bug fixes

source/sample/transformer/T2TAttention.cpp

@@ -116,12 +116,22 @@ XTensor T2TAttention::Make(XTensor &k, XTensor &q, XTensor &v, XTensor &mask)
     XTensor att;
     XTensor dot;
     XTensor scalar;
-    
+
     /* scalar = softmax(Q * K^T / sqrt(dk)) * V */
     dot = BMMul(qheads, X_NOTRANS, kheads, X_TRANS);
+
     if(isMasked)
         dot = dot + mask;
-    scalar = Softmax(Linear(dot, 1.0F/(float)sqrt((float)dk)), -1);
+
+    dot = Linear(dot, 1.0F/(float)sqrt((float)dk));
+
+    //if(llnum == 1)
+    //    dot.Dump(tf, "dot:");
+
+    scalar = Softmax(dot, -1);
+
+    //if(llnum == 1)
+    //    scalar.Dump(tf, "scalar:");
 
     //if(ignored > 0)
     //    _SetDataDim(&scalar, 0, ignored, scalar.order - 2, 1e-9F);

source/sample/transformer/T2TEncoder.cpp

@@ -103,6 +103,8 @@ XTensor AttEncoder::Make(XTensor &input, XTensor &mask, bool skipInputRes)
         XTensor fnn;
         XTensor res;
 
+        llnum = -1;
+
         /* we skip the residual connection for the first layer if
            the encoder is used in language modeling. */
         if(skipInputRes && i == 0){
@@ -115,6 +117,11 @@ XTensor AttEncoder::Make(XTensor &input, XTensor &mask, bool skipInputRes)
             x = attLayerNorms[i].Make(att); 
         }
         else{
+            //if(i == 1)
+            //    x.Dump(tf, "x:");
+            //if(i == 1)
+            //    llnum = 1;
+
             /* self attention */
             att = attentions[i].Make(x, x, x, mask);
 
@@ -125,6 +132,8 @@ XTensor AttEncoder::Make(XTensor &input, XTensor &mask, bool skipInputRes)
 
             /* layer normalization */
             x = attLayerNorms[i].Make(res);
+
+            llnum = -1;
         }
 
         /* fnn */

source/sample/transformer/T2TLayerNormal.cpp

@@ -27,7 +27,7 @@
 
 namespace transformer
 {
-    
+
 /* constructor */
 T2TLN::T2TLN()
 {

source/sample/transformer/T2TModel.cpp

@@ -130,7 +130,7 @@ void T2TModel::Make(XTensor &input, XTensor &output, XTensor &padding)
         
         _ScaleAndShiftMe(padding3, 1e9F, -1e9F);
         
-        _Sum(&mask, padding3, &mask);
+        //_Sum(&mask, padding3, &mask);
 
         encoding = MakeEncoding(input, mask, true);
         outputLayer.Make(encoding, output);

source/sample/transformer/T2TTrainer.cpp

@@ -90,7 +90,6 @@ void T2TTrainer::Init(int argc, const char ** argv)
 
 }
 
-FILE * tf = NULL;
 int tc = 0;
 
 /* 
@@ -257,7 +256,7 @@ void T2TTrainer::Test(const char * fn, const char * ofn, T2TModel * model)
     
     ClearBuf();
 
-    while(LoadBatch(file, true, &batch, &padding, &gold, seqs, 1, vSize, 1, 512, isLenSorted, wc, devID, mem)){
+    while(LoadBatch(file, true, &batch, &padding, &gold, seqs, 1, vSize, 1, 1, isLenSorted, wc, devID, mem)){
 
         CheckNTErrors(batch.order == 3, "wrong tensor order of the sequence batch");
             
@@ -503,11 +502,11 @@ int T2TTrainer::LoadBatch(FILE * file, bool isLM,
                 if(w == seqLen[s] - 1)
                     output->Set3D(1.0F, s - seq, w, buf[seqOffset[s] + w]);
                 wCount++;
-                //fprintf(tf, "%d", buf[seqOffset[s] + w]);
-                //if(w < seqLen[s] - 1)
-                //    fprintf(tf, " ");
-                //else
-                //    fprintf(tf, "\n");
+                /*fprintf(tf, "%d", buf[seqOffset[s] + w]);
+                if(w < seqLen[s] - 1)
+                    fprintf(tf, " ");
+                else
+                    fprintf(tf, "\n");*/
                 if(seqs != NULL)
                     seqs[seqSize++] = buf[seqOffset[s] + w];
             }

source/sample/transformer/T2TUtility.cpp

@@ -27,6 +27,8 @@ namespace transformer
 {
 
 FILE * tmpFILE;
+int llnum = 0;
+FILE * tf = NULL;
 
 void LoadParamString(int argc, const char ** argv, const char * name, char * p, const char * defaultP)
 {

source/sample/transformer/T2TUtility.h

@@ -38,6 +38,9 @@ void LoadParamFloat(int argc, const char ** argv, const char * name, float * p,
 /* show arguments */
 void ShowParams(int argc, const char ** argv);
 
