ void T2TSearch::Score(T2TStateBundle * prev, T2TStateBundle * beam)
     _ScaleAndShift(&lenPrev, &len, 1.0F, 1.0F);
 
     /* the GNMT-like length penalty */
-    T2TLengthPenalizer::GNMT(len, lp, alpha);
+    lp = T2TLengthPenalizer::GNMT(len, alpha);
 
     /* score = log-prob/lp */
     _Div(&score, &lp, &score);
@@ -188,17 +188,43 @@ expand the search graph
 */
 void T2TSearch::Expand(T2TStateBundle * prev, T2TStateBundle * beam)
 {
+    CheckNTErrors(beam->prediction.unitNum == beam->preID.unitNum, "A problem occurs in the beam!");
+    
     beam->MakeStates(beam->prediction.unitNum);
 
     T2TState * states = beam->states;
-    XTensor &predict = beam->prediction;
-    
-    XTensor index = *NewTensorBuf(predict.order - 1, predict.dimSize, X_FLOAT, 1.0F, 
-                                  predict.devID, predict.mem);
-
-    index.SetAscendingOrder(-1);
-
-    DelTensorBuf(&index);
+    XTensor & idRef = beam->preID;
+    XTensor & modelScoreRef = beam->modelScore;
+    XTensor & probRef = beam->prob;
+    XTensor & probPathRef = beam->probPath;
+    XTensor   id;
+    XTensor   modelScore;
+    XTensor   prob;
+    XTensor   probPath;
+    
+    InitTensorOnCPU(&id, &idRef);
+    InitTensorOnCPU(&modelScore, &modelScoreRef);
+    InitTensorOnCPU(&prob, &probRef);
+    InitTensorOnCPU(&probPath, &probPathRef);
+    
+    /* we copy the data to CPU because the frequent access to GPU is slow
+       and we can speed-up the process by doing the job on CPU. */
+    CopyValues(idRef, id);
+    CopyValues(modelScoreRef, modelScore);
+    CopyValues(prob, probRef);
+    CopyValues(probPathRef, probPath);
+    
+    for(int i = 0; i < id.unitNum; i++){
+        T2TState & state = states[i];
+        
+        /* pointer to the previous state */
+        state.last = prev->states + id.GetInt(i);
+        
+        /* scores */
+        state.modelScore = modelScore.Get(i);
+        state.prob = prob.Get(i);
+        state.probPath = probPath.Get(i);
+    }
 }
 
 /*