-
Notifications
You must be signed in to change notification settings - Fork 9
Description
Hello, thank you for your work. I ran this code, but found that memory keeps going up when I calculate the same picture iteratively.
In addition, I found that the speed of CPU single-core single-threading is: Detect 690X460 Time Using CPU: 170, which is about twice slower than the c_api version of MTCNN (of course, speed is not so important if memory problems can be solved).
This function has obvious problems:
inline const std::vector<Blob*> output_blobs(){
out_blobs.erase( out_blobs.begin(),out_blobs.end() );
for (auto &output:executor_->outputs) {
Blob *data=new Blob(output.GetShape(),Context::cpu());
output.SyncCopyToCPU(data->cpu_data(),data->Size());
out_blobs.push_back(data);
}
return out_blobs;
}The new Blob pointer is stored in the vector, and the erase vector's content only destroys the pointer but does not release the memory pointed to by the pointer.
I try to delete all the pointers in vector before erase, but core dumped. It is estimated that the pointer is still used by other functions, so I'm afraid the whole code structure needs to be modified.