bug fixes

source/sample/transformer/T2TLengthPenalty.cpp

@@ -36,7 +36,6 @@ XTensor T2TLengthPenalizer::GNMT(const XTensor & length, float alpha)
     XTensor lp;
 
     base = (length + 5)/(1 + 5);
-
     lp = Power(base, alpha);
     
     return lp;

source/sample/transformer/T2TPredictor.cpp

@@ -227,6 +227,8 @@ void T2TPredictor::Predict(T2TStateBundle * next, XTensor * encoding,
 
     /* generate the output probabilities */
     m->outputLayer->Make(decodingStep, output);
+
+    _LogMe(&output);
     
     next->layersEnc.AddList(&s->layersEnc);
     next->layersDec.Add(&inputDec);

source/sample/transformer/T2TSearch.cpp

@@ -73,8 +73,10 @@ search for the most promising states
 >> input - input of the model
 >> padding - padding of the input
 >> output - output that represents the sequences as rows
+>> score - score of the sequences
 */
-void T2TSearch::Search(T2TModel * model, XTensor * input, XTensor * padding, XTensor * output)
+void T2TSearch::Search(T2TModel * model, XTensor * input, XTensor * padding, 
+                       XTensor * output, XTensor * score)
 {
     T2TPredictor predictor;
     XTensor maskEnc;
@@ -155,7 +157,7 @@ void T2TSearch::Search(T2TModel * model, XTensor * input, XTensor * padding, XTe
     /* fill the heap with imcomplete hypotheses if neccesary */
     FillHeap(next);
     
-    Dump(output);
+    Dump(output, score);
 
     delete[] states;
 }
@@ -237,6 +239,7 @@ void T2TSearch::Score(T2TStateBundle * prev, T2TStateBundle * beam)
     InitTensor(&mask, 
                prev->endMark.order, prev->endMark.dimSize, X_FLOAT, 1.0F, 
                prev->endMark.devID, prev->endMark.mem);
+    mask.SetZeroAll();
     _SetDataFixedCond(&mask, &prev->endMark, -1e9F);
     
     mask.Reshape(mask.unitNum);
@@ -514,23 +517,28 @@ void T2TSearch::FillHeap(T2TStateBundle * beam)
 /* 
 save the output sequences in a tensor 
 >> output - output sequences (for return)
+>> score - score of thes sequences
 */
-void T2TSearch::Dump(XTensor * output)
+void T2TSearch::Dump(XTensor * output, XTensor * score)
 {
     int dims[3] = {batchSize, beamSize, maxLength};
     int * words = new int[maxLength];
 
     InitTensor(output, 3, dims, X_INT);
+    InitTensor(score, 2, dims, X_FLOAT);
     SetDataFixedInt(*output, -1);
+    score->SetZeroAll();
 
     /* heap for an input sentence in the batch */
     for(int h = 0; h < batchSize; h++){
 
         XHeap<MIN_HEAP, float> &heap = fullHypos[h];
+        int c = heap.Count();
 
         /* for each output in the beam */
         for(int i = 0; i < beamSize && heap.Count() > 0; i++){
-            T2TState * state = (T2TState *)heap.Pop().index;
+            HeapNode<float> node = heap.Pop();
+            T2TState * state = (T2TState *)node.index;
             
             int count = 0;
             bool isCompleted = true;
@@ -548,7 +556,9 @@ void T2TSearch::Dump(XTensor * output)
 
             /* dump the sentence to the output tensor */
             for(int w = 0; w < count; w++)
-                output->Set3DInt(words[count - w - 1], h, beamSize - i - 1, w);
+                output->Set3DInt(words[count - w - 1], h, c - i - 1, w);
+
+            score->Set2D(node.value, h, c - i - 1);
         }
     }
 

source/sample/transformer/T2TSearch.h

@@ -73,7 +73,8 @@ public:
     void Init(int argc, char ** argv);
 
     /* search for the most promising states */
-    void Search(T2TModel * model, XTensor * input, XTensor * padding, XTensor * output);
+    void Search(T2TModel * model, XTensor * input, XTensor * padding, 
+                XTensor * output, XTensor * score);
 
     /* preparation */
     void Prepare(int myBatchSize,int myBeamSize);
@@ -94,7 +95,7 @@ public:
     void FillHeap(T2TStateBundle * beam);
 
     /* save the output sequences in a tensor */
-    void Dump(XTensor * output);
+    void Dump(XTensor * output, XTensor * score);
 
