{
"info": {
"author": "Tim Zaman",
"author_email": "timbobel@gmail.com",
"bugtrack_url": null,
"classifiers": [
"Intended Audience :: Education",
"Intended Audience :: Science/Research",
"License :: Freeware",
"License :: Other/Proprietary License",
"Operating System :: MacOS",
"Operating System :: Unix",
"Programming Language :: Python :: 3.7",
"Programming Language :: Python :: 3 :: Only",
"Topic :: Scientific/Engineering",
"Topic :: Scientific/Engineering :: Artificial Intelligence",
"Topic :: Software Development",
"Topic :: Software Development :: Libraries",
"Topic :: Software Development :: Libraries :: Python Modules"
],
"description": "# DotaService\n\n![\"dotaservice](\"dotaservice-icon.png\")
\n\n---\n\nNOTE: The project that uses the dotaservice in a k8s environment is the [DotaClient](https://github.com/TimZaman/dotaclient) repo.\n\nDotaService is a service to play Dota 2 through gRPC. There are first class python bindings\nand examples, so you can play dota as you would use the OpenAI gym API.\n\nIt's fully functional and super lightweight. Starting Dota `obs = env.reset()` takes 5 seconds,\nand each `obs = env.step(action)` in the environment takes between 10 and 30 ms.\n\nYou can even set the config of `render=True` and you can watch the game play live. Each game will\nhave a uuid and folder associated where there's a Dota demo (replay) and console logs.\n\n![\"demo\"](\"demo.gif\")
\n\n## Run DotaService Locally\n\nRun the DotaService so you can connect your client to it later. Only one client per server\nis supported, and only one DotaService per VM (eg local or one per docker container).\n\n```sh\npython3 -m dotaservice\n>>> Serving on 127.0.0.1:13337\n```\n\n## Run DotaService Distributed\n\nSee [docker/README.md](docker/README.md).\n\nTo run two dockerservice instances, one on port `13337` and one on `13338`, f.e. run:\n\n```sh\ndocker run -dp 13337:13337 ds\ndocker run -dp 13338:13337 ds\n```\n\nYou can run as many as you want, until you run out of ports or ip addresses. If you are wearing\nyour fancy pants, use Kubernetes to deploy gazillions.\n\n## Client Code\n\n```py\nfrom grpclib.client import Channel\nfrom protobuf.DotaService_grpc import DotaServiceStub\nfrom protobuf.DotaService_pb2 import Action\nfrom protobuf.DotaService_pb2 import Config\n\n# Connect to the DotaService.\nenv = DotaServiceStub(Channel('127.0.0.1', 13337))\n\n# Get the initial observation.\nobservation = await env.reset(Config())\nfor i in range(8):\n # Sample an action from the action protobuf\n action = Action.MoveToLocation(x=.., y=.., z=..)\n # Take an action, returning the resulting observation.\n observation = await env.step(action)\n```\n\nThis is very useful to provide an environment for reinforcement learning, and service aspect of it makes it\nespecially useful for distributed training. I am planning to provide a client python\nmodule for this (`PyDota`) that mimics typical OpenAI gym APIs. Maybe I won't even make PyDota\nand the gRPC client is enough.\n\n\n
![\"dotaservice](\"dotaservice.png\")
\n