def main(unused_argv):
config = Config(FLAGS.sz, FLAGS.map, -1) # 进行参数的设置
os.makedirs('weights/' + config.full_id(), exist_ok=True)
cfg_path = 'weights/%s/config.json' % config.full_id() # 保存参数的位置
config.build(cfg_path) # 建立和设置参数
if FLAGS.human:
human()
else:
agent(config)
parser.add_argument("--map", type=str, default='MoveToBeacon')
parser.add_argument("--cfg_path", type=str, default='config.json.dist')
parser.add_argument("--restrict", type=bool, default=False)
parser.add_argument("--imitation", type=bool, default=False)
parser.add_argument("--test", type=bool, nargs='?', const=True, default=False)
parser.add_argument("--restore", type=bool, nargs='?', const=True, default=False)
parser.add_argument('--save_replay', type=bool, nargs='?', const=True, default=False)
args = parser.parse_args()
os.environ["CUDA_VISIBLE_DEVICES"] = str(args.gpu) #environ是一个字符串所对应环境的映像对象，输入为要使用的GPU number
tf.reset_default_graph()
sess = tf.Session()
# config = Config(args.sz, args.map, lambda _: 1)
config = Config(args.sz, args.map, args.run_id, restrict=args.restrict, imitation=args.imitation) # 进行参数的设置
os.makedirs('weights/' + config.full_id(), exist_ok=True)
cfg_path = 'weights/%s/config.json' % config.full_id() # 保存参数的位置
config.build(cfg_path if args.restore else args.cfg_path) # 建立和设置参数
if not args.restore and not args.test:
config.save(cfg_path) # 保存参数
envs = EnvWrapper(make_envs(args), config) # 创建环境,封装一层
agent = A2CAgent(sess, fully_conv, config, args.restore, args.discount, args.lr, args.vf_coef, args.ent_coef, args.clip_grads) # 创建agent
runner = Runner(envs, agent, args.steps) # 创建进程
runner.run(args.updates, not args.test) # 开始运行
if args.save_replay: #是否保存回放
envs.save_replay()
envs.close() # 关闭环境
# preprocess args
if args.maps == '':
args.maps = '{"' + args.map + '": ' + str(args.envs) + '}'
args.maps = json.loads(args.maps, object_pairs_hook=OrderedDict)
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '3'
os.environ["CUDA_VISIBLE_DEVICES"] = str(args.gpu)
# Allow gpu growth
tf_config = tf.ConfigProto()
tf_config.gpu_options.allow_growth = True
config = Config(args.sz, '_'.join(args.maps), args.run_id)
weight_dir = os.path.join(args.work_dir, 'weights', config.full_id())
log_dir = os.path.join(args.work_dir, 'logs', config.full_id())
print('weights are saved at ', weight_dir)
os.makedirs(weight_dir, exist_ok=True)
cfg_path = os.path.join(weight_dir, 'config.json')
if args.restore and (not os.path.isfile(os.path.join(weight_dir, 'checkpoint'))) and args.remote_restore != '':
args.restore_path = args.remote_restore
config.build(os.path.join(args.restore_path, 'config.json'))
config.save(cfg_path)
assert os.path.isfile(os.path.join(args.restore_path, 'checkpoint'))
elif args.restore:
if os.path.isfile(cfg_path):
config.build(cfg_path)
else:
config.build(args.cfg_path)
args.action_dim = opt_dim
# max_step from float --> int
if len(args.map) > 5 and (args.map.startswith('Build')):
args.max_step = int(args.num_timesteps / args.step_mul)
else:
args.max_step = int(args.num_timesteps / args.step_mul)
args.game_steps_per_episode = args.max_step * args.step_mul
# pytorch device setup
processor = 'cpu' if args.eval or args.gpu == -1 else 'cuda:%s'%args.gpu
device = torch.device(processor)
data_format = "NHWC" if args.nhwc else "NCHW"
config = Config(args, device, data_format)
os.makedirs('weights/' + config.full_id(), exist_ok=True)
with open('weights/' + config.full_id() + '/arguments.txt', 'w') as f:
for key,value in vars(args).items():
string = str(key) + ' : ' + str(value) + '\n'
f.write( string )
# make/clean directory for save mode
dirname = os.path.join('results', args.map, args.meta, 'run_%d'%(args.run_id))
if args.save_ilp or (args.meta in ['random', 'hrl'] and not args.eval):
if os.path.exists(dirname):
print("The savepath already exists. Deleting {}".format(dirname))
shutil.rmtree(dirname)
os.makedirs(dirname, exist_ok=True)
# set up agent & ilp
ilp = None