def test_dist_info_sym_include_action(self, obs_dim, action_dim,
hidden_dim):
env = TfEnv(DummyBoxEnv(obs_dim=obs_dim, action_dim=action_dim))
obs_ph = tf.compat.v1.placeholder(
tf.float32, shape=(None, None, env.observation_space.flat_dim))
with mock.patch(('metarl.tf.policies.'
'gaussian_gru_policy.GaussianGRUModel'),
new=SimpleGaussianGRUModel):
policy = GaussianGRUPolicy(env_spec=env.spec,
state_include_action=True)
policy.reset()
obs = env.reset()
dist_sym = policy.dist_info_sym(
obs_var=obs_ph,
state_info_vars={'prev_action': np.zeros((2, 1) + action_dim)},
name='p2_sym')
dist = self.sess.run(
dist_sym, feed_dict={obs_ph: [[obs.flatten()], [obs.flatten()]]})
assert np.array_equal(dist['mean'], np.full((2, 1) + action_dim, 0.5))
assert np.array_equal(dist['log_std'], np.full((2, 1) + action_dim,
0.5))
def test_get_action_state_include_action(self, obs_dim, action_dim,
hidden_dim):
env = MetaRLEnv(DummyBoxEnv(obs_dim=obs_dim, action_dim=action_dim))
obs_var = tf.compat.v1.placeholder(
tf.float32,
shape=[
None, None,
env.observation_space.flat_dim + np.prod(action_dim)
],
name='obs')
policy = GaussianGRUPolicy(env_spec=env.spec,
hidden_dim=hidden_dim,
state_include_action=True)
policy.build(obs_var)
policy.reset()
obs = env.reset()
action, _ = policy.get_action(obs.flatten())
assert env.action_space.contains(action)
policy.reset()
actions, _ = policy.get_actions([obs.flatten()])
for action in actions:
assert env.action_space.contains(action)
def test_gaussian_gru_policy(self):
gaussian_gru_policy = GaussianGRUPolicy(env_spec=self.env,
hidden_dim=1)
self.sess.run(tf.compat.v1.global_variables_initializer())
gaussian_gru_policy.reset()
obs = self.env.observation_space.high
assert gaussian_gru_policy.get_action(obs)
def test_get_action_state_include_action(self, mock_normal, obs_dim,
action_dim, hidden_dim):
mock_normal.return_value = 0.5
env = TfEnv(DummyBoxEnv(obs_dim=obs_dim, action_dim=action_dim))
with mock.patch(('metarl.tf.policies.'
'gaussian_gru_policy.GaussianGRUModel'),
new=SimpleGaussianGRUModel):
policy = GaussianGRUPolicy(env_spec=env.spec,
state_include_action=True)
policy.reset()
obs = env.reset()
expected_action = np.full(action_dim, 0.5 * np.exp(0.5) + 0.5)
action, agent_info = policy.get_action(obs)
assert env.action_space.contains(action)
assert np.allclose(action, expected_action)
expected_mean = np.full(action_dim, 0.5)
assert np.array_equal(agent_info['mean'], expected_mean)
expected_log_std = np.full(action_dim, 0.5)
assert np.array_equal(agent_info['log_std'], expected_log_std)
expected_prev_action = np.full(action_dim, 0)
assert np.array_equal(agent_info['prev_action'], expected_prev_action)
policy.reset()
actions, agent_infos = policy.get_actions([obs])
for action, mean, log_std, prev_action in zip(
actions, agent_infos['mean'], agent_infos['log_std'],
agent_infos['prev_action']):
assert env.action_space.contains(action)
assert np.allclose(action, expected_action)
assert np.array_equal(mean, expected_mean)
assert np.array_equal(log_std, expected_log_std)
assert np.array_equal(prev_action, expected_prev_action)
def test_is_pickleable(self):
env = TfEnv(DummyBoxEnv(obs_dim=(1, ), action_dim=(1, )))
with mock.patch(('metarl.tf.policies.'
