hypothesis id

source/sample/transformer/T2TSearch.cpp

@@ -126,12 +126,19 @@ void T2TSearch::Generate(T2TStateBundle * beam)
     
     
     CopyValues(scoreTopK, preID);
     CopyValues(scoreTopK, preID);
     
     
-    int sizePredict = score.GetDim(-1);
-    
-    /* pre id !!! */
-    Descale(preID, sizePredict);
-    
-    /* mod !!! */
+    int sizeVocab = score.GetDim(-1);
+    
+    /* "preID" represents the id (or the offset) of previous state used to make the current
+       hypothesis. Note that we reshape the "score" tensor into a matrix where each
+       row means a previous state. The column number is size-of-beam * vocab-size. We,
+       therefore, divide the top-k index by vocab-size to compute the id of previous state for
+       each hypothesis in the top-k list. */
+    Descale(preID, sizeVocab);
+    
+    /* Then we are going to do something similar to "preID". For the top-k predictions, we
+       need to know their indices in the vocabulary. We compute the offset of each prediction
+       in the vocabulary by dividing it with vocab-size and computing the remainder. */
+    Mod(index, sizeVocab);
 
 
     score.Reshape(order, dims);
     score.Reshape(order, dims);
 }
 }