debugging

source/sample/transformer/T2TPredictor.cpp

@@ -177,7 +177,7 @@ void T2TPredictor::Predict(T2TStateBundle * next, XTensor * encoding, XTensor *
     else{
         inputDec = GeneratePaths(s);
         inputDec.SetDevice(inputEnc->devID, inputEnc->mem);
-        inputDec = Concatenate(inputDec, first, inputDec.order - 1);
+        inputDec = Concatenate(first, inputDec, inputDec.order - 1);
     }
 
     /* prediction probabilities */
@@ -219,7 +219,6 @@ void T2TPredictor::Predict(T2TStateBundle * next, XTensor * encoding, XTensor *
     InitTensor1D(&selectTgt, 1, X_INT);
 
     selectSrc.SetInt(stride - 1, 0);
-
     selectTgt.SetInt(0, 0);
 
     selectSrc.SetDevice(decoding.devID, decoding.mem);

source/sample/transformer/T2TSearch.cpp

@@ -372,14 +372,15 @@ void T2TSearch::Expand(T2TStateBundle * prev, T2TStateBundle * beam)
     CheckNTErrors(beam->stateNum == id.unitNum, "Errors occur in counting!");
     
     /* Related variables are kept on the states of the graph. All these are 
-       maintained on CPUs to ease the implementation of requent access and 
+       maintained on CPUs to ease the implementation of frequent access and 
        modification of the states. An alternative is to do this on GPUs but 
        it needs much more coding work and the speed-up is not obvious. */
-    for(int i = 0; i < beam->stateNum; i++){
-        T2TState & state = states[i];
+    for(int i = 0; i < beam->stateNum; i += beamSize){
+        for(int j = 0; j < beamSize; j++){
+            T2TState & state = states[i + j];
 
-        int offset = id.GetInt(i);
-        T2TState * last = prev->states + offset;
+            int offset = id.GetInt(i + j);
+            T2TState * last = prev->states + i * beamSize + offset;
 
             CheckNTErrors(offset >= 0, "Wrong state index!");
         
@@ -412,6 +413,7 @@ void T2TSearch::Expand(T2TStateBundle * prev, T2TStateBundle * beam)
             /* set the ending mark */
             endMarkCPU.SetInt(state.isEnd, i);
         }
+    }
 
     /* copy the ending mark from CPU to the target device */
     CopyValues(endMarkCPU, endMark);