'gaussian_gru_policy.GaussianGRUModel'),
new=SimpleGaussianGRUModel):
policy = GaussianGRUPolicy(env_spec=env.spec,
state_include_action=False)
env.reset()
obs = env.reset()
with tf.compat.v1.variable_scope('GaussianGRUPolicy/GaussianGRUModel',
reuse=True):
return_var = tf.compat.v1.get_variable('return_var')
# assign it to all one
return_var.load(tf.ones_like(return_var).eval())
output1 = self.sess.run(
policy.model.networks['default'].mean,
feed_dict={policy.model.input: [[obs.flatten()], [obs.flatten()]]})
p = pickle.dumps(policy)
with tf.compat.v1.Session(graph=tf.Graph()) as sess:
policy_pickled = pickle.loads(p)
# yapf: disable
output2 = sess.run(
policy_pickled.model.networks['default'].mean,
feed_dict={
policy_pickled.model.input: [[obs.flatten()],
[obs.flatten()]]
})
assert np.array_equal(output1, output2)
def test_ppo_pendulum_gru(self):
"""Test PPO with Pendulum environment and recurrent policy."""
with LocalTFRunner(snapshot_config) as runner:
env = MetaRLEnv(normalize(gym.make('InvertedDoublePendulum-v2')))
gru_policy = GaussianGRUPolicy(env_spec=env.spec)
baseline = GaussianMLPBaseline(
env_spec=env.spec,
regressor_args=dict(hidden_sizes=(32, 32)),
)
algo = PPO(
env_spec=env.spec,
policy=gru_policy,
baseline=baseline,
max_path_length=100,
discount=0.99,
gae_lambda=0.95,
lr_clip_range=0.2,
optimizer_args=dict(
batch_size=32,
max_epochs=10,
),
stop_entropy_gradient=True,
entropy_method='max',
policy_ent_coeff=0.02,
center_adv=False,
)
runner.setup(algo, env)
last_avg_ret = runner.train(n_epochs=10, batch_size=2048)
assert last_avg_ret > 80
def rl2_ppo_metaworld_ml1_push(ctxt, seed, max_path_length, meta_batch_size,
n_epochs, episode_per_task):
"""Train PPO with ML1 environment.
Args:
ctxt (metarl.experiment.ExperimentContext): The experiment
configuration used by LocalRunner to create the snapshotter.
seed (int): Used to seed the random number generator to produce
determinism.
max_path_length (int): Maximum length of a single rollout.
meta_batch_size (int): Meta batch size.
n_epochs (int): Total number of epochs for training.
episode_per_task (int): Number of training episode per task.
"""
set_seed(seed)
with LocalTFRunner(snapshot_config=ctxt) as runner:
tasks = task_sampler.SetTaskSampler(lambda: RL2Env(
env=mwb.ML1.get_train_tasks('push-v1')))
env_spec = RL2Env(env=mwb.ML1.get_train_tasks('push-v1')).spec
policy = GaussianGRUPolicy(name='policy',
hidden_dim=64,
env_spec=env_spec,
state_include_action=False)
baseline = LinearFeatureBaseline(env_spec=env_spec)
algo = RL2PPO(rl2_max_path_length=max_path_length,
meta_batch_size=meta_batch_size,
task_sampler=tasks,
env_spec=env_spec,
policy=policy,
baseline=baseline,
discount=0.99,
gae_lambda=0.95,
lr_clip_range=0.2,
optimizer_args=dict(
batch_size=32,
max_epochs=10,
),
stop_entropy_gradient=True,
entropy_method='max',
policy_ent_coeff=0.02,
center_adv=False,
max_path_length=max_path_length * episode_per_task)
runner.setup(algo,
tasks.sample(meta_batch_size),
sampler_cls=LocalSampler,
n_workers=meta_batch_size,
worker_class=RL2Worker,
worker_args=dict(n_paths_per_trial=episode_per_task))
runner.train(n_epochs=n_epochs,
batch_size=episode_per_task * max_path_length *
meta_batch_size)
def test_is_pickleable(self):
env = MetaRLEnv(DummyBoxEnv(obs_dim=(1, ), action_dim=(1, )))
obs_var = tf.compat.v1.placeholder(
tf.float32,
shape=[None, None, env.observation_space.flat_dim],
name='obs')
policy = GaussianGRUPolicy(env_spec=env.spec,
state_include_action=False)
policy.build(obs_var)
env.reset()
obs = env.reset()
with tf.compat.v1.variable_scope('GaussianGRUPolicy/GaussianGRUModel',
reuse=True):
param = tf.compat.v1.get_variable(
'dist_params/log_std_param/parameter')
# assign it to all one
param.load(tf.ones_like(param).eval())
output1 = self.sess.run(
[policy.distribution.loc,
policy.distribution.stddev()],
feed_dict={policy.model.input: [[obs.flatten()], [obs.flatten()]]})
p = pickle.dumps(policy)
with tf.compat.v1.Session(graph=tf.Graph()) as sess:
policy_pickled = pickle.loads(p)
obs_var = tf.compat.v1.placeholder(
tf.float32,
shape=[None, None, env.observation_space.flat_dim],
name='obs')
policy_pickled.build(obs_var)
# yapf: disable
output2 = sess.run(
[
policy_pickled.distribution.loc,
policy_pickled.distribution.stddev()
],
feed_dict={
policy_pickled.model.input: [[obs.flatten()],
[obs.flatten()]]
})
assert np.array_equal(output1, output2)
def gaussian_gru_policy(ctxt, env_id, seed):
"""Create Gaussian GRU Policy on TF-PPO.
