def test_continuous_mlp_policy(self):
continuous_mlp_policy = ContinuousMLPPolicy(env_spec=self.env,
hidden_sizes=(1, ))
self.sess.run(tf.compat.v1.global_variables_initializer())
obs = self.env.observation_space.high
assert continuous_mlp_policy.get_action(obs)
def test_get_regularizable_vars(self, obs_dim, action_dim):
"""Test get_regularizable_vars method"""
env = TfEnv(DummyBoxEnv(obs_dim=obs_dim, action_dim=action_dim))
policy = ContinuousMLPPolicy(env_spec=env.spec)
reg_vars = policy.get_regularizable_vars()
assert len(reg_vars) == 2
for var in reg_vars:
assert ('bias' not in var.name) and ('output' not in var.name)
def test_ddpg_double_pendulum(self):
"""Test DDPG with Pendulum environment."""
with LocalTFRunner(snapshot_config, sess=self.sess) as runner:
env = MetaRLEnv(gym.make('InvertedDoublePendulum-v2'))
policy = ContinuousMLPPolicy(env_spec=env.spec,
hidden_sizes=[64, 64],
hidden_nonlinearity=tf.nn.relu,
output_nonlinearity=tf.nn.tanh)
exploration_policy = AddOrnsteinUhlenbeckNoise(env.spec,
policy,
sigma=0.2)
qf = ContinuousMLPQFunction(env_spec=env.spec,
hidden_sizes=[64, 64],
hidden_nonlinearity=tf.nn.relu)
replay_buffer = PathBuffer(capacity_in_transitions=int(1e5))
algo = DDPG(
env_spec=env.spec,
policy=policy,
policy_lr=1e-4,
qf_lr=1e-3,
qf=qf,
replay_buffer=replay_buffer,
steps_per_epoch=20,
target_update_tau=1e-2,
n_train_steps=50,
discount=0.9,
min_buffer_size=int(5e3),
exploration_policy=exploration_policy,
)
runner.setup(algo, env)
last_avg_ret = runner.train(n_epochs=10, batch_size=100)
assert last_avg_ret > 60
env.close()
def run_metarl(env, seed, log_dir):
'''
Create metarl model and training.
Replace the ddpg with the algorithm you want to run.
:param env: Environment of the task.
:param seed: Random seed for the trial.
:param log_dir: Log dir path.
:return:
'''
deterministic.set_seed(seed)
with LocalTFRunner(snapshot_config) as runner:
env = TfEnv(normalize(env))
# Set up params for ddpg
action_noise = OUStrategy(env.spec, sigma=params['sigma'])
policy = ContinuousMLPPolicy(
env_spec=env.spec,
hidden_sizes=params['policy_hidden_sizes'],
hidden_nonlinearity=tf.nn.relu,
output_nonlinearity=tf.nn.tanh)
qf = ContinuousMLPQFunction(env_spec=env.spec,
hidden_sizes=params['qf_hidden_sizes'],
hidden_nonlinearity=tf.nn.relu)
replay_buffer = SimpleReplayBuffer(
env_spec=env.spec,
size_in_transitions=params['replay_buffer_size'],
time_horizon=params['n_rollout_steps'])
ddpg = DDPG(env_spec=env.spec,
policy=policy,
qf=qf,
replay_buffer=replay_buffer,
steps_per_epoch=params['steps_per_epoch'],
policy_lr=params['policy_lr'],
qf_lr=params['qf_lr'],
target_update_tau=params['tau'],
n_train_steps=params['n_train_steps'],
discount=params['discount'],
min_buffer_size=int(1e4),
exploration_strategy=action_noise,
policy_optimizer=tf.train.AdamOptimizer,
qf_optimizer=tf.train.AdamOptimizer)
# Set up logger since we are not using run_experiment
tabular_log_file = osp.join(log_dir, 'progress.csv')
tensorboard_log_dir = osp.join(log_dir)
dowel_logger.add_output(dowel.StdOutput())
dowel_logger.add_output(dowel.CsvOutput(tabular_log_file))
dowel_logger.add_output(dowel.TensorBoardOutput(tensorboard_log_dir))
runner.setup(ddpg, env)
runner.train(n_epochs=params['n_epochs'],
batch_size=params['n_rollout_steps'])
dowel_logger.remove_all()
return tabular_log_file
def test_is_pickleable(self, obs_dim, action_dim):
"""Test if ContinuousMLPPolicy is pickleable"""
env = TfEnv(DummyBoxEnv(obs_dim=obs_dim, action_dim=action_dim))
with mock.patch(('metarl.tf.policies.'
