def dqn_cartpole(ctxt=None, seed=1):
"""Train TRPO with CubeCrash-v0 environment.
Args:
ctxt (garage.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:
n_epochs = 10
steps_per_epoch = 10
sampler_batch_size = 500
num_timesteps = n_epochs * steps_per_epoch * sampler_batch_size
env = GarageEnv(gym.make('CartPole-v0'))
replay_buffer = PathBuffer(capacity_in_transitions=int(1e4))
qf = DiscreteMLPQFunction(env_spec=env.spec, hidden_sizes=(64, 64))
policy = DiscreteQfDerivedPolicy(env_spec=env.spec, qf=qf)
exploration_policy = EpsilonGreedyPolicy(env_spec=env.spec,
policy=policy,
total_timesteps=num_timesteps,
max_epsilon=1.0,
min_epsilon=0.02,
decay_ratio=0.1)
algo = DQN(env_spec=env.spec,
policy=policy,
qf=qf,
exploration_policy=exploration_policy,
max_episode_length=100,
replay_buffer=replay_buffer,
steps_per_epoch=steps_per_epoch,
qf_lr=1e-4,
discount=1.0,
min_buffer_size=int(1e3),
double_q=True,
n_train_steps=500,
target_network_update_freq=1,
buffer_batch_size=32)
runner.setup(algo, env)
runner.train(n_epochs=n_epochs, batch_size=sampler_batch_size)
def test_fixed_alpha():
"""Test if using fixed_alpha ensures that alpha is non differentiable."""
# pylint: disable=unexpected-keyword-arg
env = normalize(GymEnv('InvertedDoublePendulum-v2',
max_episode_length=100))
deterministic.set_seed(0)
policy = TanhGaussianMLPPolicy(
env_spec=env.spec,
hidden_sizes=[32, 32],
hidden_nonlinearity=torch.nn.ReLU,
output_nonlinearity=None,
min_std=np.exp(-20.),
max_std=np.exp(2.),
)
qf1 = ContinuousMLPQFunction(env_spec=env.spec,
hidden_sizes=[32, 32],
hidden_nonlinearity=F.relu)
qf2 = ContinuousMLPQFunction(env_spec=env.spec,
hidden_sizes=[32, 32],
hidden_nonlinearity=F.relu)
replay_buffer = PathBuffer(capacity_in_transitions=int(1e6), )
runner = LocalRunner(snapshot_config=snapshot_config)
sac = SAC(env_spec=env.spec,
policy=policy,
qf1=qf1,
qf2=qf2,
gradient_steps_per_itr=100,
replay_buffer=replay_buffer,
min_buffer_size=100,
target_update_tau=5e-3,
discount=0.99,
buffer_batch_size=64,
reward_scale=1.,
steps_per_epoch=1,
fixed_alpha=np.exp(0.5))
runner.setup(sac, env, sampler_cls=LocalSampler)
sac.to()
runner.train(n_epochs=1, batch_size=100, plot=False)
assert torch.allclose(torch.Tensor([0.5]), sac._log_alpha.cpu())
assert not sac._use_automatic_entropy_tuning
def test_pickling(self):
"""Test pickle and unpickle."""
deterministic.set_seed(0)
n_epochs = 10
steps_per_epoch = 20
sampler_batch_size = 100
num_timesteps = n_epochs * steps_per_epoch * sampler_batch_size
env = normalize(
GymEnv('InvertedDoublePendulum-v2', max_episode_length=100))
policy = DeterministicMLPPolicy(env_spec=env.spec,
hidden_sizes=[64, 64],
hidden_nonlinearity=F.relu,
output_nonlinearity=None)
exploration_policy = AddGaussianNoise(env.spec,
policy,
total_timesteps=num_timesteps,
max_sigma=0.1,
min_sigma=0.1)
qf1 = ContinuousMLPQFunction(env_spec=env.spec,
hidden_sizes=[256, 256],
hidden_nonlinearity=F.relu)
qf2 = ContinuousMLPQFunction(env_spec=env.spec,
hidden_sizes=[256, 256],
hidden_nonlinearity=F.relu)
replay_buffer = PathBuffer(capacity_in_transitions=int(1e6))
td3 = TD3(env_spec=env.spec,
policy=policy,
qf1=qf1,
qf2=qf2,
replay_buffer=replay_buffer,
exploration_policy=exploration_policy,
steps_per_epoch=steps_per_epoch,
grad_steps_per_env_step=1,
num_evaluation_episodes=1,
discount=0.99)
prefer_gpu()
td3.to()
pickled = pickle.dumps(td3)
unpickled = pickle.loads(pickled)
assert unpickled
def trpo_cubecrash(ctxt=None, seed=1, batch_size=4000):
"""Train TRPO with CubeCrash-v0 environment.
Args:
ctxt (garage.experiment.ExperimentContext): The experiment
configuration used by LocalRunner to create the snapshotter.
seed (int): Used to seed the random number generator to produce
determinism.
batch_size (int): Number of timesteps to use in each training step.