Args:
ctxt (metarl.experiment.ExperimentContext): The experiment
configuration used by LocalRunner to create the
snapshotter.
env_id (str): Environment id of the task.
seed (int): Random positive integer for the trial.
"""
deterministic.set_seed(seed)
with LocalTFRunner(ctxt) as runner:
env = MetaRLEnv(normalize(gym.make(env_id)))
policy = GaussianGRUPolicy(
env_spec=env.spec,
hidden_dim=32,
hidden_nonlinearity=tf.nn.tanh,
output_nonlinearity=None,
)
baseline = GaussianMLPBaseline(
env_spec=env.spec,
regressor_args=dict(
hidden_sizes=(64, 64),
use_trust_region=False,
optimizer=FirstOrderOptimizer,
optimizer_args=dict(
batch_size=32,
max_epochs=10,
learning_rate=1e-3,
),
),
)
algo = PPO(
env_spec=env.spec,
policy=policy,
baseline=baseline,
max_path_length=100,
discount=0.99,
gae_lambda=0.95,
lr_clip_range=0.2,
policy_ent_coeff=0.0,
optimizer_args=dict(
batch_size=32,
max_epochs=10,
learning_rate=1e-3,
),
)
runner.setup(algo, env, sampler_args=dict(n_envs=12))
runner.train(n_epochs=5, batch_size=2048)
def setup_method(self):
super().setup_method()
self.max_path_length = 100
self.meta_batch_size = 10
self.episode_per_task = 4
self.tasks = task_sampler.SetTaskSampler(
lambda: RL2Env(env=normalize(HalfCheetahDirEnv())))
self.env_spec = RL2Env(env=normalize(HalfCheetahDirEnv())).spec
self.policy = GaussianGRUPolicy(env_spec=self.env_spec,
hidden_dim=64,
state_include_action=False)
self.baseline = LinearFeatureBaseline(env_spec=self.env_spec)
def rl2_trpo_halfcheetah(ctxt, seed, max_path_length, meta_batch_size,
n_epochs, episode_per_task):
"""Train TRPO with HalfCheetah environment.
Args:
ctxt (metarl.experiment.ExperimentContext): The experiment
configuration used by LocalRunner to create the snapshotter.
seed (int): Used to seed the random number generator to produce
determinism.
max_path_length (int): Maximum length of a single rollout.
meta_batch_size (int): Meta batch size.
n_epochs (int): Total number of epochs for training.
episode_per_task (int): Number of training episode per task.