'continuous_mlp_policy.MLPModel'),
new=SimpleMLPModel):
policy = ContinuousMLPPolicy(env_spec=env.spec)
env.reset()
obs, _, _, _ = env.step(1)
with tf.compat.v1.variable_scope('ContinuousMLPPolicy/MLPModel',
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.outputs,
feed_dict={policy.model.input: [obs.flatten()]})
p = pickle.dumps(policy)
with tf.compat.v1.Session(graph=tf.Graph()) as sess:
policy_pickled = pickle.loads(p)
output2 = sess.run(
policy_pickled.model.outputs,
feed_dict={policy_pickled.model.input: [obs.flatten()]})
assert np.array_equal(output1, output2)
def her_metarl_tf(ctxt, env_id, seed):
"""Create metarl TensorFlow HER model and training.
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 = ContinuousMLPPolicy(
env_spec=env.spec,
hidden_sizes=hyper_parameters['policy_hidden_sizes'],
hidden_nonlinearity=tf.nn.relu,
output_nonlinearity=tf.nn.tanh,
)
exploration_policy = AddOrnsteinUhlenbeckNoise(
env_spec=env.spec, policy=policy, sigma=hyper_parameters['sigma'])
qf = ContinuousMLPQFunction(
env_spec=env.spec,
hidden_sizes=hyper_parameters['qf_hidden_sizes'],
hidden_nonlinearity=tf.nn.relu,
)
replay_buffer = HERReplayBuffer(
env_spec=env.spec,
capacity_in_transitions=hyper_parameters['replay_buffer_size'],
replay_k=4,
reward_fn=env.compute_reward,
)
algo = DDPG(
env_spec=env.spec,
policy=policy,
qf=qf,
replay_buffer=replay_buffer,
steps_per_epoch=hyper_parameters['steps_per_epoch'],
policy_lr=hyper_parameters['policy_lr'],
qf_lr=hyper_parameters['qf_lr'],
target_update_tau=hyper_parameters['tau'],
n_train_steps=hyper_parameters['n_train_steps'],
discount=hyper_parameters['discount'],
exploration_policy=exploration_policy,
policy_optimizer=tf.compat.v1.train.AdamOptimizer,
qf_optimizer=tf.compat.v1.train.AdamOptimizer,
buffer_batch_size=256,
)
runner.setup(algo, env)
runner.train(n_epochs=hyper_parameters['n_epochs'],
batch_size=hyper_parameters['n_rollout_steps'])
def run_task(snapshot_config, *_):
"""Run task.
Args:
snapshot_config (metarl.experiment.SnapshotConfig): The snapshot
configuration used by LocalRunner to create the snapshotter.
*_ (object): Ignored by this function.
