def get_config():
agent_names = ['agent%d' % i for i in range(1, 4)]
model = Model(agent_names, STATE_SHAPE, METHOD, NUM_ACTIONS, GAMMA)
exps = [
ExpReplay(
# model=model,
agent_name=name,
player=Env(agent_names),
state_shape=STATE_SHAPE,
num_actions=[MAX_NUM_COMBS, MAX_NUM_GROUPS],
batch_size=BATCH_SIZE,
memory_size=MEMORY_SIZE,
init_memory_size=INIT_MEMORY_SIZE,
init_exploration=1.,
update_frequency=UPDATE_FREQ) for name in agent_names
]
df = MyDataFLow(exps)
bl_evaluators = [
BLEvaluator(EVAL_EPISODE, agent_names[0], 2, lambda: CEnv()),
BLEvaluator(EVAL_EPISODE, agent_names[1], 3, lambda: CEnv()),
BLEvaluator(EVAL_EPISODE, agent_names[2], 1, lambda: CEnv())
]
return AutoResumeTrainConfig(
# always_resume=False,
data=QueueInput(df),
model=model,
callbacks=[
ModelSaver(),
PeriodicTrigger(RunOp(model.update_target_param, verbose=True),
every_k_steps=STEPS_PER_EPOCH //
10), # update target network every 10k steps
*exps,
# ScheduledHyperParamSetter('learning_rate',
# [(60, 5e-5), (100, 2e-5)]),
*[
ScheduledHyperParamSetter(
ObjAttrParam(exp, 'exploration'),
[(0, 1), (30, 0.5), (100, 0.3),
(320, 0.1)], # 1->0.1 in the first million steps
interp='linear') for exp in exps
],
*bl_evaluators,
Evaluator(EVAL_EPISODE, agent_names, lambda: Env(agent_names)),
HumanHyperParamSetter('learning_rate'),
],
# session_init=ChainInit([SaverRestore('../Hierarchical_Q/train_log/DQN-9-3-LASTCARDS/model-240000', 'agent1'),
# SaverRestore('./train_log/DQN-60-MA/model-355000')]),
# starting_epoch=0,
# session_init=SaverRestore('train_log/DQN-54-AUG-STATE/model-75000'),
steps_per_epoch=STEPS_PER_EPOCH,
max_epoch=1000,
)
def _build_player(self):
return CEnv()
def get_player():
return CEnv()
v = self._player_scores
try:
mean, max = v.average, v.max
self.trainer.monitors.put_scalar('expreplay/mean_score', mean)
self.trainer.monitors.put_scalar('expreplay/max_score', max)
except Exception:
logger.exception("Cannot log training scores.")
v.reset()
if __name__ == '__main__':
def predictor(x):
return [np.random.random([1, 100])]
player = CEnv()
E = ExpReplay(predictor_io_names=(['state', 'comb_mask'], ['Qvalue']),
player=CEnv(),
state_shape=(100, 21, 256),
num_actions=[100, 21],
batch_size=16,
memory_size=1e4,
init_memory_size=1e4,
init_exploration=0.,
update_frequency=4)
E.predictor = predictor
E._init_memory()
# for k in E.get_data():
# pass
# for k in E.get_data():
print(action)
winner, done = env.step(action)
if done:
for agent_name in agent_names:
if agent_name == winner:
cnt[agent_name] += 1
print(agent_name, ' wins')
else:
if env.get_all_agent_names().index(winner) + env.get_all_agent_names().index(agent_name) == 3:
cnt[agent_name] += 1
print(agent_name, winner, ' all wins')
print(cnt)
# C env usage
# TODO
env = CEnv()
# for _ in range(1):
# env.reset()
# env.prepare()
# done = False
# while not done:
# handcards = to_char(env.get_curr_handcards())
#
# chandcards = [CCard(to_value(c) - 3) for c in handcards]
# unseen_cards = env.player_cards[agent_names[(env.get_current_idx() + 1) % len(env.agent_names)]].copy() \
# + env.player_cards[agent_names[(env.get_current_idx() + 2) % len(env.agent_names)]].copy()
# cunseen_cards = [CCard(to_value(c) - 3) for c in unseen_cards]
# next_handcards_cnt = len(env.player_cards[agent_names[(env.get_current_idx() + 1) % len(env.agent_names)]])
#
# last_cg = char2ccardgroup(env.get_last_outcards())
# caction = mcsearch(chandcards, cunseen_cards, next_handcards_cnt, last_cg, env.agent_names.index(env.curr_player), env.agent_names.index(env.controller))
