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
const=True,
default=False)
parser.add_argument('--skip_steps', type=int, default=0)
args = parser.parse_args()
os.environ["CUDA_VISIBLE_DEVICES"] = str(args.gpu)
tf.reset_default_graph()
sess = tf.Session()
# config = Config(args.sz, args.map, lambda _: 1)
config = Config(args.sz, args.map, args.run_id)
os.makedirs('weights/' + config.full_id(), exist_ok=True)
cfg_path = 'weights/%s/config.json' % config.full_id()
config.build(cfg_path if args.restore else args.cfg_path)
if not args.restore and not args.test:
config.save(cfg_path)
envs = EnvWrapper(make_envs(args), config)
agent = A2CAgent(sess, fully_conv, config, args.restore, args.discount,
args.lr, args.vf_coef, args.ent_coef, args.clip_grads,
args.save_best_only, not args.test)
runner = Runner(envs, agent, not args.test, args.steps)
runner.run(args.updates)
if args.save_replay:
envs.save_replay()
envs.close()
weightsList = os.listdir(args.ckptPath)
weightsList.remove('config.json')
weightsList.remove('checkpoint')
weightsList = [oneWeight.split('.')[0] for oneWeight in weightsList]
weightsList = list(set(weightsList))
npyList = os.listdir(args.nextStatePath)
npyList1 = [oneNpy.split('_')[1][:-4] for oneNpy in npyList]
npyList2 = list(set(npyList1))
for oneCkpt in weightsList:
tf.reset_default_graph()
if oneCkpt.split('.')[0][4:] in npyList2:
continue
try:
sess = tf.Session(config=tf_config)
args.ckptfile = oneCkpt
agent = A2CAgent(sess, fully_conv, config, args.discount, args.lr, args.vf_coef, args.ent_coef, args.clip_grads,
weight_dir, log_dir, args)
runner = Runner(None, agent, args, args.steps)
NextStateValue = runner.forwardState()
np.save(args.nextStatePath+'nextState_'+str(oneCkpt.split('-')[1]) + '.npy', NextStateValue)
except:
pass
time.sleep(10)
os._exit(0)
else:
args.ckptfile = None
# begin to train the agent
if not args.distill:
try:
sess = tf.Session(config=tf_config)
parser.add_argument('--clip_grads', type=float, default=1.)
parser.add_argument("--run_id", type=int, default=-1)
parser.add_argument("--game", type=str, default='Pong-v0')
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("--agenttype", type=str, default="vpg") # or you can use "a2c"
args = parser.parse_args()
os.environ["CUDA_VISIBLE_DEVICES"] = str(args.gpu)
tf.reset_default_graph()
sess = tf.Session()
envs = EnvWrapper(make_envs(args))
# not specify which feature to use when initialization
if args.game == 'Pong-v0':
num_action = 3
input_size = 80
agent = A2CAgent(sess, fully_conv, input_size, num_action, args.game, args.restore, args.discount, args.lr, args.vf_coef, args.ent_coef, args.clip_grads, agenttype = agenttype)
runner = Runner(envs, agent, args.steps, args.game, agenttype = agenttype)
runner.run(args.updates, not args.test)
else:
num_action = 2
input_size = None
agent = A2CAgent(sess, carpole_net, input_size, num_action, args.game, args.restore, args.discount, args.lr,
args.vf_coef, args.ent_coef, args.clip_grads)
runner = Runner(envs, agent, args.steps, args.game)
runner.run(args.updates, not args.test)
envs.close()