"""
set_seed(seed)
with LocalTFRunner(ctxt) as runner:
env = TfEnv(normalize(gym.make('CubeCrash-v0')))
policy = CategoricalCNNPolicy(env_spec=env.spec,
num_filters=(32, 64),
filter_dims=(8, 4),
strides=(4, 2),
padding='VALID',
hidden_sizes=(32, 32))
baseline = GaussianCNNBaseline(env_spec=env.spec,
regressor_args=dict(
num_filters=(32, 64),
filter_dims=(8, 4),
strides=(4, 2),
padding='VALID',
hidden_sizes=(32, 32),
use_trust_region=True))
algo = TRPO(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,
flatten_input=False)
runner.setup(algo, env)
runner.train(n_epochs=100, batch_size=batch_size)
def maml_trpo(ctxt, seed, epochs, rollouts_per_task, meta_batch_size):
"""Set up environment and algorithm and run the task.
Args:
ctxt (garage.experiment.ExperimentContext): The experiment
configuration used by LocalRunner to create the snapshotter.
seed (int): Used to seed the random number generator to produce
determinism.
epochs (int): Number of training epochs.
rollouts_per_task (int): Number of rollouts per epoch per task
for training.
meta_batch_size (int): Number of tasks sampled per batch.
"""
set_seed(seed)
env = GarageEnv(normalize(HalfCheetahDirEnv(), expected_action_scale=10.))
policy = GaussianMLPPolicy(
env_spec=env.spec,
hidden_sizes=(64, 64),
hidden_nonlinearity=torch.tanh,
output_nonlinearity=None,
)
baseline = LinearFeatureBaseline(env_spec=env.spec)
max_path_length = 100
runner = LocalRunner(ctxt)
algo = MAMLTRPO(env=env,
policy=policy,
baseline=baseline,
max_path_length=max_path_length,
meta_batch_size=meta_batch_size,
discount=0.99,
gae_lambda=1.,
inner_lr=0.1,
num_grad_updates=1)
runner.setup(algo, env)
runner.train(n_epochs=epochs,
batch_size=rollouts_per_task * max_path_length)
def test_ddpg_pendulum(self):
"""Test DDPG with Pendulum environment.
This environment has a [-3, 3] action_space bound.
"""
deterministic.set_seed(0)
runner = LocalRunner(snapshot_config)
env = GarageEnv(normalize(gym.make('InvertedPendulum-v2')))
policy = DeterministicMLPPolicy(env_spec=env.spec,
hidden_sizes=[64, 64],
hidden_nonlinearity=F.relu,
output_nonlinearity=torch.tanh)
exploration_policy = AddOrnsteinUhlenbeckNoise(env.spec,
policy,
sigma=0.2)
qf = ContinuousMLPQFunction(env_spec=env.spec,
hidden_sizes=[64, 64],
hidden_nonlinearity=F.relu)
replay_buffer = SimpleReplayBuffer(env_spec=env.spec,
size_in_transitions=int(1e6),
time_horizon=100)
algo = DDPG(env_spec=env.spec,
policy=policy,
qf=qf,
replay_buffer=replay_buffer,
steps_per_epoch=20,
n_train_steps=50,
min_buffer_size=int(1e4),
exploration_policy=exploration_policy,
target_update_tau=1e-2,
discount=0.9)
runner.setup(algo, env)
last_avg_ret = runner.train(n_epochs=10, batch_size=100)
assert last_avg_ret > 10
env.close()
def categorical_lstm_policy(ctxt, env_id, seed):
"""Create Categorical LSTM Policy on TF-PPO.
Args:
ctxt (garage.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 = normalize(GymEnv(env_id))
policy = CategoricalLSTMPolicy(
env_spec=env.spec,
hidden_dim=32,
hidden_nonlinearity=tf.nn.tanh,
)
baseline = LinearFeatureBaseline(env_spec=env.spec)
algo = PPO(
env_spec=env.spec,
policy=policy,
baseline=baseline,
max_episode_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_episode_length=10,
learning_rate=1e-3,
),
)
runner.setup(algo, env, sampler_args=dict(n_envs=12))
runner.train(n_epochs=488, batch_size=2048)
def tf_gym_music(ctxt=None, seed=1):
"""Train Policy Gradient LSTM with Music-v0 environment.
Args:
ctxt (garage.experiment.ExperimentContext): The experiment
configuration used by Trainer to create the snapshotter.
created by @wrap_experiment
seed (int): Used to seed the random number generator to produce
determinism.
