diff --git a/paddle/cuda/src/hl_cuda_cudnn.cc b/paddle/cuda/src/hl_cuda_cudnn.cc
index 19c94b2453981301bcb632ecbe5d322369009973..c2dce1977bdf5daefb6c5b8032bb6b12563e9425 100644
--- a/paddle/cuda/src/hl_cuda_cudnn.cc
+++ b/paddle/cuda/src/hl_cuda_cudnn.cc
@@ -150,7 +150,7 @@ CUDNN_DNN_ROUTINE_EACH_AFTER_R3(DYNAMIC_LOAD_CUDNN_WRAP)
 
 
 // APIs available after R4:
-#if CUDNN_VERSION >= 4000
+#if CUDNN_VERSION >= 4007
 #define CUDNN_DNN_ROUTINE_EACH_AFTER_R4(__macro)             \
   __macro(cudnnBatchNormalizationForwardTraining)            \
   __macro(cudnnBatchNormalizationForwardInference)           \
@@ -999,7 +999,7 @@ void hl_batch_norm_forward_training(hl_tensor_descriptor inputDesc,
                                     double epsilon,
                                     real *savedMean,
                                     real *savedVar) {
-#if CUDNN_VERSION >= 4000
+#if CUDNN_VERSION >= 4007
   if ((NULL != runningMean && NULL == runningInvVar) ||
       (NULL == runningMean && NULL != runningInvVar)) {
     LOG(FATAL) << "runningMean and runningInvVar can be NULL "
@@ -1024,7 +1024,7 @@ void hl_batch_norm_forward_training(hl_tensor_descriptor inputDesc,
 
   CHECK_SYNC("hl_batch_norm_forward_training failed");
 #else
-  LOG(FATAL) << "CudnnBatchNorm requires cudnn version >= 4000. "
+  LOG(FATAL) << "CudnnBatchNorm requires cudnn version >= 4007. "
              << "But cudnn lib version is " << g_cudnn_lib_version;
 #endif
 }
@@ -1039,7 +1039,7 @@ void hl_batch_norm_forward_inference(hl_tensor_descriptor inputDesc,
                                     real *estimatedMean,
                                     real *estimatedInvVar,
                                     double epsilon) {
-#if CUDNN_VERSION >= 4000
+#if CUDNN_VERSION >= 4007
   cudnnTensorDescriptor_t xDesc = GET_TENSOR_DESCRIPTOR(inputDesc);
   cudnnTensorDescriptor_t yDesc = GET_TENSOR_DESCRIPTOR(outputDesc);
   cudnnTensorDescriptor_t bnDesc = GET_TENSOR_DESCRIPTOR(bnParamDesc);
@@ -1053,7 +1053,7 @@ void hl_batch_norm_forward_inference(hl_tensor_descriptor inputDesc,
 
   CHECK_SYNC("hl_batch_norm_forward_inference failed");
 #else
-  LOG(FATAL) << "CudnnBatchNorm requires cudnn version >= 4000. "
+  LOG(FATAL) << "CudnnBatchNorm requires cudnn version >= 4007. "
              << "But cudnn lib version is " << g_cudnn_lib_version;
 #endif
 }
@@ -1071,7 +1071,7 @@ void hl_batch_norm_backward(hl_tensor_descriptor inputDesc,
                             double epsilon,
                             real *savedMean,
                             real *savedInvVar) {
-#if CUDNN_VERSION >= 4000
+#if CUDNN_VERSION >= 4007
   if ((NULL != savedMean && NULL == savedInvVar) ||
       (NULL == savedMean && NULL != savedInvVar)) {
     LOG(FATAL) << "savedMean and savedVar can be NULL "
@@ -1087,16 +1087,14 @@ void hl_batch_norm_backward(hl_tensor_descriptor inputDesc,
   cudnnBatchNormMode_t mode = CUDNN_BATCHNORM_SPATIAL;
   CHECK_CUDNN(dynload::cudnnBatchNormalizationBackward(
               t_resource.cudnn_handle, mode, &alpha, &beta,
-#if CUDNN_VERSION >= 5000
               &alpha, &beta,
-#endif
               xDesc, input, dyDesc, outGrad, dxDesc, inGrad,
               bnDesc, scale, scaleGrad, biasGrad, epsilon,
               savedMean, savedInvVar));
 
