add padding on the encoder side for t2t MT

source/sample/transformer/T2TModel.cpp

@@ -174,10 +174,10 @@ make the network for machine translation (with the output softmax layer)
 >> inputEnc - input tensor of the encoder
 >> inputDec - input tensor of the decoder
 >> output - output tensor (distribution)
->> padding - padding of the sequences
+>> paddingEnc - padding of the sequences (on the encoder side)
 >> isTraining - indicates whether the model is for training
 */
-void T2TModel::MakeMT(XTensor &inputEnc, XTensor &inputDec, XTensor &output, XTensor &padding, bool isTraining)
+void T2TModel::MakeMT(XTensor &inputEnc, XTensor &inputDec, XTensor &output, XTensor &paddingEnc, bool isTraining)
 {
     XTensor encoding;
     XTensor decoding;
@@ -199,11 +199,44 @@ void T2TModel::MakeMT(XTensor &inputEnc, XTensor &inputDec, XTensor &output, XTe
     _SetDataLowTri(&maskDec, 1e9F, 0);
     _ScaleAndShiftMe(&maskDec, 1.0F, -1e9F);
 
+    /* padding on the source side */
+    int * dimsPadding = new int[paddingEnc.order + 2];
+    for (int i = 0; i < paddingEnc.order - 1; i++)
+        dimsPadding[i] = paddingEnc.GetDim(i);
+    dimsPadding[paddingEnc.order - 1] = paddingEnc.GetDim(-1);
+    dimsPadding[paddingEnc.order] = paddingEnc.GetDim(-1);
+
+    XTensor * padding2 = NewTensorBuf(paddingEnc.order + 1, dimsPadding, paddingEnc.dataType,
+                                      paddingEnc.denseRatio, paddingEnc.devID, paddingEnc.mem);
+
+    for (int i = 0; i < padding2->order; i++)
+        dimsPadding[i + 1] = padding2->GetDim(i);
+    dimsPadding[0] = nhead;
+
+    XTensor * padding3 = NewTensorBuf(paddingEnc.order + 2, dimsPadding, paddingEnc.dataType,
+                                      paddingEnc.denseRatio, paddingEnc.devID, paddingEnc.mem);
+
+    /* mask of the padding */
+    _Unsqueeze(&paddingEnc, padding2, paddingEnc.order - 1, paddingEnc.GetDim(-1));
+    _Unsqueeze(padding2, padding3, 0, nhead);
+
+    _ScaleAndShiftMe(padding3, 1e9F, -1e9F);
+
+    InitTensor(&maskEnc, padding3);
+    maskEnc.SetZeroAll();
+
+    /* generate the mask on the source language side (for padding) */
+    _Sum(&maskEnc, padding3, &maskEnc);
+
     encoding = MakeEncoder(inputEnc, maskEnc, isTraining);
     decoding = MakeDecoder(inputDec, encoding, maskDec, isTraining);
     outputLayer.Make(decoding, output);
 
     delete[] dims;
+    delete[] dimsPadding;
+
+    DelTensorBuf(padding2);
+    DelTensorBuf(padding3);
 }
 
 /* 

source/sample/transformer/T2TModel.h

@@ -78,7 +78,7 @@ public:
     void MakeLM(XTensor &input, XTensor &output, XTensor &padding, bool isTraining);
 
     /* make the network for machine translation (with the output softmax layer) */
-    void MakeMT(XTensor &inputEnc, XTensor &inputDec, XTensor &output, XTensor &padding, bool isTraining);
+    void MakeMT(XTensor &inputEnc, XTensor &inputDec, XTensor &output, XTensor &paddingEnc, bool isTraining);
 
     /* get parameter matrics */
     void GetParams(XList &list);

source/sample/transformer/T2TTrainer.h

@@ -166,7 +166,8 @@ public:
 
     /* load a batch of sequences */
     int LoadBatch(FILE * file, bool isLM,
-                  XTensor * batch, XTensor * padding, XTensor * output, 
+                  XTensor * batchEnc, XTensor * paddingEnc, 
+                  XTensor * batchDec, XTensor * paddingDec,
                   int * seqs,
                   int vsEnc, int vsDec, int sBatch, int wBatch, 
                   bool isSorted, int &wCount,
@@ -174,14 +175,16 @@ public:
 
     /* load a batch of sequences (for language modeling) */
     int LoadBatchLM(FILE * file, 
-                    XTensor * batch, XTensor * padding, XTensor * output, 
+                    XTensor * batchEnc, XTensor * paddingEnc,
+                    XTensor * batchDec, XTensor * paddingDec,
                     int * seqs, int vs, int sBatch, int wBatch, 
                     bool isSorted, int &wCount,
                     int devID, XMem * mem);
 
     /* load a batch of sequences (for machine translation) */
     int LoadBatchMT(FILE * file, 
-                    XTensor * batch, XTensor * padding, XTensor * output, 
+                    XTensor * batchEnc, XTensor * paddingEnc, 
+                    XTensor * batchDec, XTensor * paddingDec,
                     int * seqs, int vsEnc, int vsDec, int sBatch, int wBatch, 
                     bool isSorted, int &wCount,
                     int devID, XMem * mem);