"""
set_seed(seed)
with LocalTFRunner(snapshot_config=ctxt) as runner:
tasks = task_sampler.SetTaskSampler(
lambda: RL2Env(env=HalfCheetahVelEnv()))
env_spec = RL2Env(env=HalfCheetahVelEnv()).spec
policy = GaussianGRUPolicy(name='policy',
hidden_dim=64,
env_spec=env_spec,
state_include_action=False)
baseline = LinearFeatureBaseline(env_spec=env_spec)
algo = RL2TRPO(rl2_max_path_length=max_path_length,
meta_batch_size=meta_batch_size,
task_sampler=tasks,
env_spec=env_spec,
policy=policy,
baseline=baseline,
max_path_length=max_path_length * episode_per_task,
discount=0.99,
max_kl_step=0.01,
optimizer=ConjugateGradientOptimizer,
optimizer_args=dict(hvp_approach=FiniteDifferenceHvp(
base_eps=1e-5)))
runner.setup(algo,
tasks.sample(meta_batch_size),
sampler_cls=LocalSampler,
n_workers=meta_batch_size,
worker_class=RL2Worker,
worker_args=dict(n_paths_per_trial=episode_per_task))
runner.train(n_epochs=n_epochs,
batch_size=episode_per_task * max_path_length *
meta_batch_size)
def run_task(self, snapshot_config, *_):
config = tf.ConfigProto(device_count={'GPU': 0},
allow_soft_placement=True,
intra_op_parallelism_threads=12,
inter_op_parallelism_threads=12)
sess = tf.Session(config=config)
with LocalTFRunner(snapshot_config=snapshot_config,
sess=sess) as runner:
env = gym.make(self._env)
env = TfEnv(normalize(env))
env.reset()
policy = GaussianGRUPolicy(
env_spec=env.spec,
hidden_dim=32,
hidden_nonlinearity=tf.nn.tanh,
output_nonlinearity=None,
)
baseline = GaussianMLPBaseline(
env_spec=env.spec,
regressor_args=dict(
hidden_sizes=(64, 64),
use_trust_region=False,
optimizer=FirstOrderOptimizer,
optimizer_args=dict(
batch_size=32,
max_epochs=10,
tf_optimizer_args=dict(learning_rate=1e-3),
),
),
)
algo = PPO(
env_spec=env.spec,
policy=policy,
baseline=baseline,
max_path_length=100,
discount=0.99,
gae_lambda=0.95,
lr_clip_range=0.2,
policy_ent_coeff=0.0,
optimizer_args=dict(
batch_size=32,
max_epochs=10,
tf_optimizer_args=dict(learning_rate=1e-3),
),
)
runner.setup(algo, env, sampler_args=dict(n_envs=12))
runner.train(n_epochs=5, batch_size=2048)
def setup_method(self):
super().setup_method()
self.env = MetaRLEnv(normalize(gym.make('InvertedDoublePendulum-v2')))
self.policy = GaussianMLPPolicy(
env_spec=self.env.spec,
hidden_sizes=(64, 64),
hidden_nonlinearity=tf.nn.tanh,
output_nonlinearity=None,
)
self.lstm_policy = GaussianLSTMPolicy(env_spec=self.env.spec)
self.gru_policy = GaussianGRUPolicy(env_spec=self.env.spec)
self.baseline = GaussianMLPBaseline(
env_spec=self.env.spec,
regressor_args=dict(hidden_sizes=(32, 32)),
)
def test_dist_info_sym_wrong_input(self):
env = TfEnv(DummyBoxEnv(obs_dim=(1, ), action_dim=(1, )))
obs_ph = tf.compat.v1.placeholder(
tf.float32, shape=(None, None, env.observation_space.flat_dim))
with mock.patch(('metarl.tf.policies.'
'gaussian_gru_policy.GaussianGRUModel'),
new=SimpleGaussianGRUModel):
policy = GaussianGRUPolicy(env_spec=env.spec,
state_include_action=True)
policy.reset()
obs = env.reset()
policy.dist_info_sym(
obs_var=obs_ph,
state_info_vars={'prev_action': np.zeros((3, 1, 1))},
name='p2_sym')
# observation batch size = 2 but prev_action batch size = 3
with pytest.raises(tf.errors.InvalidArgumentError):
self.sess.run(
policy.model.networks['p2_sym'].input,
feed_dict={obs_ph: [[obs.flatten()], [obs.flatten()]]})
def run_metarl(env, seed, log_dir):
"""Create metarl Tensorflow PPO model and training.
Args:
env (dict): Environment of the task.
seed (int): Random positive integer for the trial.
log_dir (str): Log dir path.
