def test_param_space_noise(self):
env = TfEnv(normalize(PointEnvMAML()))
obs = env.reset()
policy = MAMLImprovedGaussianMLPPolicy(name="policy33",
env_spec=env.spec,
hidden_sizes=(16, 16),
hidden_nonlinearity=tf.nn.tanh,
param_noise_std=0.0)
baseline = LinearFeatureBaseline(env_spec=env.spec)
import rllab.misc.logger as logger
logger.set_snapshot_dir('/tmp/')
logger.set_snapshot_mode('last')
algo = MAMLTRPO(
env=env,
policy=policy,
baseline=baseline,
batch_size=3,
max_path_length=10,
meta_batch_size=4,
num_grad_updates=1,
n_itr=1,
discount=0.99,
step_size=0.01,
)
algo.train()
tf.reset_default_graph()
pkl_file = os.path.join('/tmp/', 'params.pkl')
with tf.Session() as sess:
data = joblib.load(pkl_file)
policy = data['policy']
action_1 = policy.get_action(obs)[1]['mean']
action_2 = policy.get_action(obs)[1]['mean']
diff = np.sum((action_1 - action_2)**2)
self.assertAlmostEquals(diff, 0.0)
action_1 = policy.get_action(obs, param_noise_std=1.0)[1]['mean']
action_2 = policy.get_action(obs, param_noise_std=1.0)[1]['mean']
diff = np.sum((action_1 - action_2)**2)
self.assertGreaterEqual(diff, 0.1)
policy.param_noise_std = 1.0
action_1 = policy.get_action(obs)[1]['mean']
action_2 = policy.get_action(obs)[1]['mean']
diff = np.sum((action_1 - action_2)**2)
self.assertGreaterEqual(diff, 0.1)
def test_serialization(self):
env = TfEnv(normalize(PointEnvMAML()))
obs = env.reset()
policy = MAMLImprovedGaussianMLPPolicy(
name="policy56",
env_spec=env.spec,
hidden_sizes=(16, 16),
hidden_nonlinearity=tf.nn.tanh,
)
baseline = LinearFeatureBaseline(env_spec=env.spec)
import rllab.misc.logger as logger
logger.set_snapshot_dir('/tmp/')
logger.set_snapshot_mode('last')
algo = MAMLTRPO(
env=env,
policy=policy,
baseline=baseline,
batch_size=2,
max_path_length=10,
meta_batch_size=4,
num_grad_updates=1,
n_itr=1,
discount=0.99,
step_size=0.01,
)
algo.train()
tf.reset_default_graph()
pkl_file = os.path.join('/tmp/', 'params.pkl')
with tf.Session() as sess:
data = joblib.load(pkl_file)
policy = data['policy']
action_before = policy.get_action(obs)[1]['mean']
dump_string = pickle.dumps(policy)
tf.reset_default_graph()
with tf.Session() as sess:
policy_loaded = pickle.loads(dump_string)
action_after = policy_loaded.get_action(obs)[1]['mean']
diff = np.sum(np.abs(action_before - action_after))
self.assertAlmostEquals(diff, 0.0, places=3)
def test_get_mean_stepsize(self):
env = TfEnv(normalize(PointEnvMAML()))
obs = env.reset()
policy = MAMLImprovedGaussianMLPPolicy(name="policy2",
env_spec=env.spec,
hidden_sizes=(16, 16),
hidden_nonlinearity=tf.nn.tanh,
trainable_step_size=True,
grad_step_size=0.7)
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
mean_stepsize_1 = policy.get_mean_step_size()
self.assertAlmostEquals(mean_stepsize_1, 0.7, places=5)
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,
load_policy=initial_params_file,
baseline=baseline,
from sandbox_maml.rocky.tf.algos.trpo import TRPO
from rllab_maml.baselines.linear_feature_baseline import LinearFeatureBaseline
from rllab_maml.envs.box2d.cartpole_env import CartpoleEnv
from rllab_maml.envs.normalized_env import normalize
from sandbox_maml.rocky.tf.policies.gaussian_gru_policy import GaussianGRUPolicy
from sandbox_maml.rocky.tf.policies.gaussian_lstm_policy import GaussianLSTMPolicy
from sandbox_maml.rocky.tf.envs.base import TfEnv
import sandbox_maml.rocky.tf.core.layers as L
from sandbox_maml.rocky.tf.optimizers.conjugate_gradient_optimizer import ConjugateGradientOptimizer, FiniteDifferenceHvp
