Created by: wangkuiyi

I noticed that at https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/net.h#L73, there defines

class PlainNet : public Net {
  void Run(const ScopePtr& scope,
           const platform::DeviceContext& dev_ctx) const override {

and at https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/net.h#L60, there defines

class PlainNet : public Net {
  void InferShape(const ScopePtr& scope) const override {

where ScopePtr is defined at https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/scope.h#L27 as

using ScopePtr = std::shared_ptr<Scope>;

However the choice of smart pointer, e.g., shared_ptr v.s. unique_ptr, depends on the application environment of class Scope. In some cases, we want to keep a singleton/default object and we use unique_ptr<Scope>; and in some other cases, we want auto garbage-collection and we use shared_ptr<Scope>.

In above example, all what PlainNet::InferShape and PlainNet::Run want is const Scope&.

I'd suggest that we remove the assumption about what application code would want a smart pointer and keep the source code in its first-principle form.