     /* check if the token is an end symbol */
     bool IsEnd(int token);

source/sample/transformer/T2TTester.cpp

@@ -112,10 +112,12 @@ void T2TTester::Test(const char * fn, const char * ofn, T2TModel * model)
         CheckNTErrors(!model->isLM, "Only MT model is supported!");
         
         XTensor output;
+        XTensor score;
 
-        seacher.Search(model, &batchEnc, &paddingEnc, &output);
+        seacher.Search(model, &batchEnc, &paddingEnc, &output, &score);
 
         Dump(ofile, &output);
+        //score.Dump(ofile, "score:");
 
         float prob = 0;
             

source/tensor/core/getandset/SetData.cpp

@@ -238,12 +238,12 @@ void _SetDataFixedCond(XTensor * tensor, XTensor * condition, DTYPE p)
     int num = tensor->unitNum;
 
     CheckNTErrors(num == condition->unitNum, "Wrong size of the condition tensor!");
-    CheckNTErrors(condition->unitSize == sizeof(float), "TODO!");
+    CheckNTErrors(condition->unitSize == sizeof(int), "TODO!");
 
     if(tensor->dataType == DEFAULT_DTYPE){
         if(tensor->devID < 0){
             DTYPE * data = (DTYPE*)tensor->data;
-            DTYPE * cond = (DTYPE*)condition->data;
+            int * cond = (int*)condition->data;
             for(int i = 0; i < num; i++){
                 if(cond[i] != 0)
                     data[i] = p;

source/tensor/core/getandset/SetData.cu

@@ -159,7 +159,7 @@ if the condition entry is non-zero
 >> p - the initial value
 */
 __global__ 
-void KernelSetDataFixedCondFloat(float * d, float * c, int size, float p)
+void KernelSetDataFixedCondFloat(float * d, int * c, int size, float p)
 {
     int i = blockDim.x * blockIdx.x + threadIdx.x;
 
@@ -178,7 +178,7 @@ if the condition entry is non-zero
 void _CudaSetDataFixedCondFloat(XTensor * tensor, XTensor * condition, float p)
 {
     CheckNTErrors(tensor->dataType == X_FLOAT, "the tensor must be in X_FLOAT!");
-    CheckNTErrors(condition->unitSize == sizeof(float), "TODO!");
+    CheckNTErrors(condition->unitSize == sizeof(int), "TODO!");
 
     int gridSize[3];
     int blockSize[3];
@@ -191,7 +191,7 @@ void _CudaSetDataFixedCondFloat(XTensor * tensor, XTensor * condition, float p)
     int devIDBackup;
     ProtectCudaDev(tensor->devID, devIDBackup);
 
-    KernelSetDataFixedCondFloat <<<blocks, threads >>>((float*)tensor->data, (float*)condition->data, 
+    KernelSetDataFixedCondFloat <<<blocks, threads >>>((float*)tensor->data, (int*)condition->data, 
                                                                tensor->unitNum, p);
 
     BacktoCudaDev(tensor->devID, devIDBackup);
@@ -206,7 +206,7 @@ if the condition entry is non-zero
 >> p - the initial value
 */
 __global__ 
-void KernelSetDataFixedCondInt(int * d, float * c, int size, int p)
+void KernelSetDataFixedCondInt(int * d, int * c, int size, int p)
 {
     int i = blockDim.x * blockIdx.x + threadIdx.x;
 
@@ -225,7 +225,7 @@ if the condition entry is non-zero
 void _CudaSetDataFixedCondInt(XTensor * tensor, XTensor * condition, int p)
 {
     CheckNTErrors(tensor->dataType == X_FLOAT, "the tensor must be in X_FLOAT!");
-    CheckNTErrors(condition->unitSize == sizeof(float), "TODO!");
+    CheckNTErrors(condition->unitSize == sizeof(int), "TODO!");
 
     int gridSize[3];
     int blockSize[3];
@@ -238,7 +238,7 @@ void _CudaSetDataFixedCondInt(XTensor * tensor, XTensor * condition, int p)
     int devIDBackup;
     ProtectCudaDev(tensor->devID, devIDBackup);
 
-    KernelSetDataFixedCondInt <<<blocks, threads >>>((int*)tensor->data, (float*)condition->data, 
+    KernelSetDataFixedCondInt <<<blocks, threads >>>((int*)tensor->data, (int*)condition->data, 
                                                            tensor->unitNum, p);
 
     BacktoCudaDev(tensor->devID, devIDBackup);