from rllab_maml.misc.instrument import stub, run_experiment_lite
stub(globals())
env = TfEnv(normalize(CartpoleEnv()))
policy = GaussianLSTMPolicy(
name="policy",
env_spec=env.spec,
lstm_layer_cls=L.TfBasicLSTMLayer,
# gru_layer_cls=L.GRULayer,
)
baseline = LinearFeatureBaseline(env_spec=env.spec)
algo = TRPO(
env=env,
policy=policy,
baseline=baseline,
batch_size=4000,
max_path_length=100,
n_itr=10,
# should also code up alternative KL thing
variants = VG().variants()
max_path_length = 200
num_grad_updates = 1
use_maml=True
for v in variants:
direc = v['direc']
learning_rate = v['meta_step_size']
if direc:
env = TfEnv(normalize(HalfCheetahEnvRandDirec()))
else:
env = TfEnv(normalize(HalfCheetahEnvRand()))
policy = MAMLGaussianMLPPolicy(
name="policy",
env_spec=env.spec,
grad_step_size=v['fast_lr'],
hidden_nonlinearity=tf.nn.relu,
hidden_sizes=(100,100),
)
baseline = LinearFeatureBaseline(env_spec=env.spec)
algo = MAMLTRPO(
env=env,
policy=policy,
baseline=baseline,
return [2]
# should also code up alternative KL thing
variants = VG().variants()
max_path_length = 200
num_grad_updates = 1
use_maml=True
for v in variants:
task_var = v['task_var']
if task_var == 0:
env = TfEnv(normalize(AntEnvRandDirec()))
task_var = 'direc'
elif task_var == 1:
env = TfEnv(normalize(AntEnvRand()))
task_var = 'vel'
elif task_var == 2:
env = TfEnv(normalize(AntEnvRandGoal()))
task_var = 'pos'
policy = MAMLGaussianMLPPolicy(
name="policy",
env_spec=env.spec,
grad_step_size=v['fast_lr'],
hidden_nonlinearity=tf.nn.relu,
hidden_sizes=(100,100),
)
baselines = ['linear']
fast_batch_size = 20 # 10 works for [0.1, 0.2], 20 doesn't improve much for [0,0.2]
meta_batch_size = 40 # 10 also works, but much less stable, 20 is fairly stable, 40 is more stable
max_path_length = 100
num_grad_updates = 1
meta_step_size = 0.01
use_maml = True
interpreter_path = sys.executable
for fast_learning_rate in fast_learning_rates:
for learning_rate in learning_rates:
for bas in baselines:
stub(globals())
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),
)
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)
algo = MAMLTRPO(
env=env,
fast_learning_rates = [0.1]
baselines = ['linear']
fast_batch_size = 20
meta_batch_size = 60
max_path_length = 10
num_grad_updates = 1
meta_step_size = 0.01
use_maml = True
for fast_learning_rate in fast_learning_rates:
for bas in baselines:
stub(globals())
env = TfEnv(normalize(GridWorldEnvRand('four-state')))
policy = MAMLCategoricalMLPPolicy(
name="policy",
env_spec=env.spec,
grad_step_size=fast_learning_rate,
hidden_nonlinearity=tf.nn.relu,
hidden_sizes=(100,100),
)
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)
algo = MAMLTRPO(
env=env,
variants = VG().variants()
max_path_length = 200
num_grad_updates = 1
use_maml = True
for v in variants:
task_var = v['task_var']
oracle = v['oracle']
if task_var == 0:
task_var = 'direc'
exp_prefix = 'bugfix_trpo_maml_antdirec' + str(max_path_length)
if oracle:
env = TfEnv(normalize(AntEnvDirecOracle()))
else:
env = TfEnv(normalize(AntEnvRandDirec()))
elif task_var == 1:
task_var = 'vel'
exp_prefix = 'posticml_trpo_maml_ant' + str(max_path_length)
if oracle:
env = TfEnv(normalize(AntEnvOracle()))
else:
env = TfEnv(normalize(AntEnvRand()))
elif task_var == 2:
task_var = 'pos'
exp_prefix = 'posticml_trpo_maml_antpos_' + str(max_path_length)
if oracle:
env = TfEnv(normalize(AntEnvRandGoalOracle()))
else: