diff --git a/mindspore/ccsrc/backend/kernel_compiler/cpu/nnacl/infer/lstm_grad_infer.c b/mindspore/ccsrc/backend/kernel_compiler/cpu/nnacl/infer/lstm_grad_infer.c
new file mode 100644
index 00000000000..75353e64108
--- /dev/null
+++ b/mindspore/ccsrc/backend/kernel_compiler/cpu/nnacl/infer/lstm_grad_infer.c
@@ -0,0 +1,53 @@
+/**
+ * Copyright 2021 Huawei Technologies Co., Ltd
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include "nnacl/infer/infer_register.h"
+#include "nnacl/infer/common_infer.h"
+#include "nnacl/fp32/lstm_fp32.h"
+
+int LstmGradInferShape(const TensorC *const *inputs, size_t inputs_size, TensorC **outputs, size_t outputs_size,
+                       OpParameter *parameter) {
+  int check_ret = CheckAugmentNullSize(inputs, inputs_size, outputs, outputs_size, parameter, 11, 4);
+  if (check_ret != NNACL_OK) {
+    return check_ret;
+  }
+
+  const TensorC *input = inputs[0];
+  const TensorC *H = inputs[1];
+  const TensorC *C = inputs[2];
+  const TensorC *weight = inputs[3];
+  TensorC *output = outputs[0];
+  for (int i = 0; i < outputs_size; i++) {
+    SetDataTypeFormat(outputs[i], input);
+  }
+
+  if (!InferFlag(inputs, inputs_size)) {
+    return NNACL_INFER_INVALID;
+  }
+
+  if (input->shape_size_ != 3 || weight->shape_size_ != 3) {
+    return NNACL_ERR;
+  }
+
+  SetShapeArray(output, input->shape_, input->shape_size_);
+  SetShapeArray(outputs[1], H->shape_, H->shape_size_);
+  SetShapeArray(outputs[2], C->shape_, C->shape_size_);
+  SetShapeArray(outputs[3], weight->shape_, weight->shape_size_);
+
+  return NNACL_OK;
+}
+
+REG_INFER(LSTMGrad, PrimType_LSTMGrad, LstmGradInferShape)
diff --git a/mindspore/lite/src/cxx_api/train/train_support.cc b/mindspore/lite/src/cxx_api/train/train_support.cc
index a0467891ba0..74a4bbd91b9 100644
--- a/mindspore/lite/src/cxx_api/train/train_support.cc
+++ b/mindspore/lite/src/cxx_api/train/train_support.cc
@@ -87,6 +87,7 @@ std::shared_ptr<session::LiteSession> CreateTrainSession(std::shared_ptr<Graph::
 class TrainSupport {
  public:
   TrainSupport() { CreateTrainSessionCallbackHolder(CreateTrainSession); }
+  ~TrainSupport() {}
 };
 
 TrainSupport support_train_api;
diff --git a/mindspore/lite/src/runtime/kernel/arm/fp16_grad/pooling_fp16_grad.cc b/mindspore/lite/src/runtime/kernel/arm/fp16_grad/pooling_fp16_grad.cc
index 0002baf46eb..c464cb59880 100644
--- a/mindspore/lite/src/runtime/kernel/arm/fp16_grad/pooling_fp16_grad.cc
+++ b/mindspore/lite/src/runtime/kernel/arm/fp16_grad/pooling_fp16_grad.cc
@@ -35,11 +35,9 @@ constexpr int kNumShapeDim_2 = 2;
 }  // namespace
 int PoolingGradCPUKernelFp16::ReSize() {
   CHECK_LESS_RETURN(in_tensors_.size(), DIMENSION_3D);
-  CHECK_LESS_RETURN(out_tensors_.size(), 1);
+  CHECK_LESS_RETURN(out_tensors_.size(), DIMENSION_1D);
   PoolingParameter *pool_param = reinterpret_cast<PoolingParameter *>(op_parameter_);
   CHECK_NULL_RETURN(pool_param);
-  CHECK_LESS_RETURN(in_tensors_.size(), 3);
-  CHECK_LESS_RETURN(out_tensors_.size(), 1);
   CHECK_NULL_RETURN(in_tensors_.at(0));
   CHECK_NULL_RETURN(in_tensors_.at(1));
   CHECK_NULL_RETURN(in_tensors_.at(kNumInputDim_2));
diff --git a/mindspore/lite/src/runtime/kernel/arm/fp32_grad/apply_momentum.cc b/mindspore/lite/src/runtime/kernel/arm/fp32_grad/apply_momentum.cc
index 02cd2026abb..1c2f469f1bc 100644
--- a/mindspore/lite/src/runtime/kernel/arm/fp32_grad/apply_momentum.cc
+++ b/mindspore/lite/src/runtime/kernel/arm/fp32_grad/apply_momentum.cc
@@ -100,11 +100,9 @@ int ApplyMomentumCPUKernel::Run() {
 int ApplyMomentumCPUKernel::Init() {
   CHECK_NULL_RETURN(apply_momentum_param_);
   CHECK_LESS_RETURN(in_tensors_.size(), DIMENSION_5D);
-  CHECK_NULL_RETURN(in_tensors_.at(0));
-  CHECK_NULL_RETURN(in_tensors_.at(1));
-  CHECK_NULL_RETURN(in_tensors_.at(2));
-  CHECK_NULL_RETURN(in_tensors_.at(3));
-  CHECK_NULL_RETURN(in_tensors_.at(4));
+  for (int i = 0; i < DIMENSION_5D; i++) {
+    CHECK_NULL_RETURN(in_tensors_.at(i));
+  }
   auto ret = OptimizerKernel::Init();
   if (ret != RET_OK) {
     MS_LOG(ERROR) << "Failed to initialize Apply Momentum Kernel";
diff --git a/mindspore/lite/src/train/optimizer_kernel.h b/mindspore/lite/src/train/optimizer_kernel.h
index 552b805828b..abac6e1f88c 100644
--- a/mindspore/lite/src/train/optimizer_kernel.h
+++ b/mindspore/lite/src/train/optimizer_kernel.h
@@ -34,8 +34,8 @@ static __attribute__((always_inline)) inline bool MS_ISNAN(float var) {
 }
 
 namespace mindspore::kernel {
-
 enum class WeightUpdateMode { NORMAL, VIRTUAL_BATCH, ACCUMULATE_GRADS };
+
 class OptimizerKernel : public InnerKernel {
  public:
   OptimizerKernel() = default;
@@ -107,7 +107,11 @@ class OptimizerKernel : public InnerKernel {
     lite::Tensor *grad_sum_tensor = nullptr;
     if (grad_sum_ != nullptr) {
       auto shape = in_tensors_.at(grad_idx_)->shape();
-      grad_sum_tensor = new lite::Tensor(kNumberTypeFloat, shape);
+      grad_sum_tensor = new (std::nothrow) lite::Tensor(kNumberTypeFloat, shape);
+      if (grad_sum_tensor == nullptr) {
+        MS_LOG(ERROR) << "failed to allocate grad sum tensor";
+        return nullptr;
+      }
       grad_sum_tensor->set_tensor_name(in_tensors_.at(grad_idx_)->tensor_name());
       grad_sum_tensor->set_data(static_cast<void *>(grad_sum_));
       grad_sum_tensor->set_own_data(false);