fix the bug in dimension setting in the back propagation of Merge

source/network/XBackwardShape.cpp

@@ -71,9 +71,11 @@ dE/da = split(dE/dc)
 void XShapeGrad::GradMerge(XTensor * node)
 {
     XLink &income = node->income;
-    CheckNTErrors(income.tailNum == 0, "Wrong input tensor number for MERGE!");
-
     XTensor * input = income.tails[0];
+
+    CheckNTErrors(income.tailNum == 1, "Wrong input tensor number for MERGE!");
+    CheckNTErrors(node->order == input->order - 1, "wrong tensor orders!");
+
     int whereToMerge = income.GetParamInt(0);
     int leadDim = income.GetParamInt(1);
 
@@ -99,7 +101,7 @@ void XShapeGrad::GradMerge(XTensor * node)
                            input->devID, input->mem);
 
     dims[whereToMerge - leadDim] *= dims[0];
-    XTensor gradNodeSmall(node->order - leadDim, dims,
+    XTensor gradNodeSmall(node->order - leadDim, dims + leadDim + 1,
                           node->dataType, node->denseRatio, 
                           node->devID, node->mem);
 
@@ -109,7 +111,7 @@ void XShapeGrad::GradMerge(XTensor * node)
         for(int i = 0; i < blockNum; i++){
             gradNodeSmall.data = (char*)node->grad->data + i * blockSize;
             gradInputSmall.data = (char*)input->grad->data + i * blockSize;
-            _Split(&gradNodeSmall, &gradInputSmall, whereToMerge - leadDim, input->dimSize[leadDim]);
+            _Split(&gradNodeSmall, &gradInputSmall, whereToMerge - leadDim - 1, input->dimSize[leadDim]);
         }
     }
 
@@ -123,7 +125,7 @@ void XShapeGrad::GradMerge(XTensor * node)
         for(int i = 0; i < blockNum; i++){
             gradNodeSmall.data = (char*)node->grad->data + i * blockSize;
             gradInputSmall.data = (char*)input->grad->data + i * blockSize;
-            _Split(&gradNodeSmall, &gradInputSmallBuf, whereToMerge - leadDim, input->dimSize[leadDim]);
+            _Split(&gradNodeSmall, &gradInputSmallBuf, whereToMerge - leadDim - 1, input->dimSize[leadDim]);
             _Sum(&gradInputSmall, &gradInputSmallBuf, &gradInputSmall);
         }
     }

source/sample/fnnlm/FNNLM.cpp

@@ -73,8 +73,7 @@ void MakeWordBatch(XTensor &batch, NGram * ngrams, int ngramNum, int n, int vSiz
 void Forward(XTensor inputs[], XTensor &output, FNNModel &model, FNNNet &net);
 void Backward(XTensor inputs[], XTensor &output, XTensor &gold, LOSS_FUNCTION_NAME loss, 
               FNNModel &model, FNNModel &grad, FNNNet &net);
-void FBInOne(XTensor inputs[], XTensor &output, XTensor &gold, LOSS_FUNCTION_NAME loss, 
-             FNNModel &model, XNet &net);
+void ForwardAutoDiff(XTensor inputs[], XTensor &output, FNNModel &model);
 
 /* 
 entry of the program 
@@ -415,7 +414,10 @@ void Train(const char * train, bool isShuffled, FNNModel &model)
             }
             else{
                 /* forward + backward process */
-                FBInOne(inputs, output, gold, CROSSENTROPY, model, autoDiffer);
+                ForwardAutoDiff(inputs, output, model);
+
+                /* automatic differentiation */
+                autoDiffer.Backward(output, gold, CROSSENTROPY);
 
                 /* update model parameters */
                 Update(model, grad, learningRate, true);
@@ -902,17 +904,14 @@ void Backward(XTensor inputs[], XTensor &output, XTensor &gold, LOSS_FUNCTION_NA
 }
 
 /*
-forward + backward in one procedure
+forward process (with tensor connections)
 >> inputs - input word representations
 >> output - output probability
->> gold - gold standard
->> loss - loss function name
 >> model - the fnn model
 */
-void FBInOne(XTensor inputs[], XTensor &output, XTensor &gold, 
-             LOSS_FUNCTION_NAME loss, FNNModel &model, XNet &net)
+void ForwardAutoDiff(XTensor inputs[], XTensor &output, FNNModel &model)
 {
-    int batchSize = gold.GetDim(0);
+    int batchSize = inputs[0].GetDim(0);
     int n = model.n;
     int depth = model.hDepth;
 
@@ -945,9 +944,6 @@ void FBInOne(XTensor inputs[], XTensor &output, XTensor &gold,
 
     /* output layer */
     output = LogSoftmax(MMul(hidden, model.outputW) + b, 1);
-
-    /* automatic differentiation */
-    net.Backward(output);
 }
 
 /* 

source/tensor/core/shape/Split.cpp

@@ -301,7 +301,7 @@ void Split(const XTensor &big, XList &smalls, int whereToSplit, int splitNum)
         XLink::AddParamToHeadInt(s, whereToSplit);
 
         /* it is tricky here that we keep the id of each 
-           block, rather than the total number of splits */
+           block, rather than the total number of the splits */
         XLink::AddParamToHeadInt(s, i);
     }
 }