def main():
# Use alphazero self-play for data generation
agents_meta = parse_schedule()
# worker variable of main process
board = Board()
sigexit = Event()
sigexit.set() # pre-set signal so main proc generator will iterate only once
# subprocess data generator
helper = DataHelper(data_files=[])
helper.set_agents_meta(agents_meta=agents_meta)
generator = helper.generate_batch(TRAINING_CONFIG["batch_size"])
# start generating
with h5py.File(f"{DATA_CONFIG['data_path']}/latest.train.hdf5", 'a') as hf:
for state_batch, value_batch, probs_batch in generator:
for batch_name in ("state_batch", "value_batch", "probs_batch"):
if batch_name not in hf:
shape = locals()[batch_name].shape
hf.create_dataset(batch_name, (0, *shape), maxshape=(None, *shape))
hf[batch_name].resize(hf[batch_name].shape[0] + 1, axis=0)
hf[batch_name][-1] = locals()[batch_name]
# prevent main proc from generating data too quick
# since sigexit has been set, proc will iterate only once
run_proc(helper.buffer, helper.buffer_size, helper.lock,
sigexit, agents_meta, board)
board.reset()
def main():
# Use alphazero self-play for data generation
agents_meta = parse_schedule()
# worker variable of main process
board = Board()
sigexit = Event()
sigexit.set(
) # pre-set signal so main proc generator will iterate only once
# subprocess data generator
helper = DataHelper(data_files=[])
helper.set_agents_meta(agents_meta=agents_meta)
generator = helper.generate_batch(TRAINING_CONFIG["batch_size"])
# start generating
with h5py.File(f"{DATA_CONFIG['data_path']}/latest.train.hdf5", 'a') as hf:
for state_batch, value_batch, probs_batch in generator:
for batch_name in ("state_batch", "value_batch", "probs_batch"):
if batch_name not in hf:
shape = locals()[batch_name].shape
hf.create_dataset(batch_name, (0, *shape),
maxshape=(None, *shape))
hf[batch_name].resize(hf[batch_name].shape[0] + 1, axis=0)
hf[batch_name][-1] = locals()[batch_name]
# prevent main proc from generating data too quick
# since sigexit has been set, proc will iterate only once
run_proc(helper.buffer, helper.buffer_size, helper.lock, sigexit,
agents_meta, board)
board.reset()
def dual_play(agents, board=None, verbose=False, graphic=False):
"""
Play with 2 players.
Params:
agents: { Player.black: agent1, Player.white: agent2 }.
board: initial board state. Start player will be determined here.
verbose: if true, then return value will be in the form of training data.
Returns:
if verbose set to True:
[(state_inputs, final_score, action_probs)]
Each element is a numpy.array.
else:
winner
"""
if board is None:
board = Board()
elif board.status["is_end"]:
board.reset()
if verbose is True:
result = []
else:
result = Player.none
while True:
# set the current agent
cur_agent = agents[board.status["cur_player"]]
# evaluate board state and get action
if verbose is True:
_, action_probs, next_move = cur_agent.eval_state(board)
result.append([
board.encoded_states(),
board.status["cur_player"],
action_probs
])
else:
next_move = cur_agent.get_action(board)
# update board
board.apply_move(next_move)
if graphic:
print(board)
# end judge
if board.status["is_end"]:
winner = board.status["winner"]
if graphic:
print("Game ends. winner is {}.".format(winner))
# format output result
if verbose is True:
result = [(
state[0],
np.array(Player.calc_score(state[1], winner)),
state[2]
) for state in result]
else:
result = winner
return result
def eval_agents(agents, num_games=9):
"""
Eval the performance of two agents by multiple game simulation.
Params:
agents: [agent1, agent2]
num_games: number of games simulated, default to BO9.
board: a pre-init board can be passed to avoid re-construct.
Returns:
[win_rate(a) for a in agents]
"""
print("---------Evaluating agents-------------")
board = Board()
players = [Player.black, Player.white]
win_cnts = np.zeros(2)
for i in range(num_games):
winner = dual_play(dict(zip(players, agents)), board)
try:
win_idx = players.index(winner)
win_cnts[win_idx] += 1
print("Round {} ends, winner is <{}: {}>;".format(i + 1, winner, agents[win_idx]))
except ValueError: # tie
win_cnts += 0.5
print("Round {} ends, tie game;".format(i + 1))
players.reverse() # exchange the start player
board.reset()
[agent.reset() for agent in agents]
win_rates = win_cnts / num_games
print("Win rate:")
print("{}: {}".format(agents[0], win_rates[0]))
print("{}: {}".format(agents[1], win_rates[1]))
print("---------------------------------------")
return tuple(win_rates)