diff --git a/src/fpga/api.cpp b/src/fpga/api.cpp
index 10787b915594a12a826a087e5453b2c2e8c03f9a..f10aee5014d8f377ecc8e1735276aebf6418436f 100644
--- a/src/fpga/api.cpp
+++ b/src/fpga/api.cpp
@@ -311,6 +311,8 @@ int get_aligned_filter_num(int num) {
 
 void format_filter(framework::Tensor *filter_tensor, float max_value,
                    int group_num) {
+  filter_tensor->scale[0] = float(max_value / 127.0);
+  filter_tensor->scale[1] = float(127.0 / max_value);
   auto dims = filter_tensor->dims();
   auto num = dims[0], channel = dims[1], height = dims[2], width = dims[3];
   auto data_ptr = filter_tensor->data<float>();
diff --git a/src/io/executor.cpp b/src/io/executor.cpp
index c8d8f52a427bb1ee2b9fa04c9ef09f8e626f11b0..d733231ef03f74eba2f1f2e989a0bad1cf43f161 100644
--- a/src/io/executor.cpp
+++ b/src/io/executor.cpp
@@ -676,11 +676,11 @@ std::shared_ptr<framework::Tensor> Executor<Dtype, P>::FetchResult(int id) {
       to_predict_program_->Block(0);
   auto &ops = ops_of_block_[*to_predict_block.get()];
 
-  PADDLE_MOBILE_ENFORCE(id < ops.size(), "Index out of range");
-  auto last_op = id < 0 ? ops[ops.size() - 1] : ops[id];
-  auto output_map = last_op->Outputs();
-  std::vector<std::string> out_keys = last_op->GetOutKeys();
-  PADDLE_MOBILE_ENFORCE(!out_keys.empty(), "the last op contains no output");
+  PADDLE_MOBILE_ENFORCE(id < (int)ops.size(), "Index out of range");
+  auto op = id < 0 ? ops[ops.size() - 1] : ops[id];
+  auto output_map = op->Outputs();
+  std::vector<std::string> out_keys = op->GetOutKeys();
+  PADDLE_MOBILE_ENFORCE(!out_keys.empty(), "this op contains no output");
   auto *output_tensor = framework::GetVarValue<framework::LoDTensor>(
       out_keys[0], output_map, *(program_.scope));
   return std::make_shared<framework::Tensor>(framework::Tensor(*output_tensor));
diff --git a/src/operators/feed_op.h b/src/operators/feed_op.h
index cccd4f52ebdc368e4f68eaf9dc3f25ee3693fdd2..1b36461932798153af60d936dbac91817a4100df 100644
--- a/src/operators/feed_op.h
+++ b/src/operators/feed_op.h
@@ -50,8 +50,8 @@ class FeedOp : public framework::OperatorBase<DeviceType> {
 
   void RunImpl() const {
     auto input = (Tensor *)const_cast<LoDTensor *>(param_.InputX());
-    auto input_ptr = input->data<float>();
     fpga::format_image(input);
+    auto input_ptr = input->data<float>();
     Tensor *output = param_.Out();
     auto output_ptr = output->data<float>();
 
diff --git a/src/operators/kernel/fpga/concat_kernel.cpp b/src/operators/kernel/fpga/concat_kernel.cpp
index 86da2833ed6e1443707054896127e87c0ca297b9..1c48ef021945e6a7b8b53ee946a33b862766deeb 100644
--- a/src/operators/kernel/fpga/concat_kernel.cpp
+++ b/src/operators/kernel/fpga/concat_kernel.cpp
@@ -47,7 +47,7 @@ bool ConcatKernel<FPGA, float>::Init(ConcatParam<FPGA> *param) {
   concatArgs.image_num = (uint32_t)image_num;
   concatArgs.images_in = images_in;
   concatArgs.scales_in = scales_in;
-  concatArgs.image_out = (half *)out->mutable_data<float>();
+  concatArgs.image_out = (half *)out->data<float>();
   concatArgs.scale_out = out->scale;
   concatArgs.channel_num = channel_num;
   concatArgs.height = (uint32_t)height;
diff --git a/src/operators/kernel/fpga/softmax_kernel.cpp b/src/operators/kernel/fpga/softmax_kernel.cpp
index 7cfd0c7d76c1a8e73955dbec1971d86ceebde259..ef68cc3c33fdc4c0a8537cbb1dd3a49583c6c8b1 100644
--- a/src/operators/kernel/fpga/softmax_kernel.cpp
+++ b/src/operators/kernel/fpga/softmax_kernel.cpp
@@ -39,8 +39,8 @@ bool SoftmaxKernel<FPGA, float>::Init(SoftmaxParam<FPGA> *param) {
   args.image.height = 1;
   args.image.width = 1;
   args.image.channels = (uint32_t)input->dims()[1];
-  args.output.address = float_input->mutable_data<float>();
-
+  args.output.address = float_input->data<float>();
+  args.output.scale_address = float_input->scale;
   param->SetFloatInput(float_input);
   param->SetFpgaArgs(args);
   return true;