bug fixes

39eef110 · xiaotong · e9d68683 · 39eef110 · 39eef110 · 39eef110
Commit 39eef110 authored Mar 05, 2021 by xiaotong
--- a/source/sample/transformer/translate/LengthPenalty.cpp
+++ b/source/sample/transformer/translate/LengthPenalty.cpp
@@ -42,7 +42,7 @@ float LengthPenalizer::GNMT(float length, float alpha)
    base = (length + 5.0F) / (1.0F + 5.0F);
-    lp = pow(base, alpha);
+    lp = (float)pow(base, alpha);
    return lp;
 }

--- a/source/tensor/XUtility.cpp
+++ b/source/tensor/XUtility.cpp
@@ -155,13 +155,13 @@ void XMemSet(int devID, void * p, int value, size_t size)
 cudaMemcpyKind GetMemcpyKind(int devIDFrom, int devIDTo)
 {
    if(devIDFrom < 0 && devIDTo < 0)
-        return cudaMemcpyHostToHost;
+        return cudaMemcpyKind::cudaMemcpyHostToHost;
    else if(devIDFrom < 0 && devIDTo >= 0)
-        return cudaMemcpyHostToDevice;
+        return cudaMemcpyKind::cudaMemcpyHostToDevice;
    else if(devIDFrom >= 0 && devIDTo < 0)
-        return cudaMemcpyDeviceToHost;
+        return cudaMemcpyKind::cudaMemcpyDeviceToHost;
    else
-        return cudaMemcpyDeviceToDevice;
+        return cudaMemcpyKind::cudaMemcpyDeviceToDevice;
 }
 #endif
@@ -553,9 +553,9 @@ void XQSort(void * data, void * index, int num, int width, int stride, int (*com
    stackptr = 0;
    lo = (char*)data;
-    hi = (char*)data + realStride * (num - 1);
+    hi = (char*)data + (long)realStride * (num - 1);
    indexlo = (int*)index;
-    indexhi = index != NULL ? (int*)index + stride * (num - 1) : NULL;
+    indexhi = index != NULL ? (int*)index + (long)stride * (num - 1) : NULL;
 recurse:
@@ -565,8 +565,8 @@ recurse:
    if(size <= MIN_QSORT_NUM)
        XShortSort(lo, hi, indexlo, indexhi, width, stride, comp);
    else {
-        mid = lo + (size/2) * realStride;
+        mid = lo + (long)(size/2) * realStride;
-        indexmid = indexlo + (size/2) * stride;
+        indexmid = indexlo + (long)(size/2) * stride;
        /* sort the first, last and middle elements into order */
        if(comp(lo, mid) > 0)
@@ -834,8 +834,7 @@ int SplitALine(char* inputString, const char* seperator, StrList* items)
        return 0;
    if (sepLen == 0) {
+        char* item = new char[(long)inputLen + 1];
-        char* item = new char[inputLen + 1];
        strcpy(item, inputString);
        items->Add(item);
    }

--- a/source/train/TTrain.cpp
+++ b/source/train/TTrain.cpp
@@ -122,12 +122,12 @@ void TTDataLoader::SetBatchSize(int myBatchSize)
 /* start the process */
 bool TTDataLoader::Start()
 {
-    file = fopen(fileName, "wb");
+    file = fopen(fileName, "rb");
-    CheckNTErrors(file, "Cannot open the file");
+    CheckNTErrors(file != NULL, "Cannot open the file");
    /* skip the first line */
    char * line = new char[MAX_SAMPLE_LINE_LENGTH];
-    fgets(line, MAX_SAMPLE_LINE_LENGTH - 1, file);
+    fgets(line, MAX_SAMPLE_LINE_LENGTH, file);
    delete[] line;
    return true;
@@ -148,6 +148,7 @@ get a batch of samples
 */
 bool TTDataLoader::GetBatchSimple(XList * inputs, XList * golds)
 {
+    fprintf(stderr, "get batch 0\n");
    CheckNTErrors(file != NULL, "No input file specificed!");
    CheckNTErrors(inputs != NULL && inputs->count >= 1, "Wrong argument!");
    CheckNTErrors(golds != NULL && golds->count >= 1, "Wrong argument!");
@@ -162,9 +163,9 @@ bool TTDataLoader::GetBatchSimple(XList * inputs, XList * golds)
    int * goldBatch = new int[batchSize];
    int A, B, C, D;
-    while (fgets(line, MAX_SAMPLE_LINE_LENGTH - 1, file)) {
+    while (fgets(line, MAX_SAMPLE_LINE_LENGTH, file)) {
-        if (count++ == batchSize)
+        if (count == batchSize)
            break;
        if (sscanf(line, "%d %d %d %d", &A, &B, &C, &D) < 4) {
@@ -175,19 +176,28 @@ bool TTDataLoader::GetBatchSimple(XList * inputs, XList * golds)
        inputBatch[count * 3 + 1] = B;
        inputBatch[count * 3 + 2] = C;
        goldBatch[count] = D;
+        count++;
    }
+    if (count > 0) {
        InitTensor2D(input, count, 3, X_INT);
        InitTensor2D(gold, count, 1, X_INT);
        input->SetData(input, count * 3);
        gold->SetData(gold, count);
+    }
    delete[] line;
    delete[] inputBatch;
    delete[] goldBatch;
+    fprintf(stderr, "get batch 1\n");
+    if (count > 0)
        return true;
+    else
+        return false;
 }
 /*****************************
@@ -233,6 +243,8 @@ void TTModel::Forward(int devID, XTensor * input, XTensor * output)
    XTensor embeddingCat;
    XTensor hidden;
+    fprintf(stderr, "forward 0\n");
    /* [e_0, e_1, e_2] = w_e * input(one-hot) */
    embedding = Gather(embeddingW, *input);
@@ -244,6 +256,8 @@ void TTModel::Forward(int devID, XTensor * input, XTensor * output)
    /* output = Softmax(h) */
    *output = Softmax(hidden, 0);
+    fprintf(stderr, "forward 1\n");
 }
 /* clear the model */