"""
with LocalTFRunner(snapshot_config=snapshot_config) as runner:
env = TfEnv(gym.make('FetchReach-v1'))
action_noise = OUStrategy(env.spec, sigma=0.2)
policy = ContinuousMLPPolicy(
env_spec=env.spec,
name='Policy',
hidden_sizes=[256, 256, 256],
hidden_nonlinearity=tf.nn.relu,
output_nonlinearity=tf.nn.tanh,
input_include_goal=True,
)
qf = ContinuousMLPQFunction(
env_spec=env.spec,
name='QFunction',
hidden_sizes=[256, 256, 256],
hidden_nonlinearity=tf.nn.relu,
input_include_goal=True,
)
replay_buffer = HerReplayBuffer(env_spec=env.spec,
size_in_transitions=int(1e6),
time_horizon=100,
replay_k=0.4,
reward_fun=env.compute_reward)
ddpg = DDPG(
env_spec=env.spec,
policy=policy,
policy_lr=1e-3,
qf_lr=1e-3,
qf=qf,
replay_buffer=replay_buffer,
target_update_tau=0.05,
steps_per_epoch=20,
max_path_length=100,
n_train_steps=40,
discount=0.9,
exploration_strategy=action_noise,
policy_optimizer=tf.train.AdamOptimizer,
qf_optimizer=tf.train.AdamOptimizer,
buffer_batch_size=256,
input_include_goal=True,
)
runner.setup(algo=ddpg, env=env)
runner.train(n_epochs=50, batch_size=100)
def td3_pendulum(ctxt=None, seed=1):
"""Wrap TD3 training task in the run_task function.
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(ctxt) as runner:
env = MetaRLEnv(gym.make('InvertedDoublePendulum-v2'))
policy = ContinuousMLPPolicy(env_spec=env.spec,
hidden_sizes=[400, 300],
hidden_nonlinearity=tf.nn.relu,
output_nonlinearity=tf.nn.tanh)
exploration_policy = AddGaussianNoise(env.spec,
policy,
max_sigma=0.1,
min_sigma=0.1)
qf = ContinuousMLPQFunction(name='ContinuousMLPQFunction',
env_spec=env.spec,
hidden_sizes=[400, 300],
action_merge_layer=0,
hidden_nonlinearity=tf.nn.relu)
qf2 = ContinuousMLPQFunction(name='ContinuousMLPQFunction2',
env_spec=env.spec,
hidden_sizes=[400, 300],
action_merge_layer=0,
hidden_nonlinearity=tf.nn.relu)
replay_buffer = PathBuffer(capacity_in_transitions=int(1e6))
td3 = TD3(env_spec=env.spec,
policy=policy,
policy_lr=1e-4,
qf_lr=1e-3,
qf=qf,
qf2=qf2,
replay_buffer=replay_buffer,
target_update_tau=1e-2,
steps_per_epoch=20,
n_train_steps=1,
smooth_return=False,
discount=0.99,
buffer_batch_size=100,
min_buffer_size=1e4,
exploration_policy=exploration_policy,
policy_optimizer=tf.compat.v1.train.AdamOptimizer,
qf_optimizer=tf.compat.v1.train.AdamOptimizer)
runner.setup(td3, env)
runner.train(n_epochs=500, batch_size=250)
def run_task(snapshot_config, *_):
"""Wrap TD3 training task in the run_task function.
Args:
snapshot_config (metarl.experiment.SnapshotConfig): Configuration
values for snapshotting.
*_ (object): Hyperparameters (unused).
"""
with LocalTFRunner(snapshot_config) as runner:
env = TfEnv(gym.make('InvertedDoublePendulum-v2'))
action_noise = GaussianStrategy(env.spec, max_sigma=0.1, min_sigma=0.1)
policy = ContinuousMLPPolicy(env_spec=env.spec,
hidden_sizes=[400, 300],
hidden_nonlinearity=tf.nn.relu,
output_nonlinearity=tf.nn.tanh)
qf = ContinuousMLPQFunction(name='ContinuousMLPQFunction',
env_spec=env.spec,
hidden_sizes=[400, 300],
action_merge_layer=0,
hidden_nonlinearity=tf.nn.relu)
qf2 = ContinuousMLPQFunction(name='ContinuousMLPQFunction2',
env_spec=env.spec,
hidden_sizes=[400, 300],
action_merge_layer=0,
hidden_nonlinearity=tf.nn.relu)
replay_buffer = SimpleReplayBuffer(env_spec=env.spec,
size_in_transitions=int(1e6),
time_horizon=250)
td3 = TD3(env_spec=env.spec,
policy=policy,
policy_lr=1e-4,
qf_lr=1e-3,
qf=qf,
qf2=qf2,
replay_buffer=replay_buffer,
target_update_tau=1e-2,
steps_per_epoch=20,
n_train_steps=1,
smooth_return=False,
discount=0.99,
buffer_batch_size=100,
min_buffer_size=1e4,
exploration_strategy=action_noise,
policy_optimizer=tf.train.AdamOptimizer,
qf_optimizer=tf.train.AdamOptimizer)
runner.setup(td3, env)
runner.train(n_epochs=500, batch_size=250)
def test_td3_pendulum(self):
"""Test TD3 with Pendulum environment."""