Returns:
str: Path to output csv file
"""
deterministic.set_seed(seed)
snapshot_config = SnapshotConfig(snapshot_dir=log_dir,
snapshot_mode='gap',
snapshot_gap=10)
with LocalTFRunner(snapshot_config) as runner:
env, task_samplers = _prepare_meta_env(env)
policy = GaussianGRUPolicy(
hidden_dims=hyper_parameters['hidden_sizes'],
env_spec=env.spec,
state_include_action=False)
baseline = MetaRLLinearFeatureBaseline(env_spec=env.spec)
inner_algo = RL2PPO(
env_spec=env.spec,
policy=policy,
baseline=baseline,
max_path_length=hyper_parameters['max_path_length'] *
hyper_parameters['rollout_per_task'],
discount=hyper_parameters['discount'],
gae_lambda=hyper_parameters['gae_lambda'],
lr_clip_range=hyper_parameters['lr_clip_range'],
optimizer_args=dict(
max_epochs=hyper_parameters['optimizer_max_epochs'],
tf_optimizer_args=dict(
learning_rate=hyper_parameters['optimizer_lr'], ),
))
algo = RL2(policy=policy,
inner_algo=inner_algo,
max_path_length=hyper_parameters['max_path_length'],
meta_batch_size=hyper_parameters['meta_batch_size'],
task_sampler=task_samplers)
# Set up logger since we are not using run_experiment
tabular_log_file = osp.join(log_dir, 'progress.csv')
text_log_file = osp.join(log_dir, 'debug.log')
dowel_logger.add_output(dowel.TextOutput(text_log_file))
dowel_logger.add_output(dowel.CsvOutput(tabular_log_file))
dowel_logger.add_output(dowel.StdOutput())
dowel_logger.add_output(dowel.TensorBoardOutput(log_dir))
runner.setup(
algo,
task_samplers.sample(hyper_parameters['meta_batch_size']),
sampler_cls=hyper_parameters['sampler_cls'],
n_workers=hyper_parameters['meta_batch_size'],
worker_class=RL2Worker,
sampler_args=dict(
use_all_workers=hyper_parameters['use_all_workers']),
worker_args=dict(
n_paths_per_trial=hyper_parameters['rollout_per_task']))
runner.setup_meta_evaluator(
test_task_sampler=task_samplers,
n_exploration_traj=hyper_parameters['rollout_per_task'],
n_test_rollouts=hyper_parameters['test_rollout_per_task'],
n_test_tasks=hyper_parameters['n_test_tasks'],
n_workers=hyper_parameters['n_test_tasks'])
runner.train(n_epochs=hyper_parameters['n_itr'],
batch_size=hyper_parameters['meta_batch_size'] *
hyper_parameters['rollout_per_task'] *
hyper_parameters['max_path_length'])
dowel_logger.remove_all()
return tabular_log_file
def test_clone(self):
env = MetaRLEnv(DummyBoxEnv(obs_dim=(4, ), action_dim=(4, )))
policy = GaussianGRUPolicy(env_spec=env.spec)
policy_clone = policy.clone('GaussianGRUPolicyClone')
assert policy_clone.env_spec == policy.env_spec
def rl2_ppo_metaworld_ml10_meta_test(ctxt, seed, max_path_length,
meta_batch_size, n_epochs,
episode_per_task):
"""Train PPO with ML10 environment with meta-test.
Args:
ctxt (metarl.experiment.ExperimentContext): The experiment
configuration used by LocalRunner to create the snapshotter.
seed (int): Used to seed the random number generator to produce
determinism.
max_path_length (int): Maximum length of a single rollout.
meta_batch_size (int): Meta batch size.
n_epochs (int): Total number of epochs for training.
episode_per_task (int): Number of training episode per task.
