diff --git a/paddle/fluid/lite/kernels/arm/split_compute_test.cc b/paddle/fluid/lite/kernels/arm/split_compute_test.cc
index 7d2d95205fc58a092be5714720c92382a9883a17..39632bee8decfe875f0adb3c2717d58e593c400b 100644
--- a/paddle/fluid/lite/kernels/arm/split_compute_test.cc
+++ b/paddle/fluid/lite/kernels/arm/split_compute_test.cc
@@ -27,7 +27,7 @@ void splite_resize_out(const lite::Tensor* din,
                        const std::vector<lite::Tensor*>& dout, int axis,
                        int num, const std::vector<int>& sections) {
   auto in_dims = din->dims();
-  int outs_number;
+  int outs_number = dout.size();
 
   std::vector<lite::DDimLite> outs_dims;
   outs_dims.reserve(outs_number);
@@ -118,11 +118,11 @@ TEST(split_arm, compute) {
                 for (int i = 0; i < x.dims().production(); i++) {
                   x_data[i] = i;
                 }
-
                 for (auto out : output) delete out;
                 for (auto out : output_ref) delete out;
                 output.clear();
                 output_ref.clear();
+
                 int outs_number;
                 if (num > 0) {
                   outs_number = num;
@@ -133,7 +133,6 @@ TEST(split_arm, compute) {
                   output.push_back(new lite::Tensor);
                   output_ref.push_back(new lite::Tensor);
                 }
-
                 splite_resize_out(&x, output, axis, num, sections);
                 splite_resize_out(&x, output_ref, axis, num, sections);
                 param.x = &x;
diff --git a/paddle/fluid/lite/operators/split_op.cc b/paddle/fluid/lite/operators/split_op.cc
index 299be81fc0b216b9b436866ea76ff4ce04051a96..9b4b7662ab7ba7228ee215bf051601150e2b6bb7 100644
--- a/paddle/fluid/lite/operators/split_op.cc
+++ b/paddle/fluid/lite/operators/split_op.cc
@@ -31,7 +31,7 @@ bool SplitOp::CheckShape() const {
 
 bool SplitOp::InferShape() const {
   const auto &outs = param_.output;
-  auto in_dims = param_.x.dims();
+  auto in_dims = param_.x->dims();
   int axis = param_.axis;
   int num = param_.num;
   const auto &sections = param_.sections;
@@ -68,7 +68,7 @@ bool SplitOp::AttachImpl(const cpp::OpDesc &opdesc, lite::Scope *scope) {
   param_.sections = opdesc.GetAttr<std::vector<int>>("sections");
   param_.x = const_cast<lite::Tensor *>(
       &scope->FindVar(opdesc.Input("X").front())->Get<lite::Tensor>());
-  auto outs = op_desc.Output("Out").front;
+  auto outs = opdesc.Output("Out");
   for (auto var : outs) {
     param_.output.push_back(scope->FindVar(var)->GetMutable<lite::Tensor>());
   }
diff --git a/paddle/fluid/lite/operators/split_op.h b/paddle/fluid/lite/operators/split_op.h
index 177c44171e6e67214f820f04e801be6c01df01cc..20dc4b1028c27f4efab558694285e44d46182ef8 100644
--- a/paddle/fluid/lite/operators/split_op.h
+++ b/paddle/fluid/lite/operators/split_op.h
@@ -23,7 +23,7 @@ namespace paddle {
 namespace lite {
 namespace operators {
 
-class SoftmaxOp : public OpLite {
+class SplitOp : public OpLite {
  public:
   SplitOp() {}
   explicit SplitOp(const std::string &op_type) : OpLite(op_type) {}