fix some bugs in beam search

c177e3c5 · huchi · e155205c · c177e3c5 · c177e3c5 · c177e3c5
Commit c177e3c5 authored Feb 18, 2020 by huchi
--- a/source/Main.cpp
+++ b/source/Main.cpp
@@ -43,6 +43,21 @@ int main( int argc, const char ** argv )
    _CrtSetBreakAlloc(2708);*/
    TransformerMain(argc - 1, argv + 1);
+    //XTensor singleScore, singleIdx, score;
+    //InitTensor3DV2(&score, 2, 1, 136160);
+    ////score.SetDataRand(0, 1);
+    //InitTensor1DV2(&singleIdx, 1, X_INT);
+    //singleIdx.Set1DInt(1, 0);
+    //singleIdx.Dump(stderr);
+    //singleScore = Select(score, singleIdx, 0);
+    //XTensor s, i;
+    //InitTensor3DV2(&s, 2, 1, 4);
+    //InitTensor3DV2(&i, 2, 1, 4, X_INT);
+    //TopK(score, s, i, -1, 4);
+    //i.Dump(stderr, "single score:\n");
    //_CrtDumpMemoryLeaks();
    return 0;

--- a/source/sample/transformer/T2TModel.cpp
+++ b/source/sample/transformer/T2TModel.cpp
@@ -196,7 +196,7 @@ void T2TModel::MakeMT(XTensor& inputEnc, XTensor& inputDec, XTensor& output, XTe
    MakeMTMaskEnc(inputEnc, paddingEnc, maskEnc);
    /* decoder mask */
-    MakeMTMaskDec(inputEnc, inputDec, paddingEnc, paddingDec, maskDec, maskEncDec, 0);
+    MakeMTMaskDec(inputEnc, inputDec, paddingEnc, paddingDec, maskDec, maskEncDec);
    encoding = MakeEncoder(inputEnc, &maskEnc, isTraining);
@@ -225,7 +225,7 @@ void T2TModel::MakeMTMask(XTensor& inputEnc, XTensor& inputDec,
        dims[i + 1] = inputDec.GetDim(i);
    dims[0] = nhead;
    dims[inputDec.order + 1] = len;
-    InitTensorV2(&maskDec, inputDec.order + 2, dims, X_FLOAT, 1.0F, paddingDec.devID);
+    InitTensor(&maskDec, inputDec.order + 2, dims, X_FLOAT, paddingDec.devID);
    /* an upper triangular matrix where the cells of the upper triangular are set to -1e-9.
       this matrix can be used to prevent the attention to current or following words in
@@ -235,10 +235,10 @@ void T2TModel::MakeMTMask(XTensor& inputEnc, XTensor& inputDec,
    /* encoder-decoder mask that prevents the attention to padding dummy words */
    dims[inputDec.order + 1] = inputEnc.GetDim(inputEnc.order - 1);
-    InitTensorV2(&maskEncDec, inputDec.order + 2, dims, X_FLOAT, 1.0F, paddingEnc.devID);
+    InitTensor(&maskEncDec, inputDec.order + 2, dims, X_FLOAT, 1.0F, paddingEnc.devID);
-    XTensor* maskEncDecTMPEnc = NewTensorBufV2(paddingEnc.order + 1, dims + 1, paddingEnc.dataType, paddingEnc.devID);
+    XTensor * maskEncDecTMPEnc = NewTensorBuf(paddingEnc.order + 1, dims + 1, paddingEnc.dataType, paddingEnc.devID);
-    XTensor* maskEncDecTMPDec = NewTensorBufV2(paddingEnc.order + 1, dims + 1, paddingEnc.dataType, paddingEnc.devID);
+    XTensor * maskEncDecTMPDec = NewTensorBuf(maskEncDecTMPEnc, paddingEnc.devID);
    _Unsqueeze(&paddingEnc, maskEncDecTMPEnc, paddingEnc.order - 1, paddingDec.GetDim(-1));
    _ScaleAndShiftMe(maskEncDecTMPEnc, 1e9F, -1e9F);
@@ -254,15 +254,13 @@ void T2TModel::MakeMTMask(XTensor& inputEnc, XTensor& inputDec,
    dimsPadding[paddingEnc.order - 1] = paddingEnc.GetDim(-1);
    dimsPadding[paddingEnc.order] = paddingEnc.GetDim(-1);
-    XTensor* padding2 = NewTensorBufV2(paddingEnc.order + 1, dimsPadding, paddingEnc.dataType,
+    XTensor * padding2 = NewTensorBuf(paddingEnc.order + 1, dimsPadding, paddingEnc.dataType, paddingEnc.devID);