"""
set_seed(seed)
with LocalTFRunner(snapshot_config=ctxt) as runner:
ml10_train_envs = [
RL2Env(mwb.ML10.from_task(task_name))
for task_name in mwb.ML10.get_train_tasks().all_task_names
]
tasks = task_sampler.EnvPoolSampler(ml10_train_envs)
tasks.grow_pool(meta_batch_size)
ml10_test_envs = [
RL2Env(mwb.ML10.from_task(task_name))
for task_name in mwb.ML10.get_test_tasks().all_task_names
]
test_tasks = task_sampler.EnvPoolSampler(ml10_test_envs)
env_spec = ml10_train_envs[0].spec
policy = GaussianGRUPolicy(name='policy',
hidden_dim=64,
env_spec=env_spec,
state_include_action=False)
baseline = LinearFeatureBaseline(env_spec=env_spec)
meta_evaluator = MetaEvaluator(test_task_sampler=test_tasks,
n_exploration_traj=10,
n_test_rollouts=10,
max_path_length=max_path_length,
n_test_tasks=5)
algo = RL2PPO(rl2_max_path_length=max_path_length,
meta_batch_size=meta_batch_size,
task_sampler=tasks,
env_spec=env_spec,
policy=policy,
baseline=baseline,
discount=0.99,
gae_lambda=0.95,
lr_clip_range=0.2,
optimizer_args=dict(
batch_size=32,
max_epochs=10,
),
stop_entropy_gradient=True,
entropy_method='max',
policy_ent_coeff=0.02,
center_adv=False,
max_path_length=max_path_length * episode_per_task,
meta_evaluator=meta_evaluator,
n_epochs_per_eval=10)
runner.setup(algo,
tasks.sample(meta_batch_size),
sampler_cls=LocalSampler,
n_workers=meta_batch_size,
worker_class=RL2Worker,
worker_args=dict(n_paths_per_trial=episode_per_task))
runner.train(n_epochs=n_epochs,
batch_size=episode_per_task * max_path_length *
meta_batch_size)
def rl2_ppo_halfcheetah(ctxt=None, seed=1):
"""Train PPO with HalfCheetah environment.
Args:
ctxt (metarl.experiment.ExperimentContext): The experiment
configuration used by LocalRunner to create the snapshotter.
seed (int): Used to seed the random number generator to produce
determinism.
"""
set_seed(seed)
with LocalTFRunner(snapshot_config=ctxt) as runner:
max_path_length = 100
meta_batch_size = 10
n_epochs = 50
episode_per_task = 4
# ---- For ML1-push
from metaworld.benchmarks import ML1
tasks = task_sampler.SetTaskSampler(lambda: RL2Env(
env=ML1.get_train_tasks('push-v1')))
# ---- For HalfCheetahVel
# tasks = task_sampler.SetTaskSampler(lambda: RL2Env(
# env=HalfCheetahVelEnv()))
env_spec = tasks.sample(1)[0]().spec
policy = GaussianGRUPolicy(name='policy',
hidden_dim=64,
env_spec=env_spec,
state_include_action=False)
baseline = LinearFeatureBaseline(env_spec=env_spec)
inner_algo = RL2PPO(
env_spec=env_spec,
policy=policy,
baseline=baseline,
max_path_length=max_path_length * episode_per_task,
discount=0.99,
gae_lambda=0.95,
lr_clip_range=0.2,
optimizer_args=dict(
batch_size=32,
max_epochs=10,
),
stop_entropy_gradient=True,
entropy_method='max',
policy_ent_coeff=0.02,
center_adv=False,
)
algo = RL2(policy=policy,
inner_algo=inner_algo,
max_path_length=max_path_length,
meta_batch_size=meta_batch_size,
task_sampler=tasks)
runner.setup(algo,
tasks.sample(meta_batch_size),
sampler_cls=LocalSampler,
n_workers=meta_batch_size,
worker_class=RL2Worker)
runner.train(n_epochs=n_epochs,
batch_size=episode_per_task * max_path_length *
meta_batch_size)
def test_state_info_specs_with_state_include_action(self):
env = MetaRLEnv(DummyBoxEnv(obs_dim=(4, ), action_dim=(4, )))
policy = GaussianGRUPolicy(env_spec=env.spec,
state_include_action=True)
assert policy.state_info_specs == [('prev_action', (4, ))]
def test_invalid_env(self):
env = TfEnv(DummyDiscreteEnv())
with pytest.raises(ValueError):
GaussianGRUPolicy(env_spec=env.spec)
def run_metarl(env, envs, tasks, seed, log_dir):
"""Create metarl Tensorflow PPO model and training.