+extern int llnum;
+extern FILE * tf;
+
 }
 
 #endif

source/tensor/XTensor.cpp

@@ -1377,9 +1377,10 @@ dump data to a file
 >> file - where to domp the data
 >> label - label of the tensor
 >> n - number of items to dump
+>> beg - the first item id
 >> verbose - verbose level
 */
-void XTensor::Dump(FILE * file, const char * label, const int n, const int verbose)
+void XTensor::Dump(FILE * file, const char * label, const int n, const int beg, const int verbose)
 {
     if (verbose > verboseLevel)
         return;
@@ -1437,28 +1438,26 @@ void XTensor::Dump(FILE * file, const char * label, const int n, const int verbo
     }
     if (!isSparse) {
         if (dataType == DEFAULT_DTYPE) {
-            if (unitNum > 0) {
-                DTYPE f = *(DTYPE*)d;
-                fprintf(file, "%e", f);
-            }
-            int num = unitNum;
-            if (n > 0)
-                num = MIN(num, n);
-            for (int i = 1; i < num; i++) {
-                DTYPE * f = ((DTYPE*)d) + i;
-                fprintf(file, " %e", *f);
+            int end = MIN(n > 0 ? beg + n : beg + unitNum, unitNum);
+            for(int i = beg; i < end; i++){
+                DTYPE f = ((DTYPE*)d)[i];
+                if(i == beg)
+                    fprintf(file, "%e", f);
+                else
+                    fprintf(file, " %e", f);
+
             }
         }
         else {
-            ShowNTErrors("Cannot dump the tensor to the file in non-float values!");
+            ShowNTErrors("TODO!");
         }
     }
     else {
         int num = this->unitNumNonZero > 0 ? *(int*)d : 0;
-        if (n > 0)
-            num = MIN(num, n);
+        if (beg + n > 0)
+            num = MIN(num, beg + n);
         fprintf(file, "%d ", num);
-        for (int i = 0; i < num; i++) {
+        for (int i = beg; i < num; i++) {
             int key = GetKeyInSparse(i);
             DTYPE value = GetInSparse(i);
             fprintf(file, "[%d]%e ", key, value);
@@ -1481,13 +1480,14 @@ dump data to a file
 >> file - where to domp the data
 >> label - label of the tensor
 >> n - number of items to dump
+>> beg - the first item id
 >> verbose - verbose level
 */
-void XTensor::Dump(const XTensor * tensor, FILE * file, const char * label, const int n, const int verbose)
+void XTensor::Dump(const XTensor * tensor, FILE * file, const char * label, const int n, const int beg, const int verbose)
 {
     XTensor a(tensor->order, tensor->dimSize, tensor->dataType, tensor->denseRatio, tensor->devID, tensor->mem);
     _CopyValues(tensor, &a);
-    a.Dump(file, label, n, verbose);
+    a.Dump(file, label, n, beg, verbose);
 }
 
 /* 

source/tensor/XTensor.h

@@ -339,11 +339,11 @@ public:
     bool BinarySearch(int key, DTYPE &value, void * &position) const;
 
     /* dump data to a file */
-    void Dump(FILE * file, const char * label = NULL, const int n = -1, const int verbose = 0);
+    void Dump(FILE * file, const char * label = NULL, const int n = -1, const int beg = 0, const int verbose = 0);
 
     /* dump data to a file */
     static
-    void Dump(const XTensor * tensor, FILE * file, const char * label = NULL, const int n = -1, const int verbose = 0);
+    void Dump(const XTensor * tensor, FILE * file, const char * label = NULL, const int n = -1, const int beg = 0, const int verbose = 0);
 
     /* read data from a file */
     void Read(FILE * file, const char * label = NULL);

source/tensor/core/reduce/ReduceMax.cu

@@ -482,7 +482,7 @@ void KernelReduceMaxOp(DTYPE * input, DTYPE * output,int stride, int strideNum,
     if (tid < 32){
         if (tid < blockDim.y / 32)
             threadMax = data[tid];
-        else threadMax = 0;
+        else threadMax = FLOAT_MIN;
         threadMax = shflDownReduceMax(threadMax);
         if (tid == 0 && blockIdx.y < reducedStrideNum)
             output[(k * reducedStrideNum + blockIdx.y) * stride + iOffset] = threadMax;