rd.seed(seed)
###
seed %= 4294967294
global seed_
seed_ = seed
rd.seed(seed)
np.random.seed(seed)
try:
import tensorflow as tf
tf.set_random_seed(seed)
except Exception as e:
print(e)
print('using seed %s' % (str(seed)))
env = TfEnv(normalize(PointEnvRandGoal()))
policy = MAMLGaussianMLPPolicy(
name="policy",
env_spec=env.spec,
grad_step_size=fast_learning_rate,
hidden_nonlinearity=tf.nn.relu,
hidden_sizes=(100, 100),
std_modifier=pre_std_modifier,
)
if bas == 'zero':
baseline = ZeroBaseline(env_spec=env.spec)
elif 'linear' in bas:
baseline = LinearFeatureBaseline(env_spec=env.spec)
else:
baseline = GaussianMLPBaseline(env_spec=env.spec)
#expert_policy = PointEnvExpertPolicy(env_spec=env.spec)
names = ['maml', 'maml0', 'random', 'oracle']
exp_names = [gen_name + name for name in names]
all_avg_returns = []
for step_i, initial_params_file in zip(range(len(step_sizes)),
initial_params_files):
avg_returns = []
for goal in goals:
goal = list(goal)
if initial_params_file is not None and 'oracle' in initial_params_file:
env = normalize(PointEnvRandGoalOracle(goal=goal))
n_itr = 1
else:
env = normalize(PointEnvRandGoal(goal=goal))
n_itr = 5
env = TfEnv(env)
policy = GaussianMLPPolicy( # random policy
name='policy',
env_spec=env.spec,
hidden_sizes=(100, 100),
)
if initial_params_file is not None:
policy = None
baseline = LinearFeatureBaseline(env_spec=env.spec)
algo = VPG(
env=env,
policy=policy,
###
seed %= 4294967294
global seed_
seed_ = seed
rd.seed(seed)
np.random.seed(seed)
try:
import tensorflow as tf
tf.set_random_seed(seed)
except Exception as e:
print(e)
print('using seed %s' % (str(seed)))
env = TfEnv(normalize(PointEnvRandGoal()))
base_partitions = [PointEnvRandGoal(goal = goal) for goal in goals]
partitions = [TfEnv(normalize(part_env)) for part_env in base_partitions]
metaPolicy = MAMLGaussianMLPPolicy(
name="central_policy",
env_spec=env.spec,
grad_step_size=fast_learning_rate,
hidden_nonlinearity=tf.nn.relu,
hidden_sizes=(100, 100),
std_modifier=pre_std_modifier,
num_tasks = meta_batch_size,
updateMode = updateMode
)
metaBaseline = LinearFeatureBaseline(env_spec=env.spec)
