def main():
global PARTNERS
num_partners = len(PARTNERS) if PARTNERS is not None else 1
print("model path: ", model_path)
net_arch = [args.netsz,args.latentz]
partner_net_arch = [args.netsz,args.netsz]
policy_kwargs = dict(activation_fn=nn.ReLU,
net_arch=[dict(vf=net_arch, pi=net_arch)],
partner_net_arch=[dict(vf=partner_net_arch, pi=partner_net_arch)],
num_partners=num_partners,
baseline=args.baseline,
nomain=args.nomain,
)
def load_model_fn(partners, testing, try_load=True):
return load_model(model_path=model_path, policy_class=HanabiPolicy, policy_kwargs=policy_kwargs, env=env, hp=HP, partners=partners, testing=testing, try_load=try_load)
def learn_model_fn(model, timesteps, save, period):
return learn(model, model_name=model_name, model_path=model_path, timesteps=timesteps, save=save, period=period)
# TRAINING
if not args.testing:
print("#section Training")
model = load_model_fn(partners=PARTNERS, testing=False)
learn_model_fn(model, timesteps=args.timesteps, save=True, period=2000)
ts, period = 25600, HP['n_steps_testing']
# TESTING
if args.testing and not args.zeroshot:
if args.baseline: adapt_partner_baseline(load_model_fn, learn_model_fn, partners=PARTNERS, timesteps=ts, period=period, do_optimal=False)
else: adapt_partner_modular(load_model_fn, learn_model_fn, partners=PARTNERS, timesteps=ts, period=period, do_optimal=False)
adapt_partner_scratch(load_model_fn, learn_model_fn, partners=PARTNERS, timesteps=ts, period=period, do_optimal=False)
if args.testing and args.zeroshot:
adapt_task(load_model_fn, learn_model_fn, train_partners=TRAIN_PARTNERS, test_partners=TEST_PARTNERS, timesteps=args.timesteps, period=200)
def main():
env = gym.make('blocks-v0',
grid_size=args.n,
vis1=args.vis1,
vis2=args.vis2,
one_sided_reward=args.onesided,
max_move_number=args.maxmovenumber)
num_partners = len(PARTNERS) if PARTNERS is not None else 1
print("model path: ", model_path)
net_arch = [args.netsz, args.latentz]
partner_net_arch = [args.netsz, args.netsz]
policy_kwargs = dict(
activation_fn=nn.ReLU,
net_arch=[dict(vf=net_arch, pi=net_arch)],
partner_net_arch=[dict(vf=partner_net_arch, pi=partner_net_arch)],
num_partners=num_partners,
baseline=args.baseline,
nomain=args.nomain,
)
def load_model_fn(partners, testing, try_load=True):
return load_model(model_path=model_path,
policy_class=BlocksPolicy,
policy_kwargs=policy_kwargs,
env=env,
hp=HP,
partners=partners,
testing=testing,
try_load=try_load)
def learn_model_fn(model, timesteps, save, period):
save_thresh = 19.2 if args.selfplay else None
return learn(model,
model_name=model_name,
model_path=model_path,
timesteps=timesteps,
save=save,
period=period,
save_thresh=save_thresh)
# TRAINING
if not args.testing:
print("#section Training")
model = load_model_fn(partners=PARTNERS, testing=False)
learn_model_fn(model, timesteps=args.timesteps, save=True, period=5000)
ts, period = 25600, HP['n_steps_testing']
# TESTING
if args.testing and not args.zeroshot:
if args.baseline:
adapt_partner_baseline(load_model_fn,
learn_model_fn,
partners=PARTNERS,
timesteps=ts,
period=period,
do_optimal=False)
else:
adapt_partner_modular(load_model_fn,
learn_model_fn,
partners=PARTNERS,
timesteps=ts,
period=period,
do_optimal=False)
if args.testing and args.zeroshot:
adapt_task(load_model_fn,
learn_model_fn,
train_partners=TRAIN_PARTNERS,
test_partners=TEST_PARTNERS,
invert_train_partners=INVERTTRAIN_PARTNERS,
invert_test_partners=INVERTTEST_PARTNERS,
timesteps1=100000,
timesteps2=500000,
period=5000)
def main(args):
# global PARTNERS
# Get model name and path
model_name, model_path = get_model_name_and_path(args.run, mreg=args.mreg)
