def start(args):
logger = utils.default_logger_setup(logging.INFO)
logger.info('Writing to ' + args.outdir)
config.set_cuda(args.cuda)
p1dir = os.path.join(args.outdir, 'p1')
p2dir = os.path.join(args.outdir, 'p2')
utils.mkdir_p(p1dir)
utils.mkdir_p(p2dir)
prog = os.path.join(
config.get('showdown_root'),
config.get('showdown_server_dir'),
'pokemon-showdown')
game = Game(options = formats.get(args.format), prog = prog)
policy_1 = torch_policy.load(args.p1)
policy_2 = torch_policy.load(args.p2)
wins = [0, 0]
logger.info('starting...')
# TODO: make multithreaded. Maybe integrate this with RL experiment
for i in range(args.num_matches):
p1 = EnginePkmnPlayer(policy_1, '%s-p1' % i,
play_best_move = args.play_best_move in ['p1', 'both'])
p2 = EnginePkmnPlayer(policy_2, '%s-p2' % i,
play_best_move = args.play_best_move in ['p2', 'both'])
game.play(p1, p2)
for j in [0, 1]:
player = [p1, p2][j]
dirname = [p1dir, p2dir][j]
bogger = battlelogs.BattleLogger(player.gid, dirname)
for block in player.blocks:
bogger.log(block)
bogger.close()
if player.result == 'winner':
wins[j] += 1
else:
assert player.result in ['loser', 'tie']
return wins
def main():
args = parse_args()
config.set_cuda(False)
from metagrok import remote_debug
remote_debug.listen()
p1_policy = torch_policy.load(args.policy_tag)
p2_policy = p1_policy
if args.p2_policy_tag:
p2_policy = torch_policy.load(args.p2_policy_tag)
fmt = formats.get(args.fmt)
game = Game(fmt, '{}/{}/pokemon-showdown'.format(
config.get('showdown_root'),
config.get('showdown_server_dir')))
count = 0
while True:
time.sleep(0.1)
r = sys.stdin.readline().strip()
if r == 'done':
break
battle_dir = os.path.join('/tmp', args.id, '%06d' % count)
utils.mkdir_p(battle_dir)
p1 = EnginePkmnPlayer(p1_policy, 'p1', epsilon = args.epsilon)
p2 = EnginePkmnPlayer(p2_policy, 'p2', epsilon = args.epsilon)
game.play(p1, p2)
num_blocks = 0
for i, player in enumerate([p1, p2]):
blogger = battlelogs.BattleLogger('p%d' % (i + 1), battle_dir)
for block in player.blocks:
blogger.log(block)
num_blocks += 1
blogger.close()
count += 1
sys.stdout.write('%s\t%d\n' % (battle_dir, num_blocks))
sys.stdout.flush()
def create_type_chart():
rv = {}
type_chart_fname = '{}/BattleTypeChart.json'.format(config.get('dex_root'))
for defender_type, type_data in json.load(type_chart_fname).items():
defender_type = utils.to_id(defender_type)
attacker_data = {}
for attacker_type, modifier in type_data['damageTaken'].items():
attacker_type = utils.to_id(attacker_type)
if modifier == 0:
multiplier = 1.
elif modifier == 1:
multiplier = 2.
elif modifier == 2:
multiplier = 0.5
elif modifier == 3:
multiplier = 0.
else:
raise ValueError('Unknown modifier %s' % modifier)
attacker_data[attacker_type] = multiplier
rv[defender_type] = attacker_data
return rv
if modifier == 0:
multiplier = 1.
elif modifier == 1:
multiplier = 2.
elif modifier == 2:
multiplier = 0.5
elif modifier == 3:
multiplier = 0.
else:
raise ValueError('Unknown modifier %s' % modifier)
attacker_data[attacker_type] = multiplier
rv[defender_type] = attacker_data
return rv
Movedex = json.load('{}/BattleMovedex.json'.format(config.get('dex_root')))
TypeChart = create_type_chart()
class RandomPlayer(object):
def act(self, state, candidates):
probs = np.asarray([1. if c else 0. for c in candidates])
probs = probs / probs.sum()
return dict(probs = probs)
class MostDamageMovePlayer(object):
'''
Ranks actions in the following order:
- damaging moves (sort by power, then move modifier)
- zmove
- ultra
- mega
def start(args):
logger = utils.default_logger_setup(logging.INFO)
logger.info('Writing to ' + args.outdir)
config.set_cuda(args.cuda)
p1dir = os.path.join(args.outdir, 'p1')
p2dir = os.path.join(args.outdir, 'p2')
utils.mkdir_p(p1dir)
utils.mkdir_p(p2dir)
prog = os.path.join(
config.get('showdown_root'),
config.get('showdown_server_dir'),
'pokemon-showdown')
#game = Game(options = formats.get(args.format), prog = prog)
policy_1 = torch_policy.load(args.p1)
policy_2 = torch_policy.load(args.p2)
wins = [0, 0]
logger.info('starting...')
# TODO: make multithreaded. Maybe integrate this with RL experiment
p1_teams, p2_teams = tg.init_lc_thunderdome()
strategy_agent = team_choice.AgentStrategyProfile(p1_teams, p2_teams)
for i in range(args.num_matches):
#team1_ind = team_choice.teamchoice_random(formats.ou_teams)
team1_ind = strategy_agent.select_action()
team2_ind = strategy_agent.select_action_p2()
game = Game(options = strategy_agent.get_teams(team1_ind, team2_ind), prog = prog)
p1 = EnginePkmnPlayer(policy_1, '%s-p1' % i,
play_best_move = args.play_best_move in ['p1', 'both'])
p2 = EnginePkmnPlayer(policy_2, '%s-p2' % i,
play_best_move = args.play_best_move in ['p2', 'both'])
game.play(p1, p2)
for j in [0, 1]:
player = [p1, p2][j]
dirname = [p1dir, p2dir][j]
bogger = battlelogs.BattleLogger(player.gid, dirname)
for block in player.blocks:
bogger.log(block)
bogger.close()
if j == 0:
if player.result == 'winner':
strategy_agent.update(team1_ind, team2_ind, p1_win=True)
else:
strategy_agent.update(team1_ind, team2_ind, p1_win=False)
if player.result == 'winner':
wins[j] += 1
else:
assert player.result in ['loser', 'tie']
print(strategy_agent.get_utility_matrix())
with open("lc_thunderdome_results.txt", "w+") as wf:
for ct, team in enumerate(strategy_agent.p1_teams):
wf.write("{}\t{}\n".format(ct, team))
for ct, team in enumerate(strategy_agent.p2_teams):
wf.write("{}\t{}\n".format(ct, team))
# wf.write(strategy_agent.get_utility_matrix())
wf.write("\n")
wf.flush()
wf.close()
return wins
name = to_id(name)
if name.startswith('hiddenpower'):
match = HP_REGEX.match(name)
if match:
name = 'hiddenpower' + match.group(1)
elif name.startswith('return') and RETURN_REGEX.match(name):
name = 'return'
if name not in self.name_to_index:
logger.warn('Could not find %s in %s', name, self.feature_key)
return 1
return self.name_to_index[name]
def nidx(self, names):
if not isinstance(names, (list, tuple)):
names = (names,)
return [self.to_index(n) for n in names]
def nhot(self, names):
if not isinstance(names, (list, tuple)):
names = (names,)
rv = np.zeros(self.size, dtype = config.nt())
for name in names:
rv[self.to_index(name)] = 1
return rv
DEX = Dex.create(config.get('dex_root'))
if __name__ == '__main__':
print(DEX.gen7rb_stats().describe())
def script():
return fileio.read('{}/engine.js'.format(config.get('metagrok_client_root')))
def load_test_data():
return jsons.load('{}/reward-shaper-tests.jsons'.format(
config.get('test_data_root')))