Args:
env (dict): Environment of the task.
seed (int): Random positive integer for the trial.
log_dir (str): Log dir path.
Returns:
str: Path to output csv file
"""
deterministic.set_seed(seed)
snapshot_config = SnapshotConfig(snapshot_dir=log_dir,
snapshot_mode='gap',
snapshot_gap=10)
with LocalTFRunner(snapshot_config) as runner:
policy = GaussianGRUPolicy(
hidden_dims=hyper_parameters['hidden_sizes'],
env_spec=env.spec,
state_include_action=False)
baseline = MetaRLLinearFeatureBaseline(env_spec=env.spec)
inner_algo = RL2PPO(
env_spec=env.spec,
policy=policy,
baseline=baseline,
max_path_length=hyper_parameters['max_path_length'] * hyper_parameters['rollout_per_task'],
discount=hyper_parameters['discount'],
gae_lambda=hyper_parameters['gae_lambda'],
lr_clip_range=hyper_parameters['lr_clip_range'],
optimizer_args=dict(
max_epochs=hyper_parameters['optimizer_max_epochs'],
tf_optimizer_args=dict(
learning_rate=hyper_parameters['optimizer_lr'],
),
)
)
# Need to pass this if meta_batch_size < num_of_tasks
task_names = list(ML45_ENVS['train'].keys())
algo = RL2(
policy=policy,
inner_algo=inner_algo,
max_path_length=hyper_parameters['max_path_length'],
meta_batch_size=hyper_parameters['meta_batch_size'],
task_sampler=tasks,
task_names=None if hyper_parameters['meta_batch_size'] >= len(task_names) else task_names)
# Set up logger since we are not using run_experiment
tabular_log_file = osp.join(log_dir, 'progress.csv')
text_log_file = osp.join(log_dir, 'debug.log')
dowel_logger.add_output(dowel.TextOutput(text_log_file))
dowel_logger.add_output(dowel.CsvOutput(tabular_log_file))
dowel_logger.add_output(dowel.StdOutput())
dowel_logger.add_output(dowel.TensorBoardOutput(log_dir))
runner.setup(
algo,
envs,
sampler_cls=hyper_parameters['sampler_cls'],
n_workers=hyper_parameters['meta_batch_size'],
worker_class=RL2Worker,
sampler_args=dict(
use_all_workers=hyper_parameters['use_all_workers']),
worker_args=dict(
n_paths_per_trial=hyper_parameters['rollout_per_task']))
# meta evaluator
env_obs_dim = [env().observation_space.shape[0] for (_, env) in ML45_ENVS['test'].items()]
max_obs_dim = max(env_obs_dim)
ML_test_envs = [
TaskIdWrapper(NormalizedRewardEnv(RL2Env(env(*ML45_ARGS['test'][task]['args'],
**ML45_ARGS['test'][task]['kwargs']), max_obs_dim)), task_id=task_id, task_name=task)
for (task_id, (task, env)) in enumerate(ML45_ENVS['test'].items())
]
test_tasks = task_sampler.EnvPoolSampler(ML_test_envs)
test_tasks.grow_pool(hyper_parameters['n_test_tasks'])
test_task_names = list(ML45_ENVS['test'].keys())
runner.setup_meta_evaluator(test_task_sampler=test_tasks,
n_exploration_traj=hyper_parameters['rollout_per_task'],
n_test_rollouts=hyper_parameters['test_rollout_per_task'],
n_test_tasks=hyper_parameters['n_test_tasks'],
n_workers=hyper_parameters['n_test_tasks'],
test_task_names=None if hyper_parameters['n_test_tasks'] >= len(test_task_names) else test_task_names)
runner.train(n_epochs=hyper_parameters['n_itr'],
batch_size=hyper_parameters['meta_batch_size'] * hyper_parameters['rollout_per_task'] * hyper_parameters['max_path_length'])
dowel_logger.remove_all()
return tabular_log_file
def test_state_info_specs(self):
env = MetaRLEnv(DummyBoxEnv(obs_dim=(4, ), action_dim=(4, )))
policy = GaussianGRUPolicy(env_spec=env.spec,
state_include_action=False)
assert policy.state_info_specs == []