-        paddingEnc.devID);
    for (int i = 0; i < padding2->order; i++)
        dimsPadding[i + 1] = padding2->GetDim(i);
    dimsPadding[0] = nhead;
-    XTensor* padding3 = NewTensorBufV2(paddingEnc.order + 2, dimsPadding, paddingEnc.dataType,
+    XTensor * padding3 = NewTensorBuf(paddingEnc.order + 2, dimsPadding, paddingEnc.dataType, paddingEnc.devID);
-        paddingEnc.devID);
    /* mask of the padding */
    _Unsqueeze(&paddingEnc, padding2, paddingEnc.order - 1, paddingEnc.GetDim(-1));
@@ -270,7 +268,7 @@ void T2TModel::MakeMTMask(XTensor& inputEnc, XTensor& inputDec,
    _ScaleAndShiftMe(padding3, 1e9F, -1e9F);
-    InitTensorV2(&maskEnc, padding3);
+    InitTensor(&maskEnc, padding3);
    maskEnc.SetZeroAll();
    /* generate the mask on the source language side (for padding) */
@@ -298,13 +296,13 @@ void T2TModel::MakeMTMaskEnc(XTensor& inputEnc, XTensor& paddingEnc, XTensor& ma
    dimsPadding[paddingEnc.order - 1] = paddingEnc.GetDim(-1);
    dimsPadding[paddingEnc.order] = paddingEnc.GetDim(-1);
-    XTensor* padding2 = NewTensorBufV2(paddingEnc.order + 1, dimsPadding, paddingEnc.dataType, paddingEnc.devID);
+    XTensor * padding2 = NewTensorBuf(paddingEnc.order + 1, dimsPadding, paddingEnc.dataType, paddingEnc.devID);
    for (int i = 0; i < padding2->order; i++)
        dimsPadding[i + 1] = padding2->GetDim(i);
    dimsPadding[0] = nhead;
-    XTensor* padding3 = NewTensorBufV2(paddingEnc.order + 2, dimsPadding, paddingEnc.dataType, paddingEnc.devID);
+    XTensor* padding3 = NewTensorBuf(paddingEnc.order + 2, dimsPadding, paddingEnc.dataType, paddingEnc.devID);
    /* mask of the padding */
    _Unsqueeze(&paddingEnc, padding2, paddingEnc.order - 1, paddingEnc.GetDim(-1));
@@ -312,7 +310,7 @@ void T2TModel::MakeMTMaskEnc(XTensor& inputEnc, XTensor& paddingEnc, XTensor& ma
    _ScaleAndShiftMe(padding3, 1e9F, -1e9F);
-    InitTensorV2(&maskEnc, padding3);
+    InitTensor(&maskEnc, padding3);
    maskEnc.SetZeroAll();
    /* generate the mask on the source language side (for padding) */
@@ -332,32 +330,36 @@ make the mask of the decoder
 >> maksDec - mask of the decoder self-attention
 >> maksEncDec - mask of the decoder enc-dec attention
 */
-void T2TModel::MakeMTMaskDec(XTensor& inputEnc, XTensor& inputDec,
+void T2TModel::MakeMTMaskDec(XTensor &inputEnc, XTensor &inputDec,
-    XTensor& paddingEnc, XTensor& paddingDec,
+                             XTensor &paddingEnc, XTensor &paddingDec,
-    XTensor& maskDec, XTensor& maskEncDec, int incDim)
+                             XTensor &maskDec, XTensor &maskEncDec)
 {
    int len = inputDec.GetDim(inputDec.order - 1);
-    int* dims = new int[inputDec.order + 2];
+    int * dims = new int[inputDec.order + 2];
-    for (int i = 0; i < inputDec.order; i++)
+    for(int i = 0; i < inputDec.order; i++)
        dims[i + 1] = inputDec.GetDim(i);
-    //dims[inputDec.order] += incDim;
    dims[0] = nhead;
    dims[inputDec.order + 1] = len;
-    //InitTensor(&maskDec, inputDec.order + 2, dims, X_FLOAT, 1.0F, paddingDec.devID, paddingDec);
+    InitTensor(&maskDec, inputDec.order + 2, dims, X_FLOAT, paddingDec.devID);
    /* An upper triangular matrix where the cells of the upper triangular are set to -1e-9.
