After tflite invoke, the tflite model use a larger memory

I try to use the tflite model via c++ and deploy it on mobile.
Here is the code example:

void Model::Inference(const std::vector<int> &inputs,
                      std::vector<int> &outputs){
  int size = inputs.size();
  const std::vector<int> input_size = {size};
  std::cout<<"model memo 1: "<<physical_memory_used_by_process_temp()<<std::endl;
  interpreter_->ResizeInputTensorStrict(0, input_size);
  std::cout<<"model memo 2: "<<physical_memory_used_by_process_temp()<<std::endl;
  int* input_tensor = interpreter_->typed_input_tensor<int>(0);
  for(int i = 0; i < input_size[0]; i++){
    input_tensor[i] = inputs[i];
  std::cout<<"model memo 3: "<<physical_memory_used_by_process_temp()<<std::endl;
  std::cout<<"model memo 4: "<<physical_memory_used_by_process_temp()<<std::endl;
  int* output_tensor = interpreter_->typed_output_tensor<int>(0);
  std::cout<<"model memo 5: "<<physical_memory_used_by_process_temp()<<std::endl;
  auto output_size = interpreter_->output_tensor(0)->dims->data[0];
  for(int i = 0; i < output_size; i++){
    if(output_tensor[i] == end_token_id_)

The model file size is 85MB. The memory usage is only 140MB when the model is loaded. After the interpreter_->Invoke() , the memory usage increase to 200 or even 250 mb and never go down. Is this normal? Is there any way to release the increased memory usage?

Any help? I’ve been struggling with this issue for a few days. :smiley: :smiley: :smiley:

??? Any help???