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.eval_only = getattr(args, "env_eval_only", 0) == 1
opt.hist_len = args.hist_len
opt.reward_clip = args.reward_clip
GC = atari.GameContext(co, opt)
print("Version: ", GC.Version())
params = GC.GetParams()
print("Num Actions: ", params["num_action"])
desc = {}
# For actor model, 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.
desc["actor"] = dict(batchsize=args.batchsize,
input=dict(T=1,
keys=set(
["s", "last_r", "last_terminal"])),
reply=dict(T=1, keys=set(["rv", "pi", "V", "a"])))
if not args.actor_only:
# For training: group 1
# We want input, action (filled by actor models), value (filled by actor
# models) and reward.
desc["train"] = dict(batchsize=args.batchsize,
input=dict(T=args.T,
keys=set([
"rv", "id", "pi", "s", "a",
"last_r", "V", "seq",
"last_terminal"
])),
reply=None)
if args.additional_labels is not None:
extra = args.additional_labels.split(",")
for _, v in desc.items():
v["input"]["keys"].update(extra)
# Initialize shared memory (between Python and C++) based on the specification defined by desc.
params["num_group"] = 1 if args.actor_only else 2
params["action_batchsize"] = desc["actor"]["batchsize"]
if not args.actor_only:
params["train_batchsize"] = desc["train"]["batchsize"]
params["hist_len"] = args.hist_len
params["T"] = args.T
return GCWrapper(GC, co, desc, use_numpy=False, params=params)
def benchmark(ngame):
batchsize = 16
co, opt = initialize_game(
batchsize, num_games=ngame,
frame_skip=frame_skip, T=4)
GC = atari.GameContext(co, opt)
start = time.time()
sys.stderr.write("Start benchmark ngame={}...\n".format(ngame))
stats = Counter()
for i in range(ngame): stats[i] = 0
for k in range(Niter):
infos = GC.Wait(0)
#for i in range(len(infos)):
# stats[infos[i].meta().id] += 1
GC.Steps(infos)
print(stats)
return Niter / (time.time() - start) * frame_skip * batchsize
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_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.eval_only = getattr(args, "env_eval_only", 0) == 1
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 = {}
# For actor model, 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.
desc["actor"] = dict(input=dict(id="",
s=str(args.hist_len),
last_r="",
last_terminal="",
_batchsize=str(args.batchsize),
_T="1"),
reply=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.
desc["train"] = dict(input=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)),
reply=None)
# Initialize shared memory (between Python and C++) based on the specification defined by desc.
params = dict()
params["num_action"] = GC.get_num_actions()
params["num_group"] = 1 if args.actor_only else 2
params["action_batchsize"] = int(desc["actor"]["input"]["_batchsize"])
if not args.actor_only:
params["train_batchsize"] = int(
desc["train"]["input"]["_batchsize"])
params["hist_len"] = args.hist_len
params["T"] = args.T
return GCWrapper(GC, co, desc, use_numpy=False, params=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)