wisnunugroho21 / asynchronous_impala_ppo Goto Github PK

Hi, I want to ask if I can do like this, I want to add 4 train threads and 1 test thread, so I did like this:

#train thread
runners = [Runner.remote(env_name, training_mode, render, n_update, i) for i in range(n_agent)]
#test thread
runners.append(Runner.remote(args, False, render, n_update, n_agent)) # training_mode = False

But I encountered an error when the test thread read the weight file:

 super(_open_zipfile_reader, self).__init__(torch._C.PyTorchFileReader(name_or_buffer))
RuntimeError: [enforce fail at inline_container.cc:145] . PytorchStreamReader failed reading zip archive: failed finding central directory

I suspect that the learner hasn't saved the weight file completely yet, so the test thread can not load the weight file.
But when I use 5 train threads, everything is fine.

Do you know how come is that?
Thank you very much

Hi,
When I training with a more complicated CNN, I find that the Learner will be slower with each update because the memory of its will increase over time, so I think you should define a max capacity memory for each Runner.

Thanks