with LocalTFRunner(snapshot_config) as runner:
env = TfEnv(gym.make('InvertedDoublePendulum-v2'))
action_noise = GaussianStrategy(env.spec,
max_sigma=0.1,
min_sigma=0.1)
policy = ContinuousMLPPolicy(env_spec=env.spec,
hidden_sizes=[400, 300],
hidden_nonlinearity=tf.nn.relu,
output_nonlinearity=tf.nn.tanh)
qf = ContinuousMLPQFunction(name='ContinuousMLPQFunction',
env_spec=env.spec,
hidden_sizes=[400, 300],
action_merge_layer=0,
hidden_nonlinearity=tf.nn.relu)
qf2 = ContinuousMLPQFunction(name='ContinuousMLPQFunction2',
env_spec=env.spec,
hidden_sizes=[400, 300],
action_merge_layer=0,
hidden_nonlinearity=tf.nn.relu)
replay_buffer = SimpleReplayBuffer(env_spec=env.spec,
size_in_transitions=int(1e6),
time_horizon=250)
algo = TD3(env_spec=env.spec,
policy=policy,
policy_lr=1e-3,
qf_lr=1e-3,
qf=qf,
qf2=qf2,
replay_buffer=replay_buffer,
steps_per_epoch=20,
target_update_tau=0.005,
n_train_steps=50,
discount=0.99,
smooth_return=False,
min_buffer_size=int(1e4),
buffer_batch_size=100,
policy_weight_decay=0.001,
qf_weight_decay=0.001,
exploration_strategy=action_noise,
policy_optimizer=tf.compat.v1.train.AdamOptimizer,
qf_optimizer=tf.compat.v1.train.AdamOptimizer)
runner.setup(algo, env)
last_avg_ret = runner.train(n_epochs=10, batch_size=250)
assert last_avg_ret > 400
def continuous_mlp_q_function(ctxt, env_id, seed):
"""Create Continuous MLP QFunction on TF-DDPG.
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, max_cpus=12) as runner:
env = MetaRLEnv(normalize(gym.make(env_id)))
policy = ContinuousMLPPolicy(
env_spec=env.spec,
name='ContinuousMLPPolicy',
hidden_sizes=hyper_params['policy_hidden_sizes'],
hidden_nonlinearity=tf.nn.relu,
output_nonlinearity=tf.nn.tanh)
exploration_policy = AddOrnsteinUhlenbeckNoise(
env.spec, policy, sigma=hyper_params['sigma'])
qf = ContinuousMLPQFunction(
env_spec=env.spec,
hidden_sizes=hyper_params['qf_hidden_sizes'],
hidden_nonlinearity=tf.nn.relu,
name='ContinuousMLPQFunction')
replay_buffer = PathBuffer(
capacity_in_transitions=hyper_params['replay_buffer_size'])
ddpg = DDPG(env_spec=env.spec,
policy=policy,
qf=qf,
replay_buffer=replay_buffer,
steps_per_epoch=hyper_params['steps_per_epoch'],
policy_lr=hyper_params['policy_lr'],
qf_lr=hyper_params['qf_lr'],
target_update_tau=hyper_params['tau'],
n_train_steps=hyper_params['n_train_steps'],
discount=hyper_params['discount'],
min_buffer_size=int(1e4),
exploration_policy=exploration_policy,
policy_optimizer=tf.compat.v1.train.AdamOptimizer,
qf_optimizer=tf.compat.v1.train.AdamOptimizer)
runner.setup(ddpg, env, sampler_args=dict(n_envs=12))
runner.train(n_epochs=hyper_params['n_epochs'],
batch_size=hyper_params['n_rollout_steps'])
def test_get_action_sym(self, obs_dim, action_dim):
"""Test get_action_sym method"""
env = TfEnv(DummyBoxEnv(obs_dim=obs_dim, action_dim=action_dim))
with mock.patch(('metarl.tf.policies.'
