start_chips=10000,
# You can specify one or both modes. Choosing both is useful to compare them.
eval_modes_of_algo=(
EvalAgentDeepCFR.EVAL_MODE_SINGLE, # SD-CFR
# EvalAgentDeepCFR.EVAL_MODE_AVRG_NET
),
# enables simplified obs. Default works also for 3+ players
use_simplified_headsup_obs=True,
log_verbose=True,
lbr_args=LBRArgs(
lbr_bet_set=bet_sets.PL_2,
n_lbr_hands_per_seat=1,
lbr_check_to_round=Poker.TURN,
# recommended to set to Poker.TURN for 4-round games.
n_parallel_lbr_workers=1,
use_gpu_for_batch_eval=False,
DISTRIBUTED=False,
),
),
eval_methods={
"lbr": 99, # lbr, br, h2h
},
n_iterations=64)
# here we find all particular aces and their avg regrets grouped by actions where round=preflop
hand = np.array([[12, 2], [12, 0]])
player = 0
g_round = Poker.TURN
eq = hand_and_round_eq(hand=hand, round=g_round, ctrl=ctrl)
def __init__(self):
t_prof = TrainingProfile(
name="NLH_EXPLOITABILITY_PLO",
nn_type="feedforward",
DISTRIBUTED=False,
CLUSTER=False,
n_learner_actor_workers=2, # 20 workers
max_buffer_size_adv=1e6,
max_buffer_size_avrg=1e6,
export_each_net=False,
checkpoint_freq=8,
eval_agent_export_freq=4, # produces GBs!
# How many actions out of all legal on current step to branch randomly = action bredth limit
n_actions_traverser_samples=
4, # 3 is the default, 4 is current max for b_2
# number of traversals gives some amount of otcomes to train network on
# mult = 1...4, buffer appends every() step with new data
n_traversals_per_iter=30,
# number of mini_batch fetches and model updates on each step
n_batches_adv_training=801, # 1024
n_batches_avrg_training=2048, # 2048
use_pre_layers_adv=True,
n_cards_state_units_adv=192,
n_merge_and_table_layer_units_adv=64,
n_units_final_adv=64,
# amount of batch to feed to NN at once, fetched from buffer randomly.
mini_batch_size_adv=512, # 256
mini_batch_size_avrg=512, # 512
init_adv_model=
"random", # warm start neural weights with init from last iter
init_avrg_model="random",
# use_pre_layers_avrg=False, # shallower nets
lr_avrg=0.001,
game_cls=DiscretizedNLHoldem, # PLO or DiscretizedNLHoldem
env_bldr_cls=VanillaEnvBuilder,
agent_bet_set=bet_sets.PL_2,
n_seats=2,
start_chips=10000,
# You can specify one or both modes. Choosing both is useful to compare them.
eval_modes_of_algo=(
EvalAgentDeepCFR.EVAL_MODE_SINGLE, # SD-CFR
# EvalAgentDeepCFR.EVAL_MODE_AVRG_NET
),
# enables simplified obs. Default works also for 3+ players
use_simplified_headsup_obs=True,
log_verbose=True,
lbr_args=LBRArgs(
lbr_bet_set=bet_sets.PL_2,
n_lbr_hands_per_seat=100,
lbr_check_to_round=Poker.TURN,
# recommended to set to Poker.TURN for 4-round games.
n_parallel_lbr_workers=1,
use_gpu_for_batch_eval=False,
DISTRIBUTED=True,
),
)
self._eval_env_bldr = _util.get_env_builder_lbr(t_prof=t_prof)
stk = [10000, 10000]
self._env = self._eval_env_bldr.get_new_env(is_evaluating=True,
stack_size=stk)
self.t_prof = t_prof
# You can specify one or both modes. Choosing both is useful to compare them.
eval_modes_of_algo=(
EvalAgentDeepCFR.EVAL_MODE_SINGLE, # SD-CFR
),
DISTRIBUTED=True,
log_verbose=True,
rl_br_args=RLBRArgs(rlbr_bet_set=None,
n_hands_each_seat=200,
n_workers=1,
# Training
DISTRIBUTED=False,
n_iterations=100,
play_n_games_per_iter=50,
# The DDQN
batch_size=512,
),
lbr_args=LBRArgs(n_lbr_hands_per_seat=30000,
n_parallel_lbr_workers=10,
DISTRIBUTED=True,
),
),
eval_methods={'br': 1,
#'rlbr': 1,
'lbr': 1,
},
n_iterations=12)
ctrl.run()
pdb.set_trace()
# from PokerRL.eval.lbr.LBRArgs import LBRArgs
# agent_prof.module_args['lbr'] = LBRArgs()
# lbr_chief = Chief(t_prof=agent_prof)
# eval_master = LBRMaster(t_prof=agent_to_eval.t_prof,
# chief_handle=lbr_chief)
# num_workers = 3
# LBR_workers = [LBRWorker(t_prof=agent_prof, chief_handle=lbr_chief, eval_agent_cls=EvalAgentDeepCFR) for _ in range(num_workers)]
# eval_master.set_worker_handles(*LBR_workers)
# #
#
# #eval_master.evaluate(0)
#-------- Driver Approach
from DeepCFR.workers.driver.Driver import Driver
from PokerRL.eval.lbr.LBRArgs import LBRArgs
agent_prof.module_args['lbr'] = LBRArgs()
ctrl = Driver(agent_prof, eval_methods={'lbr': 1})
ctrl.chief_handle._strategy_buffers = agent_to_eval._strategy_buffers
w, info = ctrl.chief_handle._pull_current_eval_strategy()
ctrl.eval_masters['lbr'][0]._eval_agent = agent_to_eval
ctrl.eval_masters['lbr'][0].weights_for_eval_agent = w
ctrl.eval_masters['lbr'][0].evaluate(0)
# ctrl.eval_masters['br'][0]._eval_agent = agent_to_eval
# ctrl.eval_masters['br'][0].evaluate(0)
# agent_prof.module_args['lbr'] = LBRArgs()
# ctrl = Driver(agent_prof, eval_methods={'lbr': 1})
#
# ctrl.eval_masters['lbr'][0]._eval_agent = agent_to_eval