diff --git a/inference-engine/src/mkldnn_plugin/nodes/mkldnn_mvn_node.cpp b/inference-engine/src/mkldnn_plugin/nodes/mkldnn_mvn_node.cpp
index 36d9942b09932f..8d810fc259e503 100644
--- a/inference-engine/src/mkldnn_plugin/nodes/mkldnn_mvn_node.cpp
+++ b/inference-engine/src/mkldnn_plugin/nodes/mkldnn_mvn_node.cpp
@@ -691,37 +691,6 @@ MKLDNNMVNNode::MKLDNNMVNNode(const std::shared_ptr<ngraph::Node>& op, const mkld
         epsMode_ = INSIDE_SQRT;
         acrossChannels_ = mvnOp->get_across_channels();
     }
-
-    transformTo5DCase(inDataShape);
-}
-
-void MKLDNNMVNNode::transformTo5DCase(const ngraph::Shape& shape) {
-    switch (shape.size()) {
-        // for 1 and 2 rank, if acrossChannels_ is true, adjust shape to fully vectorize under unified 5d procedure.
-        // otherwise there are not enough data in spatial dimension to process in one kernel.
-        case 1 :  // C
-            if (acrossChannels_) {
-                shape5D = std::make_tuple(1, 1, 1, 1, shape[0]);
-                acrossChannels_ = false;
-                break;
-            } else {
-                shape5D = std::make_tuple(1, shape[0], 1, 1, 1);
-                break;
-            }
-        case 2 :  // NC
-            if (acrossChannels_) {
-                shape5D = std::make_tuple(1, shape[0], 1, shape[1], 1);
-                acrossChannels_ = false;
-                break;
-            } else {
-                shape5D = std::make_tuple(shape[0], shape[1], 1, 1, 1);
-                break;
-            }
-        case 3 : { shape5D = std::make_tuple(shape[0], shape[1], 1, shape[2], 1); break; }
-        case 4 : { shape5D = std::make_tuple(shape[0], shape[1], 1, shape[2], shape[3]); break; }
-        case 5 : { shape5D = std::make_tuple(shape[0], shape[1], shape[2], shape[3], shape[4]); break; }
-        default : { IE_THROW() << "MVN layer with name '" << getName() << "' doesn't support planar layout with rank: " << shape.size(); }
-    }
 }
 
 void MKLDNNMVNNode::getSupportedDescriptors() {
@@ -850,6 +819,7 @@ void MKLDNNMVNNode::createPrimitive() {
     jcp.normalize_variance = normalizeVariance_;
     jcp.across_channels = acrossChannels_;
     SizeVector in_dims = getParentEdgeAt(0)->getDims().ToSizeVector();
+    transformTo5DCase(in_dims);
     int N = 0;
     std::tie(N, jcp.C, jcp.D, jcp.H, jcp.W) = shape5D;
 
@@ -892,6 +862,35 @@ void MKLDNNMVNNode::createPrimitive() {
         mvn_variance_kernel->create_ker();
 }
 
+void MKLDNNMVNNode::transformTo5DCase(const SizeVector& shape) {
+    switch (shape.size()) {
+        // for 1 and 2 rank, if acrossChannels_ is true, adjust shape to fully vectorize under unified 5d procedure.
+        // otherwise there are not enough data in spatial dimension to process in one kernel.
+        case 1 :  // C
+            if (acrossChannels_) {
+                shape5D = std::make_tuple(1, 1, 1, 1, shape[0]);
+                acrossChannels_ = false;
+                break;
+            } else {
+                shape5D = std::make_tuple(1, shape[0], 1, 1, 1);
+                break;
+            }
+        case 2 :  // NC
+            if (acrossChannels_) {
+                shape5D = std::make_tuple(1, shape[0], 1, shape[1], 1);
+                acrossChannels_ = false;
+                break;
+            } else {
+                shape5D = std::make_tuple(shape[0], shape[1], 1, 1, 1);
+                break;
+            }
+        case 3 : { shape5D = std::make_tuple(shape[0], shape[1], 1, shape[2], 1); break; }
+        case 4 : { shape5D = std::make_tuple(shape[0], shape[1], 1, shape[2], shape[3]); break; }
+        case 5 : { shape5D = std::make_tuple(shape[0], shape[1], shape[2], shape[3], shape[4]); break; }
+        default : { IE_THROW() << "MVN layer with name '" << getName() << "' doesn't support planar layout with rank: " << shape.size(); }
+    }
+}
+
 void MKLDNNMVNNode::setPostOps(mkldnn::primitive_attr &attr, bool initWeights) {
     mkldnn::post_ops ops;
     for (auto &node : fusedWith) {
diff --git a/inference-engine/src/mkldnn_plugin/nodes/mkldnn_mvn_node.h b/inference-engine/src/mkldnn_plugin/nodes/mkldnn_mvn_node.h
index 9ce7e784e406d0..dd0090c3d72079 100644
--- a/inference-engine/src/mkldnn_plugin/nodes/mkldnn_mvn_node.h
+++ b/inference-engine/src/mkldnn_plugin/nodes/mkldnn_mvn_node.h
@@ -103,7 +103,7 @@ class MKLDNNMVNNode : public MKLDNNNode {
 
     void setPostOps(mkldnn::primitive_attr &attr, bool initWeights = false);
 
-    void transformTo5DCase(const ngraph::Shape& shape);
+    void transformTo5DCase(const InferenceEngine::SizeVector& shape);
 
     std::tuple<size_t, size_t, size_t, size_t, size_t> shape5D;