'continuous_mlp_policy.MLPModel'),
new=SimpleMLPModel):
policy = ContinuousMLPPolicy(env_spec=env.spec)
env.reset()
obs, _, _, _ = env.step(1)
obs_dim = env.spec.observation_space.flat_dim
state_input = tf.compat.v1.placeholder(tf.float32,
shape=(None, obs_dim))
action_sym = policy.get_action_sym(state_input, name='action_sym')
expected_action = np.full(action_dim, 0.5)
action = self.sess.run(action_sym,
feed_dict={state_input: [obs.flatten()]})
action = policy.action_space.unflatten(action)
assert np.array_equal(action, expected_action)
assert env.action_space.contains(action)
def test_get_action(self, obs_dim, action_dim):
"""Test get_action method"""
env = TfEnv(DummyBoxEnv(obs_dim=obs_dim, action_dim=action_dim))
with mock.patch(('metarl.tf.policies.'
'continuous_mlp_policy.MLPModel'),
new=SimpleMLPModel):
policy = ContinuousMLPPolicy(env_spec=env.spec)
env.reset()
obs, _, _, _ = env.step(1)
action, _ = policy.get_action(obs.flatten())
expected_action = np.full(action_dim, 0.5)
assert env.action_space.contains(action)
assert np.array_equal(action, expected_action)
actions, _ = policy.get_actions(
[obs.flatten(), obs.flatten(),
obs.flatten()])
for action in actions:
assert env.action_space.contains(action)
assert np.array_equal(action, expected_action)
def ddpg_pendulum(ctxt=None, seed=1):
"""Train DDPG with InvertedDoublePendulum-v2 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:
env = MetaRLEnv(gym.make('InvertedDoublePendulum-v2'))
policy = ContinuousMLPPolicy(env_spec=env.spec,
hidden_sizes=[64, 64],
hidden_nonlinearity=tf.nn.relu,
output_nonlinearity=tf.nn.tanh)
exploration_policy = AddOrnsteinUhlenbeckNoise(env.spec,
policy,
sigma=0.2)
qf = ContinuousMLPQFunction(env_spec=env.spec,
hidden_sizes=[64, 64],
hidden_nonlinearity=tf.nn.relu)
replay_buffer = PathBuffer(capacity_in_transitions=int(1e6))
ddpg = DDPG(env_spec=env.spec,
policy=policy,
policy_lr=1e-4,
qf_lr=1e-3,
qf=qf,
replay_buffer=replay_buffer,
steps_per_epoch=20,
target_update_tau=1e-2,
n_train_steps=50,
discount=0.9,
min_buffer_size=int(1e4),
exploration_policy=exploration_policy,
policy_optimizer=tf.compat.v1.train.AdamOptimizer,
qf_optimizer=tf.compat.v1.train.AdamOptimizer)
runner.setup(algo=ddpg, env=env)
runner.train(n_epochs=500, batch_size=100)
def test_ddpg_pendulum_with_decayed_weights(self):
"""Test DDPG with Pendulum environment and decayed weights.
This environment has a [-3, 3] action_space bound.