--- a/source/train/XLeader.cpp
+++ b/source/train/XLeader.cpp
@@ -264,7 +264,57 @@ bool XLeader::Run(XConfig * config, DataDistributeBase * dataDistributor,
        ShowNTErrors("No data-broadcasting workers!");
    }
+    WaitForFinishing();
    return isDataOK;
 }
+/* wait until all workers finish their job */
+void XLeader::WaitForFinishing(int sleepTime)
+{
+    while (1) {
+        bool finished = true;
+        if (finished) {
+            for (int i = 0; i < jworkers.count; i++) {
+                XWorkerJob* worker = (XWorkerJob*)jworkers[i];
+                if (worker->GetJobNum() > 0) {
+                    finished = false;
+                    break;
+                }
+            }
+        }
+        if (finished) {
+            for (int i = 0; i < cworkers.count; i++) {
+                XWorkerJob* worker = (XWorkerJob*)cworkers[i];
+                if (worker->GetJobNum() > 0) {
+                    finished = false;
+                    break;
+                }
+            }
+        }
+        if (finished) {
+            for (int i = 0; i < uworkers.count; i++) {
+                XWorkerJob* worker = (XWorkerJob*)uworkers[i];
+                if (worker->GetJobNum() > 0) {
+                    finished = false;
+                    break;
+                }
+            }
+        }
+        if (finished)
+            break;
+#ifdef _WIN32
+        Sleep((DWORD)sleepTime);
+#else
+        sleep((unsigned)sleepTime / 1000);
+#endif
+    }
+}
 } /* end of the nts (NiuTrans.Tensor) namespace */
--- a/source/train/XLeader.h
+++ b/source/train/XLeader.h
@@ -48,6 +48,7 @@
 namespace nts { // namespace nts(NiuTrans.Tensor)
 #define MAX_NUM_OF_WORKERS 1024
+#define SLEEP_TIME_IN_WAITING_FOR_JOBS 10
 /* 
 conmmunication mode of a leader. This offers a way of organizing a hierachy of the work
@@ -126,6 +127,9 @@ public:
    /* run the model (for one time) */
    bool Run(XConfig * config, DataDistributeBase * dataDistributor, 
             XModel * model, XOptimizer * optimizer);
+    /* wait until all workers finish their job */
+    void WaitForFinishing(int sleepTime = SLEEP_TIME_IN_WAITING_FOR_JOBS);
 };
 }

--- a/source/train/XModel.cpp
+++ b/source/train/XModel.cpp
@@ -124,8 +124,10 @@ bool XModel::Run(XList * args)
    XModel * model = (XModel*)args->GetItem(0);
    XList newArgs;
-    for (int i = 1; i < args->count; i++)
+    for (int i = 1; i < args->count; i++) {
-        newArgs.Add(args->GetItem(i));
+        void * arg = args->GetItem(i);
+        newArgs.Add(arg);
+    }
    return model->RunMe(&newArgs);
 }

--- a/source/train/XWorkerJob.cpp
+++ b/source/train/XWorkerJob.cpp
@@ -105,6 +105,7 @@ add a new job of model refreshment
 */
 bool XWorkerJob::AddJobRefresh(XModel * myModel)
 {
+    fprintf(stderr, "refresh 0\n");
    CheckNTErrors(myModel != NULL, "no parameter keeper!");
    XList args(1);
@@ -112,6 +113,8 @@ bool XWorkerJob::AddJobRefresh(XModel * myModel)
    queue.EnqueueJob((void*)(char*)XModel::Refresh, &args);
+    fprintf(stderr, "refresh 1\n");
    return true;
 }