diff --git a/paddle/cuda/include/hl_matrix.h b/paddle/cuda/include/hl_matrix.h
index c7f25109972195fb56b9e96c4b68d952363e6338..7daca18761b80eac0f876b21377a6ccc6a853485 100644
--- a/paddle/cuda/include/hl_matrix.h
+++ b/paddle/cuda/include/hl_matrix.h
@@ -300,4 +300,12 @@ extern void hl_matrix_col2Vol(real* dataDst,
                               real alpha,
                               real beta);
 
+/**
+ * @brief  Matrix col2Vol: Convert col matrix into 3D volume
+ * @param[out]  out     output int vector.
+ * @param[in]   vec     input float vector.
+ * @param[in]   size    size of the vector.
+ */
+extern void hl_vector_cast2int(int* out, real* vec, int size);
+
 #endif /* HL_MATRIX_H_ */
diff --git a/paddle/cuda/include/stub/hl_matrix_stub.h b/paddle/cuda/include/stub/hl_matrix_stub.h
index 6ac332945c8f09fef23f35680ba5bb1d9ba9f4fd..46e77e140768dd80fd327dd4eb3b0f62a3370950 100644
--- a/paddle/cuda/include/stub/hl_matrix_stub.h
+++ b/paddle/cuda/include/stub/hl_matrix_stub.h
@@ -133,4 +133,6 @@ inline void hl_matrix_col2Vol(real* dataDst,
                               real alpha,
                               real beta) {}
 
+inline void hl_vector_cast2int(int* out, real* vec, int size) {}
+
 #endif  // HL_MATRIX_STUB_H_
diff --git a/paddle/cuda/src/hl_cuda_matrix.cu b/paddle/cuda/src/hl_cuda_matrix.cu
index b41a3a1e06db7b2566acef19ce430645f79d486d..607efb4f6b0aa0d22a2789397b8743f7a5271d5b 100644
--- a/paddle/cuda/src/hl_cuda_matrix.cu
+++ b/paddle/cuda/src/hl_cuda_matrix.cu
@@ -793,3 +793,14 @@ void hl_matrix_col2Vol(real* dataDst,
 
   CHECK_SYNC("hl_matrix_col2Vol failed");
 }
+
+__global__ void keVectorCast2Int(int* out, real* vec, int size) {
+  for (int i = threadIdx.x; i < (size); i += blockDim.x) {
+    out[i] = int(vec[i]);
+  }
+}
+
+void hl_vector_cast2int(int* out, real* vec, int size) {
+  keVectorCast2Int<<<1, 512, 0, STREAM_DEFAULT>>>(out, vec, size);
+  CHECK_SYNC("hl_vector_cast2int failed");
+}
diff --git a/paddle/gserver/evaluators/Evaluator.cpp b/paddle/gserver/evaluators/Evaluator.cpp
index 9db6d252d97bfeee3fe376bcda431fe94c65a678..87cb2d280866ac2be2d6f85e872e547e12548feb 100644
--- a/paddle/gserver/evaluators/Evaluator.cpp
+++ b/paddle/gserver/evaluators/Evaluator.cpp
@@ -395,14 +395,24 @@ real AucEvaluator::evalImp(std::vector<Argument>& arguments) {
   CHECK_LE(arguments.size(), (size_t)3);
   MatrixPtr output = arguments[0].value;
   IVectorPtr label = arguments[1].ids;
+  MatrixPtr labelval = arguments[1].value;
   bool supportWeight = (3 == arguments.size()) ? true : false;
   MatrixPtr weight = supportWeight ? arguments[2].value : nullptr;
-  if (nullptr == output || nullptr == label ||
-      (supportWeight && nullptr == weight)) {
+
+  if (nullptr == output || (supportWeight && nullptr == weight)) {
     return 0;
   }
   size_t insNum = output->getHeight();
   size_t outputDim = output->getWidth();
+  // Copy label from value to a vector.
+  if (nullptr == label && nullptr != labelval) {
+    // label width is 1
+    CHECK_EQ(1, labelval->getWidth());
+    VectorPtr vec =
+        Vector::create(labelval->getData(), insNum, output->useGpu());
+    label = vec->castToInt();
+  }
+
   CHECK_EQ(insNum, label->getSize());
   if (supportWeight) {
     CHECK_EQ(insNum, weight->getHeight());
@@ -443,6 +453,7 @@ real AucEvaluator::evalImp(std::vector<Argument>& arguments) {
   int* labelD = label->getData();
   real* weightD = supportWeight ? weight->getData() : nullptr;
   size_t pos = realColumnIdx_;
+
   for (size_t i = 0; i < insNum; ++i) {
     real value = outputD[pos];
     uint32_t binIdx = static_cast<uint32_t>(value * kBinNum_);
diff --git a/paddle/math/Vector.cpp b/paddle/math/Vector.cpp
index ff72672e3ab77212b309fcfea835839a916fa632..346008439c35a2bcbcd2e9dfd36d689e01d7495f 100644
--- a/paddle/math/Vector.cpp
+++ b/paddle/math/Vector.cpp
@@ -18,6 +18,7 @@ limitations under the License. */
 #include <memory>
 #include "Matrix.h"
 #include "hl_gpu.h"
+#include "hl_matrix.h"
 #include "hl_table_apply.h"
 #include "paddle/utils/Flags.h"
 #include "paddle/utils/Logging.h"
@@ -99,6 +100,19 @@ MatrixPtr VectorT<int>::toOneHotSparseMatrix(size_t idRange, bool useGpu) {
   return mat;
 }
 
+template <>
+std::shared_ptr<VectorT<int>> VectorT<real>::castToInt() {
+  std::shared_ptr<VectorT<int>> ret = IVector::create(this->getSize(), useGpu_);
+  if (useGpu_) {
+    hl_vector_cast2int(ret->getData(), this->getData(), this->getSize());
+  } else {
+    for (size_t i = 0; i < getSize(); ++i) {
+      ret->getData()[i] = int(this->getData()[i]);
+    }
+  }
+  return ret;
+}
+
 template <class T>
 GpuVectorT<T>::GpuVectorT(size_t size)
     : VectorT<T>(size,
diff --git a/paddle/math/Vector.h b/paddle/math/Vector.h
index 80b9775fccf10c57bb48145ef56165ec7c86d8b8..f965a5809209da313c78a545c44e7aa39e95ac65 100644
--- a/paddle/math/Vector.h
+++ b/paddle/math/Vector.h
@@ -162,6 +162,13 @@ public:
    */
   std::shared_ptr<Matrix> toOneHotSparseMatrix(size_t idRange, bool useGpu);
 
+  /**
+   * @brief cast vector of "real" elements to "int" elements.
+   *
+   * @note: float -> int must be casted, or you'll get wrong data.
+   */
+  std::shared_ptr<VectorT<int>> castToInt();
+
   /**
    * This function will crash if the size of src and dest is different.
    */