"""
with LocalTFRunner(snapshot_config, sess=self.sess) as runner:
env = TfEnv(normalize(gym.make('InvertedPendulum-v2')))
action_noise = OUStrategy(env.spec, sigma=0.2)
policy = ContinuousMLPPolicy(env_spec=env.spec,
hidden_sizes=[64, 64],
hidden_nonlinearity=tf.nn.relu,
output_nonlinearity=tf.nn.tanh)
qf = ContinuousMLPQFunction(env_spec=env.spec,
hidden_sizes=[64, 64],
hidden_nonlinearity=tf.nn.relu)
replay_buffer = SimpleReplayBuffer(env_spec=env.spec,
size_in_transitions=int(1e6),
time_horizon=100)
algo = DDPG(
env_spec=env.spec,
policy=policy,
policy_lr=1e-4,
qf_lr=1e-3,
qf=qf,
replay_buffer=replay_buffer,
steps_per_epoch=20,
target_update_tau=1e-2,
n_train_steps=50,
discount=0.9,
policy_weight_decay=0.01,
qf_weight_decay=0.01,
min_buffer_size=int(5e3),
exploration_strategy=action_noise,
)
runner.setup(algo, env)
last_avg_ret = runner.train(n_epochs=10, batch_size=100)
assert last_avg_ret > 10
env.close()
def td3_metarl_tf(ctxt, env_id, seed):
"""Create metarl TensorFlow TD3 model and training.
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 = ContinuousMLPPolicy(
env_spec=env.spec,
hidden_sizes=hyper_parameters['policy_hidden_sizes'],
hidden_nonlinearity=tf.nn.relu,
output_nonlinearity=tf.nn.tanh)
exploration_policy = AddGaussianNoise(
env.spec,
policy,
max_sigma=hyper_parameters['sigma'],
min_sigma=hyper_parameters['sigma'])
qf = ContinuousMLPQFunction(
name='ContinuousMLPQFunction',
env_spec=env.spec,
hidden_sizes=hyper_parameters['qf_hidden_sizes'],
action_merge_layer=0,
hidden_nonlinearity=tf.nn.relu)
qf2 = ContinuousMLPQFunction(
name='ContinuousMLPQFunction2',
env_spec=env.spec,
hidden_sizes=hyper_parameters['qf_hidden_sizes'],
action_merge_layer=0,
hidden_nonlinearity=tf.nn.relu)
replay_buffer = PathBuffer(
capacity_in_transitions=hyper_parameters['replay_buffer_size'])
td3 = TD3(env.spec,
policy=policy,
qf=qf,
qf2=qf2,
replay_buffer=replay_buffer,
steps_per_epoch=hyper_parameters['steps_per_epoch'],
policy_lr=hyper_parameters['policy_lr'],
qf_lr=hyper_parameters['qf_lr'],
target_update_tau=hyper_parameters['tau'],
n_train_steps=hyper_parameters['n_train_steps'],
discount=hyper_parameters['discount'],
smooth_return=hyper_parameters['smooth_return'],
min_buffer_size=hyper_parameters['min_buffer_size'],
buffer_batch_size=hyper_parameters['buffer_batch_size'],
exploration_policy=exploration_policy,
policy_optimizer=tf.compat.v1.train.AdamOptimizer,
qf_optimizer=tf.compat.v1.train.AdamOptimizer)
runner.setup(td3, env)
runner.train(n_epochs=hyper_parameters['n_epochs'],
batch_size=hyper_parameters['n_rollout_steps'])
def test_no_reset(self):
with LocalTFRunner(snapshot_config, sess=self.sess) as runner:
# This tests if off-policy sampler respect batch_size
# when no_reset is set to True
env = MetaRLEnv(normalize(gym.make('InvertedDoublePendulum-v2')))
policy = ContinuousMLPPolicy(env_spec=env.spec,
hidden_sizes=[64, 64],
hidden_nonlinearity=tf.nn.relu,
output_nonlinearity=tf.nn.tanh)
exploration_policy = AddOrnsteinUhlenbeckNoise(env.spec,
policy,
sigma=0.2)
qf = ContinuousMLPQFunction(env_spec=env.spec,
hidden_sizes=[64, 64],
hidden_nonlinearity=tf.nn.relu)
replay_buffer = PathBuffer(capacity_in_transitions=int(1e6))
algo = DDPG(
env_spec=env.spec,
policy=policy,
policy_lr=1e-4,
qf_lr=1e-3,
qf=qf,
replay_buffer=replay_buffer,
