diff --git a/mindspore/ccsrc/plugin/device/cpu/kernel/clip_by_norm_cpu_kernel.cc b/mindspore/ccsrc/plugin/device/cpu/kernel/clip_by_norm_cpu_kernel.cc
index ee086e3e039..f16a97392de 100644
--- a/mindspore/ccsrc/plugin/device/cpu/kernel/clip_by_norm_cpu_kernel.cc
+++ b/mindspore/ccsrc/plugin/device/cpu/kernel/clip_by_norm_cpu_kernel.cc
@@ -48,9 +48,7 @@ bool ClipByNormCpuKernelMod::Init(const BaseOperatorPtr &base_operator, const st
                                   const std::vector<KernelTensorPtr> &outputs) {
   // Get C++ primitive and kernel_name
   MS_EXCEPTION_IF_NULL(base_operator);
-  auto prim = std::dynamic_pointer_cast<ops::ClipByNorm>(base_operator);
-  MS_EXCEPTION_IF_NULL(prim);
-  kernel_name_ = prim->name();
+  kernel_name_ = base_operator->name();
   // Check whether current input and output data types are valid.
   auto kernel_attr = GetKernelAttrFromTensors(inputs, outputs);
   if (!MatchKernelAttr(kernel_attr, GetOpSupport()).first) {
@@ -67,15 +65,15 @@ int ClipByNormCpuKernelMod::Resize(const BaseOperatorPtr &base_operator, const s
   if (auto ret = KernelMod::Resize(base_operator, inputs, outputs, inputsOnHost); ret != KRET_OK) {
     return ret;
   }
-  auto prim = std::dynamic_pointer_cast<ops::ClipByNorm>(base_operator);
   ResetResource();
   // Init basic variables
   InitIOShape(inputs, outputs);
-  InitAxisAndEpsilon(prim);
   // Init the `l2_norm` reduce shape according to `axis`
   l2_norm_output_shape_ = x_shape_;
+  auto prim = std::dynamic_pointer_cast<ops::ClipByNorm>(base_operator);
+  MS_EXCEPTION_IF_NULL(prim);
+  InitAxisAndEpsilon(prim);
   (void)std::for_each(axis_.begin(), axis_.end(), [this](const size_t &idx) { l2_norm_output_shape_[idx] = 1; });
-
   InitSizeLists();
   return KRET_OK;
 }
@@ -109,14 +107,13 @@ std::vector<KernelAttr> ClipByNormCpuKernelMod::GetOpSupport() { return clip_by_
 void ClipByNormCpuKernelMod::ResetResource() {
   epsilon_ = 0.000001f;
   x_dim_ = 0;
-  data_type_ = std::make_pair(kNumberTypeFloat32, kNumberTypeFloat32);
-  axis_.clear();
   x_shape_.clear();
   clip_norm_shape_.clear();
   l2_norm_output_shape_.clear();
   output_shape_.clear();
   input_size_list_.clear();
   output_size_list_.clear();
+  workspace_size_list_.clear();
 }
 
 void ClipByNormCpuKernelMod::InitIOShape(const std::vector<KernelTensorPtr> &inputs,
@@ -161,6 +158,7 @@ void ClipByNormCpuKernelMod::InitAxisAndEpsilon(const ops::ClipByNormPtr &prim)
     MS_EXCEPTION(TypeError) << "For `" << kernel_name_ << "`, the type of attribute `axis` is invalid.";
   }
   // Init `axis_`
+  axis_.clear();
   if (temp_axis.empty()) {
     for (size_t i = 0; i < x_dim_; ++i) {
       (void)axis_.emplace_back(i);  // Reduce for all dimensions.