# Hyperparameters TODO -> Work out a gin config for these
HP = {
'n_steps': 640,
'n_steps_testing': 640,
'batch_size': 160,
'n_epochs': 5,
'n_epochs_testing': 5,
'mreg': args.mreg,
}
# Game config
config = {
"colors": args.colors,
"ranks": args.ranks,
"players": 2,
"hand_size": args.hand_sz,
"max_information_tokens": args.info,
"max_life_tokens": args.life,
"observation_type": pyhanabi.AgentObservationType.CARD_KNOWLEDGE.value
}
# TODO Need to integrate this with Anton's hmf
# NOTE He is using gymt to make the env, which menas it is written
# in a gym style ?
env = gym.make('hanabi-v0', config=config)
# This is the list of partners. It will be None for Selfplay, adn set to
# the other partnerrs for Multiplay
PARTNERS = None
if not args.selfplay:
# TODO Inject Rulebased Partner Config
setting, partner_type = "", "ppo"
# get the config parameters of the hanabi partners
TRAIN_PARTNERS, TEST_PARTNERS = get_hanabi_partners(setting, partner_type)
PARTNERS = [TEST_PARTNERS[args.k % len(TEST_PARTNERS)]] if args.testing else TRAIN_PARTNERS
num_partners = len(PARTNERS) if PARTNERS is not None else 1
print("model path: ", model_path)
net_arch = [ args.netsz, # Network Dimensions
args.latentz # Latent Dimensions (Z)
]
# NOTE insertion point for rule-based agents -> Look at how to interface this with Umut's work
# Define the architechture for hte parnter network
partner_net_arch = [ args.netsz,
args.netsz
]
policy_kwargs = dict(activation_fn=nn.ReLU,
net_arch=[dict(vf=net_arch, pi=net_arch)],
# NOTE this needs to be changed somehow to incorporate multiple partners
partner_net_arch=[dict(vf=partner_net_arch, pi=partner_net_arch)],
num_partners=num_partners,
baseline=args.baseline,
nomain=args.nomain,
)
def load_model_fn( partners,
testing,
try_load=True
):
return load_model( model_path=model_path,
policy_class=HanabiPolicy,
policy_kwargs=policy_kwargs,
env=env,
hp=HP,
partners=partners,
testing=testing,
try_load=try_load
)
def learn_model_fn( model,
timesteps,
save,
period
):
return learn( model,
model_name=model_name,
model_path=model_path,
timesteps=timesteps,
save=save,
period=period
)
# TRAINING
if not args.testing:
print("#Section Training")
# PARTNERS is passed to load the model. Question is what role does it play ?
model = load_model_fn( partners=PARTNERS,
testing=False
)
learn_model_fn( model,
timesteps=args.timesteps,
save=True,
period=2000
)
ts, period = 25600, HP['n_steps_testing']
# TESTING
if args.testing and not args.zeroshot:
if args.baseline:
adapt_partner_baseline( load_model_fn,
learn_model_fn,
partners=PARTNERS,
timesteps=ts,
period=period,
do_optimal=False
)
else:
adapt_partner_modular( load_model_fn,
learn_model_fn,
partners=PARTNERS,
timesteps=ts,
period=period,
do_optimal=False
)
adapt_partner_scratch( load_model_fn,
learn_model_fn,
partners=PARTNERS,
timesteps=ts,
period=period,
do_optimal=False
)
if args.testing and args.zeroshot:
adapt_task( load_model_fn,
learn_model_fn,
train_partners=TRAIN_PARTNERS,
test_partners=TEST_PARTNERS,
timesteps=args.timesteps,
period=200
)