target_update_tau=1e-2,
n_train_steps=50,
discount=0.9,
min_buffer_size=int(1e4),
exploration_policy=exploration_policy,
)
sampler = OffPolicyVectorizedSampler(algo, env, 1, no_reset=True)
sampler.start_worker()
runner.initialize_tf_vars()
paths1 = sampler.obtain_samples(0, 5)
paths2 = sampler.obtain_samples(0, 5)
len1 = sum([len(path['rewards']) for path in paths1])
len2 = sum([len(path['rewards']) for path in paths2])
assert len1 == 5 and len2 == 5, 'Sampler should respect batch_size'
# yapf: disable
# When done is False in 1st sampling, the next sampling should be
# stacked with the last batch in 1st sampling
case1 = (len(paths1[-1]['rewards']) + len(paths2[0]['rewards'])
== paths2[0]['running_length'])
# When done is True in 1st sampling, the next sampling should be
# separated
case2 = len(paths2[0]['rewards']) == paths2[0]['running_length']
done = paths1[-1]['dones'][-1]
assert (
(not done and case1) or (done and case2)
), 'Running length should be the length of full path'
# yapf: enable
case1 = np.isclose(
paths1[-1]['rewards'].sum() + paths2[0]['rewards'].sum(),
paths2[0]['undiscounted_return'])
case2 = np.isclose(paths2[0]['rewards'].sum(),
paths2[0]['undiscounted_return'])
assert (
(not done and case1) or (done and case2)
), 'Undiscounted_return should be the sum of rewards of full path'
def run_metarl(env, seed, log_dir):
"""Create metarl model and training.
Replace the ddpg with the algorithm you want to run.
Args:
env (gym.Env): Environment of the task.
seed (int): Random seed for the trial.
log_dir (str): Log dir path.
Returns:
str: Log file path.
"""
deterministic.set_seed(seed)
config = tf.ConfigProto(allow_soft_placement=True,
intra_op_parallelism_threads=12,
inter_op_parallelism_threads=12)
sess = tf.Session(config=config)
with LocalTFRunner(snapshot_config, sess=sess, max_cpus=12) as runner:
env = TfEnv(normalize(env))
# Set up params for ddpg
action_noise = OUStrategy(env.spec, sigma=params['sigma'])
policy = ContinuousMLPPolicy(
env_spec=env.spec,
name='ContinuousMLPPolicy',
hidden_sizes=params['policy_hidden_sizes'],
hidden_nonlinearity=tf.nn.relu,
output_nonlinearity=tf.nn.tanh)
qf = ContinuousMLPQFunction(env_spec=env.spec,
hidden_sizes=params['qf_hidden_sizes'],
hidden_nonlinearity=tf.nn.relu,
name='ContinuousMLPQFunction')
replay_buffer = SimpleReplayBuffer(
env_spec=env.spec,
size_in_transitions=params['replay_buffer_size'],
time_horizon=params['n_rollout_steps'])
ddpg = DDPG(env_spec=env.spec,
policy=policy,
qf=qf,
replay_buffer=replay_buffer,
steps_per_epoch=params['steps_per_epoch'],
policy_lr=params['policy_lr'],
qf_lr=params['qf_lr'],
target_update_tau=params['tau'],
n_train_steps=params['n_train_steps'],
discount=params['discount'],
min_buffer_size=int(1e4),
exploration_strategy=action_noise,
policy_optimizer=tf.train.AdamOptimizer,
qf_optimizer=tf.train.AdamOptimizer)
# Set up logger since we are not using run_experiment
tabular_log_file = osp.join(log_dir, 'progress.csv')
dowel_logger.add_output(dowel.StdOutput())
dowel_logger.add_output(dowel.CsvOutput(tabular_log_file))
dowel_logger.add_output(dowel.TensorBoardOutput(log_dir))
runner.setup(ddpg, env, sampler_args=dict(n_envs=12))
runner.train(n_epochs=params['n_epochs'],
batch_size=params['n_rollout_steps'])
dowel_logger.remove_all()
return tabular_log_file