   CHECK_SYNC("hl_batch_norm_backward failed");
 #else
-  LOG(FATAL) << "CudnnBatchNorm requires cudnn version >= 4000. "
+  LOG(FATAL) << "CudnnBatchNorm requires cudnn version >= 4007. "
              << "But cudnn lib version is " << g_cudnn_lib_version;
 #endif
 }
diff --git a/paddle/gserver/layers/CudnnBatchNormLayer.cpp b/paddle/gserver/layers/CudnnBatchNormLayer.cpp
index cef8772fc254f98d676e6fb89042487315280c61..3c6d13b0bf92ea98eb5c3331a1fdff6b177529b6 100644
--- a/paddle/gserver/layers/CudnnBatchNormLayer.cpp
+++ b/paddle/gserver/layers/CudnnBatchNormLayer.cpp
@@ -115,29 +115,11 @@ void CudnnBatchNormLayer::backward(const UpdateCallback& callback) {
     create(tmpBiasGrad_, 1, channels_, &betaGrad);
   }
 
-  // because of the different api of cudnn v4 and v5.
-  if (hl_get_cudnn_lib_version() < 5000) {
-    if (weight_->getWGrad()) {
-      create(tmpWGrad_, 1, channels_, &gammaGrad);
-    }
-    if (biases_ && biases_->getWGrad()) {
-      create(tmpBiasGrad_, 1, channels_, &betaGrad);
-    }
-  }
-
   hl_batch_norm_backward(ioDesc_, input, ioDesc_, outGrad,
                          ioDesc_, inGrad, bnParamDesc_,
                          gamma, gammaGrad, betaGrad,
                          EPS, savedMean, savedInvVar);
 
-  // because of the different api of cudnn v4 and v5.
-  if (hl_get_cudnn_lib_version() < 5000) {
-    if (weight_->getWGrad() && biases_->getWGrad()) {
-      weight_->getWGrad()->add(*tmpWGrad_);
-      biases_->getWGrad()->add(*tmpBiasGrad_);
-    }
-  }
-
   {
     REGISTER_TIMER_INFO("WeightUpdate", getName().c_str());
     biases_->getParameterPtr()->incUpdate(callback);
diff --git a/python/paddle/trainer/config_parser.py b/python/paddle/trainer/config_parser.py
index b26a63e7f3c1d2fdfc5fea0a034a2f2c5238d1f0..3656d9e7d8242733aa62294012f4faae1d72c1c8 100644
--- a/python/paddle/trainer/config_parser.py
+++ b/python/paddle/trainer/config_parser.py
@@ -1614,7 +1614,7 @@ class BatchNormLayer(LayerBase):
         # Also based on cudnn version.
         use_cudnn = use_gpu and batch_norm_type != "batch_norm" and \
             ((not parallel_nn) or self.config.device > -1) and \
-            cudnn_version >= 4000
+            cudnn_version >= 4007
         self.layer_type = "cudnn_batch_norm" if use_cudnn else "batch_norm"
         super(BatchNormLayer, self).__init__(name, self.layer_type, 0,
                                              active_type=active_type,
diff --git a/python/paddle/trainer_config_helpers/layers.py b/python/paddle/trainer_config_helpers/layers.py
index b7e5f566bb8c39fa6ea9ed491f28fa046bba71ee..bbd8891ce6952e920835c98484351d1423b187ed 100644
--- a/python/paddle/trainer_config_helpers/layers.py
+++ b/python/paddle/trainer_config_helpers/layers.py
@@ -171,6 +171,8 @@ class LayerOutput(object):
         assert LayerType.is_layer_type(layer_type)
         self.name = name
         self.layer_type = layer_type
+        if parents is not None and type(parents) != list:
+            parents = [parents]
         self.parents = [] if parents is None else parents
         self.activation = activation
         self.num_filters = num_filters
@@ -512,7 +514,7 @@ class MixedLayerType(LayerOutput):
         :rtype: MixedLayerType
         """
         if not self.finalized:
-            assert isinstance(other, Projection)
+            assert isinstance(other, Projection) or isinstance(other, Operator)
             self.inputs.append(other)
             self.parents.append(other.origin)
             return self