diff --git a/openbr/plugins/cuda/cudapca.cpp b/openbr/plugins/cuda/cudapca.cpp
index 2d3038c..9341c24 100644
--- a/openbr/plugins/cuda/cudapca.cpp
+++ b/openbr/plugins/cuda/cudapca.cpp
@@ -33,6 +33,7 @@ using namespace cv;
 // definitions from the CUDA source file
 namespace br { namespace cuda { namespace pca {
   void initializeWrapper(float* evPtr, int evRows, int evCols, float* meanPtr, int meanElems);
+  void trainWrapper(void* cudaSrc, float* dst, int rows, int cols);
   void wrapper(void* src, void** dst);
 }}}
 
@@ -94,6 +95,7 @@ private:
           int type = *((int*)srcDataPtr[3]);
 
           Mat mat = Mat(rows, cols, type);
+          br::cuda::pca::trainWrapper(cudaMemPtr, mat.ptr<float>(), rows, cols);
           trainingQlist.append(Template(mat));
         }
 
diff --git a/openbr/plugins/cuda/cudapca.cu b/openbr/plugins/cuda/cudapca.cu
index dbc2877..1a2861f 100644
--- a/openbr/plugins/cuda/cudapca.cu
+++ b/openbr/plugins/cuda/cudapca.cu
@@ -111,6 +111,11 @@ namespace br { namespace cuda { namespace pca {
     CUDA_SAFE_MALLOC(&intermediaryBuffer, _numSteps*_evCols*sizeof(float), &err);
   }
 
+  void trainWrapper(void* cudaSrc, float* data, int rows, int cols) {
+    cudaError_t err;
+    CUDA_SAFE_MEMCPY(data, cudaSrc, rows*cols*sizeof(float), cudaMemcpyDeviceToHost, &err);
+  }
+
   void wrapper(void* src, void** dst) {
     cudaError_t err;
     CUDA_SAFE_MALLOC(dst, _evCols*sizeof(float), &err);