def selfplay_thread():
nonlocal net, temperature, num_games
torch.set_num_threads(1)
with torch.no_grad():
while True:
if net.metadata["iteration"] == 1 and options.get(
"fast_first_iteration"):
game_options = {
**options,
"zero_value": True, # ignore value net
"num_visits": 10,
"detailed_visits_prob": 1.0,
"kl_surprise_weights": False,
}
temp_fn = lambda mv: 1.0
else:
game_options = options
temp_fn = temperature
game_states = play_game(
net_evaluator=evaluate,
game_class=game_class,
temperature=temp_fn,
**game_options,
)
if game_states:
trainer.add_sample.remote(game_states)
num_games += 1
def selfplay_proc(cpu, game_class, options):
num_games_before_check = 5
temperature = lambda mv: 1.0 if mv < 4 else 0.1 # selfplay param
net = game_class.create_net(cuda=True, **options)
print(net.device)
with torch.no_grad():
for iter in range(5):
start = time.time()
samples = 0
for i in range(num_games_before_check):
game_states, endstate = play_game(net,
game_class,
temperature=temperature,
**options)
samples += len(game_states)
print(
f"CPU {cpu} self-play generated {samples} samples in {time.time()-start:.1f}s"
)
def selfplay_thread(tid):
nonlocal games_played
torch.set_num_threads(1)
games_played = 0
with torch.no_grad():
while True:
start = time.time()
samples = 0
game_states, endstate = play_game(
net_evaluator=evaluate,
game_class=game_class,
temperature=temperature,
**options,
)
samples += len(game_states)
dt = time.time() - start
# games_q.put((game_states, endstate))
games_played += 1
print(
f"[{games_played}] CPU {cpu} thread {tid} self-play generated {samples} samples (out of {endstate['end_move']} moves) in {dt:.1f}s"
)
def test_play():
game = GoMokuState
net = game.create_net()
play_game(net, game, verbose=True)
parser = argparse.ArgumentParser(description="Self-play visualization.")
parser.add_argument("--game", type=str, help="Game to play")
parser.add_argument("--tag", type=str, help="Tag for experiment", default="")
args = parser.parse_args()
game = args.game
if game == "cg":
game_class = CaptureGoState
elif game == "pxcg":
game_class = PixelCaptureGoState
elif game == "nim":
game_class = NimState
elif game == "oth":
game_class = OthelloState
else:
raise Exception("unknown game")
net = game_class.create_net(tag=args.tag)
options = {}
print(f"Loaded net {net.metadata['filename']} on cuda? {net.device}")
temp_fn = lambda mv: 1.0 if mv < 2 else 0.1
with torch.no_grad():
game_states = play_game(
net_evaluator=net.evaluate_sample,
game_class=game_class,
temperature=temp_fn,
verbose=True,
)
players.append(cls.create_net(net_ts=ts, tag=tag, cuda=False))
except Exception as e:
print(e)
print(len(players), "players loaded")
options = {"num_visits": 1, "cpuct": 1.5, "force_win": True}
options = {"num_visits": 1, "cpuct": 1.1}
elocalc = BayesElo(players)
for p1 in tqdm(players, ascii=True):
for p2 in tqdm(players, ascii=True):
if p1 is not p2:
for _ in range(num_games):
game_states, endstate = play_game([p1, p2], game_class, temperature=temp_fn, **options)
result = endstate["value"][0] - endstate["value"][1]
training_samples = play_game(
net_evaluator=[p1.evaluate_sample, p2.evaluate_sample],
game_class=game_class,
temperature=temp_fn,
**options
)
v = training_samples[-1]["value"]
result = v[0] - v[1]
elocalc.add_result(p1, p2, result)
df, aux = elocalc.summary_df()
pd.set_option("display.max_rows", 500)
df, aux = elocalc.summary_df(aux_zero=True)