README.md

Overview

dipcc is a Diplomacy game engine written from scratch in c++. It aims to be a faster alternative to the MILA Diplomacy Engine with a similar API.

dipcc also provides a python interface pydipcc via pybind11. The ground truth python API is at dipcc/pybind/pybind.cc

Build

Run ./compile.sh, or from the repo root run make dipcc

To compile with verbose adjudicator logging, run MODE=DEBUG ./compile.sh

JSON Encoding/Decoding

The biggest API change from the MILA engine to dipcc is that the json encoding/decoding functions operate on strings, and the json processing is done by dipcc.

New dipcc API:

with open("/path/to/game.json", "r") as f:
    s = f.read()  # string

game = pydipcc.Game.from_json(s)

json_string = game.to_json()

Old API:

with open("/path/to/game.json", "r") as f:
    d = json.load(f)  # python dict

game = diplomacy.Game.from_saved_game_format(d)

python_dict = diplomacy.utils.export.to_saved_game_format(game)

pydipcc.ThreadPool

A c++ thread pool allows for parallelizing some of the slower methods of dipcc when operating concurrently on multiple Game objects. The ground truth python API is at dipcc/pybind/pybind.cc

The main methods are:

pydipcc.ThreadPool.process_multi(List[Game]): calls game.process() on each of the games passed to it

pydipcc.ThreadPool.encode_inputs_multi(List[Game], <data buffers>): more easily called from encode_batch_inputs(ThreadPool, List[Game]) -> List[DataFields] in model_sampled_agent.py, this produces the tensors that are provided to the no-text pytorch model.

pydipcc.ThreadPool.decode_order_idxs(order_idxs): converts a LongTensor of order vocabulary idxs to their corresponding order strings.