To implement a stream inference server, when to call start_stream?

[gaoxt1983]

I want to implement a stream inference server which using grpc protocol, but I'm confused with start_stream, stop_stream and async_stream_infer. Basically, I'm using an instance of python class to hold a triton client instance, but when do I call start_stream? Before EVERY inference call or just call it in init function?

I hope I have explained my doubt, since I am not an English speaker. So endulge my poor English please :(