       This matrix can be used to block the attention to current or following words in
       a given sequence. */
-       //_SetDataLowTri(&maskDec, 1e9F, 0);
+    _SetDataLowTri(&maskDec, 1e9F, 0);
+    //maskDec.Dump(stderr, "mask: ");
-       //_ScaleAndShiftMe(&maskDec, 1.0F, -1e9F);
+    _ScaleAndShiftMe(&maskDec, 1.0F, -1e9F);
+    //maskDec.Dump(stderr, "mask: ");
    /* encoder-decoder mask that prevents the attention to padding dummy words */
    dims[inputDec.order + 1] = inputEnc.GetDim(inputEnc.order - 1);
-    InitTensorV2(&maskEncDec, inputDec.order + 2, dims, X_FLOAT, 1.0F, paddingEnc.devID);
+    InitTensor(&maskEncDec, inputDec.order + 2, dims, X_FLOAT, paddingEnc.devID);
-    XTensor* maskEncDecTMPEnc = NewTensorBufV2(paddingEnc.order + 1, dims + 1, paddingEnc.dataType, paddingEnc.devID);
+    XTensor * maskEncDecTMPEnc = NewTensorBuf(paddingEnc.order + 1, dims + 1, paddingEnc.dataType,
-    XTensor* maskEncDecTMPDec = NewTensorBufV2(paddingEnc.order + 1, dims + 1, paddingEnc.dataType, paddingEnc.devID);
+                                              paddingEnc.devID);
+    XTensor * maskEncDecTMPDec = NewTensorBuf(maskEncDecTMPEnc, paddingEnc.devID);
    _Unsqueeze(&paddingEnc, maskEncDecTMPEnc, paddingEnc.order - 1, paddingDec.GetDim(-1));

--- a/source/sample/transformer/T2TModel.h
+++ b/source/sample/transformer/T2TModel.h
@@ -90,9 +90,9 @@ public:
    void MakeMTMaskEnc(XTensor &inputEnc, XTensor &paddingEnc, XTensor &maskEnc);
    /* make the mask of the decoder */
-    void MakeMTMaskDec(XTensor &inputEnc, XTensor &inputDec,
+    void MakeMTMaskDec(XTensor& inputEnc, XTensor& inputDec,
-                       XTensor &paddingEnc, XTensor &paddingDec,
+        XTensor& paddingEnc, XTensor& paddingDec,
-                       XTensor &maskDec, XTensor &maskEncDec, int incDim);
+        XTensor& maskDec, XTensor& maskEncDec);
    /* get parameter matrics */
    void GetParams(TensorList &list);

--- a/source/sample/transformer/T2TPredictor.cpp
+++ b/source/sample/transformer/T2TPredictor.cpp
@@ -166,7 +166,6 @@ void T2TPredictor::Predict(T2TStateBundle* next, XTensor* encoding, XTensor* inp
        inputDec = GetLastPrediction(s);
        inputDec.SetDevice(inputEnc->devID);
    }
-    inputDec.Dump(stderr, "inputDec");
    /* prediction probabilities */
    XTensor& output = next->prob;
@@ -184,10 +183,10 @@ void T2TPredictor::Predict(T2TStateBundle* next, XTensor* encoding, XTensor* inp
    XTensor maskEncDec;
    /* decoder mask */
-    //m->MakeMTMaskDec(*inputEnc, inputDec, *paddingEnc, paddingDec, maskDec, maskEncDec, 0);
+    m->MakeMTMaskDec(*inputEnc, inputDec, *paddingEnc, paddingDec, maskDec, maskEncDec);
    /* make the decoding network */
-    decoding = m->decoder->Make(inputDec, *encoding, NULL, maskEncDec, false);
