def initialize(self):
args = self.args
co = minirts.ContextOptions()
self.context_params.initialize(co)
opt = minirts.Options()
opt.seed = args.seed
opt.frame_skip_ai = args.fs_ai
opt.frame_skip_opponent = args.fs_opponent
opt.simulation_type = minirts.ST_NORMAL
opt.ai_type = getattr(minirts, args.ai_type)
if args.ai_type == "AI_NN":
opt.backup_ai_type = minirts.AI_SIMPLE
if args.ai_type == "AI_FLAG_NN":
opt.backup_ai_type = minirts.AI_FLAG_SIMPLE
opt.opponent_ai_type = getattr(minirts, args.opponent_type)
opt.latest_start = args.latest_start
opt.latest_start_decay = args.latest_start_decay
opt.mcts_threads = args.mcts_threads
opt.mcts_rollout_per_thread = 50
opt.max_tick = args.max_tick
opt.handicap_level = args.handicap_level
opt.simple_ratio = args.simple_ratio
opt.ratio_change = args.ratio_change
# opt.output_filename = b"simulators.txt"
# opt.cmd_dumper_prefix = b"cmd-dump"
# opt.save_replay_prefix = b"replay"
GC = minirts.GameContext(co, opt)
print("Version: ", GC.Version())
num_action = GC.get_num_actions()
print("Num Actions: ", num_action)
num_unittype = GC.get_num_unittype()
print("Num unittype: ", num_unittype)
desc = []
name2idx = {}
# For actor model: group 0
# No reward needed, we only want to get input and return distribution of actions.
# sampled action and and value will be filled from the reply.
name2idx["actor"] = len(desc)
desc.append((dict(id="",
s=str(2),
r0="",
r1="",
last_r="",
last_terminal="",
_batchsize=str(args.batchsize),
_T="1"),
dict(rv="",
pi=str(num_action),
V="1",
a="1",
_batchsize=str(args.batchsize),
_T="1")))
if not args.actor_only:
# For training: group 1
# We want input, action (filled by actor models), value (filled by actor
# models) and reward.
name2idx["train"] = len(desc)
desc.append((dict(rv="",
id="",
pi=str(num_action),
s=str(2),
r0="",
r1="",
a="1",
r="1",
V="1",
seq="",
terminal="",
_batchsize=str(args.batchsize),
_T=str(args.T)), None))
inputs, replies = utils_elf.init_collectors(GC,
co,
desc,
use_numpy=False)
params = dict(
num_action=num_action,
num_unit_type=num_unittype,
num_group=1 if args.actor_only else 2,
action_batchsize=int(desc[name2idx["actor"]][0]["_batchsize"]),
train_batchsize=int(desc[name2idx["train"]][0]["_batchsize"])
if not args.actor_only else None,
T=args.T)
return utils_elf.GCWrapper(GC, inputs, replies, name2idx, params)
def initialize(self):
args = self.args
co = atari.ContextOptions()
self.context_args.initialize(co)
opt = atari.Options()
opt.frame_skip = args.frame_skip
opt.rom_file = os.path.join(args.rom_dir, args.rom_file)
opt.seed = 42
opt.hist_len = args.hist_len
opt.reward_clip = args.reward_clip
GC = atari.GameContext(co, opt)
print("Version: ", GC.Version())
num_action = GC.get_num_actions()
print("Num Actions: ", num_action)
desc = []
name2idx = {}
# For actor model: group 0
# No reward needed, we only want to get input and return distribution of actions.
# sampled action and and value will be filled from the reply.
name2idx["actor"] = len(desc)
# Descriptor of each group.
# desc = [(input_group0, reply_group0), (input_group1, reply_group1), ...]
# GC.Wait(0) will return a batch of game states in the same group.
# For example, if you register group 0 as the actor group, which has the history length of 1, and group 1 as the optimizer group which has the history length of T
# Then you can check the group id to decide which Python procedure to use to deal with the group, by checking their group_id.
# For self-play, we can register one group each player.
desc.append((dict(id="",
s=str(args.hist_len),
last_r="",
last_terminal="",
_batchsize=str(args.batchsize),
_T="1"),
dict(rv="",
pi=str(num_action),
V="1",
a="1",
_batchsize=str(args.batchsize),
_T="1")))
if not args.actor_only:
# For training: group 1
# We want input, action (filled by actor models), value (filled by actor
# models) and reward.
name2idx["train"] = len(desc)
desc.append((dict(rv="",
id="",
pi=str(num_action),
s=str(args.hist_len),
a="1",
r="1",
V="1",
seq="",
terminal="",
_batchsize=str(args.batchsize),
_T=str(args.T)), None))
# Initialize shared memory (between Python and C++) based on the specification defined by desc.
inputs, replies = utils_elf.init_collectors(GC,
co,
desc,
use_numpy=False)
params = dict()
params["num_action"] = GC.get_num_actions()
params["num_group"] = 1 if args.actor_only else 2
params["action_batchsize"] = int(
desc[name2idx["actor"]][0]["_batchsize"])
if not args.actor_only:
params["train_batchsize"] = int(
desc[name2idx["train"]][0]["_batchsize"])
params["hist_len"] = args.hist_len
params["T"] = args.T
return utils_elf.GCWrapper(GC, inputs, replies, name2idx, params)
def initialize(self):
args = self.args
co = atari.ContextOptions()
self.context_params.initialize(co)
opt = atari.Options()
opt.frame_skip = args.frame_skip
opt.rom_file = os.path.join(args.rom_dir, args.rom_file)
opt.seed = 42
opt.hist_len = args.hist_len
opt.reward_clip = args.reward_clip
GC = atari.GameContext(co, opt)
print("Version: ", GC.Version())
num_action = GC.get_num_actions()
print("Num Actions: ", num_action)
desc = []
name2idx = {}
# For actor model: group 0
# No reward needed, we only want to get input and return distribution of actions.
# sampled action and and value will be filled from the reply.
name2idx["actor"] = len(desc)
desc.append((dict(id="",
s=str(args.hist_len),
last_r="",
last_terminal="",
_batchsize=str(args.batchsize),
_T="1"),
dict(rv="",
pi=str(num_action),
V="1",
a="1",
_batchsize=str(args.batchsize),
_T="1")))
if not args.actor_only:
# For training: group 1
# We want input, action (filled by actor models), value (filled by actor
# models) and reward.
name2idx["train"] = len(desc)
desc.append((dict(rv="",
id="",
pi=str(num_action),
s=str(args.hist_len),
a="1",
r="1",
V="1",
seq="",
terminal="",
_batchsize=str(args.batchsize),
_T=str(args.T)), None))
inputs, replies = utils_elf.init_collectors(GC,
co,
desc,
use_numpy=False)
params = dict()
params["num_action"] = GC.get_num_actions()
params["num_group"] = 1 if args.actor_only else 2
params["action_batchsize"] = int(
desc[name2idx["actor"]][0]["_batchsize"])
if not args.actor_only:
params["train_batchsize"] = int(
desc[name2idx["train"]][0]["_batchsize"])
params["hist_len"] = args.hist_len
params["T"] = args.T
return utils_elf.GCWrapper(GC, inputs, replies, name2idx, params)