+    decoding = m->decoder->Make(inputDec, *encoding, &maskDec, maskEncDec, false);
    CheckNTErrors(decoding.order >= 2, "The tensor must be of order 2 or larger!");

--- a/source/sample/transformer/T2TSearch.cpp
+++ b/source/sample/transformer/T2TSearch.cpp
--- a/source/sample/transformer/T2TSearch.h
+++ b/source/sample/transformer/T2TSearch.h
@@ -62,6 +62,12 @@ private:
    /* start symbol */
    int startSymbol;
+    /* scalar of the input sequence (for max number of search steps) */
+    float scalarMaxLength;
+    /* indicate whether the early stop strategy is used */
+    bool isEarlyStop;
 public:
    /* constructor */
    T2TSearch();
@@ -73,7 +79,7 @@ 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);
@@ -93,12 +99,15 @@ public:
    /* fill the hypotheis heap with incomplete hypothses */
    void FillHeap(T2TStateBundle* beam);
-    /* save the output sequences in a tensor */
+    /* save the output sequences and score */
-    void Dump(XTensor* output);
+    void Dump(XTensor* output, XTensor* score);
    /* check if the token is an end symbol */
    bool IsEnd(int token);
+    /*check whether all hypotheses are completed*/
+    bool IsAllCompleted(T2TStateBundle* beam);
    /* set end symbols for search */
    void SetEnd(const int* tokens, const int tokenNum);

--- a/source/sample/transformer/T2TTester.cpp
+++ b/source/sample/transformer/T2TTester.cpp
@@ -101,9 +101,10 @@ void T2TTester::Test(const char* fn, const char* ofn, T2TModel* model)
        vector<int> indices = batchLoader.LoadBatch(&batchEnc, &paddingEnc, sentBatch, devID);
        XTensor output;
+        XTensor score;
+        seacher.Search(model, &batchEnc, &paddingEnc, &output, &score);
-        seacher.Search(model, &batchEnc, &paddingEnc, &output);
-        output.Dump(stderr);
        for (int i = 0; i < indices.size(); ++i) {
            Result res;
            XTensor sent, srcIdx, tgtIdx;
@@ -127,9 +128,7 @@ void T2TTester::Test(const char* fn, const char* ofn, T2TModel* model)
        if (batchCount % 1 == 0) {
            double elapsed = GetClockSec() - startT;
-            XPRINT3(0, stderr,
+            XPRINT3(0, stderr, "[INFO] elapsed=%.1fs, sentence=%d, sword=%d\n", elapsed, sentCount, wordCount);
-                "[INFO] elapsed=%.1fs, sentence=%d, sword=%d\n",
-                elapsed, sentCount, wordCount);
        }
    }
@@ -160,9 +159,10 @@ void T2TTester::Dump(FILE* file, XTensor* output)
    for (int i = 0; i < output->unitNum; i += seqLength) {
        for (int j = 0; j < seqLength; j++) {
            int w = output->GetInt(i + j);
-            fprintf(file, "%d ", w);
+            if (w < 0 || w == 1)
-            if (w < 0)
                break;
+            fprintf(file, "%d ", w);
        }
        fprintf(file, "\n");