"""
set_seed(seed)
with TFTrainer(snapshot_config=ctxt) as trainer:
env = GymEnv(MusicEnv(monitor = HeartMonitor('DC:39:39:66:26:1F')),max_episode_length = 35)
policy = GaussianLSTMPolicy(name='policy',
env_spec=env.spec,
hidden_dim= 32)
baseline = GaussianMLPBaseline(
env_spec = env.spec,
hidden_sizes=(32, 32),
)
sampler = LocalSampler(agents=policy,
envs=env,
max_episode_length=env.spec.max_episode_length,
is_tf_worker=False,
n_workers = 1,
)
algo = NPO(env_spec = env.spec,
policy = policy,
baseline = baseline,
sampler = sampler,
)
trainer.setup(algo, env)
trainer.train(n_epochs=120, batch_size=1,store_episodes = True)
def test_ddpg_double_pendulum(self):
"""Test DDPG with Pendulum environment."""
deterministic.set_seed(0)
trainer = Trainer(snapshot_config)
env = GymEnv('InvertedDoublePendulum-v2', max_episode_length=100)
policy = DeterministicMLPPolicy(env_spec=env.spec,
hidden_sizes=[64, 64],
hidden_nonlinearity=F.relu,
output_nonlinearity=torch.tanh)
exploration_policy = AddOrnsteinUhlenbeckNoise(env.spec,
policy,
sigma=0.2)
qf = ContinuousMLPQFunction(env_spec=env.spec,
hidden_sizes=[64, 64],
hidden_nonlinearity=F.relu)
replay_buffer = PathBuffer(capacity_in_transitions=int(1e6))
sampler = LocalSampler(agents=exploration_policy,
envs=env,
max_episode_length=env.spec.max_episode_length,
worker_class=FragmentWorker)
algo = DDPG(env_spec=env.spec,
policy=policy,
qf=qf,
replay_buffer=replay_buffer,
sampler=sampler,
steps_per_epoch=20,
n_train_steps=50,
min_buffer_size=int(1e4),
exploration_policy=exploration_policy,
target_update_tau=1e-2,
discount=0.9)
trainer.setup(algo, env)
last_avg_ret = trainer.train(n_epochs=10, batch_size=100)
assert last_avg_ret > 45
env.close()
def vpg_garage_tf(ctxt, env_id, seed):
"""Create garage TensorFlow VPG model and training.
Args:
ctxt (garage.experiment.ExperimentContext): The experiment
configuration used by Trainer 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 TFTrainer(ctxt) as trainer:
env = normalize(GymEnv(env_id))
policy = TF_GMP(
env_spec=env.spec,
hidden_sizes=hyper_parameters['hidden_sizes'],
hidden_nonlinearity=tf.nn.tanh,
output_nonlinearity=None,
)
baseline = LinearFeatureBaseline(env_spec=env.spec)
sampler = RaySampler(agents=policy,
envs=env,
max_episode_length=env.spec.max_episode_length,
is_tf_worker=True)
algo = TF_VPG(env_spec=env.spec,
policy=policy,
baseline=baseline,
sampler=sampler,
discount=hyper_parameters['discount'],
center_adv=hyper_parameters['center_adv'],
optimizer_args=dict(
learning_rate=hyper_parameters['learning_rate'], ))
trainer.setup(algo, env)
trainer.train(n_epochs=hyper_parameters['n_epochs'],
batch_size=hyper_parameters['batch_size'])
def __init__(self,
algo,
env,
seed,
should_render=False,
num_processors=None,
sampler_worker_cls=None):
self._sampler_worker = ray.remote(SamplerWorker if sampler_worker_cls
is None else sampler_worker_cls)
self._env = env
self._algo = algo
self._seed = seed
deterministic.set_seed(self._seed)
self._max_path_length = self._algo.max_path_length
self._should_render = should_render
if not ray.is_initialized():
ray.init(log_to_driver=False)
self._num_workers = (num_processors if num_processors else
psutil.cpu_count(logical=False))
self._all_workers = defaultdict(None)
def trpo_garage_tf(ctxt, env_id, seed):
"""Create garage Tensorflow TROI model and training.
Args:
ctxt (garage.experiment.ExperimentContext): The experiment
configuration used by Trainer 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 TFTrainer(ctxt) as trainer:
env = normalize(GymEnv(env_id))
policy = GaussianMLPPolicy(
env_spec=env.spec,
hidden_sizes=hyper_parameters['hidden_sizes'],
hidden_nonlinearity=tf.nn.tanh,
output_nonlinearity=None,
)
baseline = LinearFeatureBaseline(env_spec=env.spec)
sampler = RaySampler(agents=policy,
envs=env,
max_episode_length=env.spec.max_episode_length,
is_tf_worker=True)
algo = TRPO(env_spec=env.spec,
policy=policy,
baseline=baseline,
sampler=sampler,
discount=hyper_parameters['discount'],
gae_lambda=hyper_parameters['gae_lambda'],
max_kl_step=hyper_parameters['max_kl'])
trainer.setup(algo, env)
trainer.train(n_epochs=hyper_parameters['n_epochs'],
batch_size=hyper_parameters['batch_size'])
def trpo_pendulum_ray_sampler(ctxt=None, seed=1):
"""Set up environment and algorithm and run the task.
Args:
ctxt (garage.experiment.ExperimentContext): The experiment
configuration used by LocalRunner to create the snapshotter.
seed (int): Used to seed the random number generator to produce
determinism.
"""
# Since this is an example, we are running ray in a reduced state.
# One can comment this line out in order to run ray at full capacity
ray.init(memory=52428800,
object_store_memory=78643200,
ignore_reinit_error=True,
log_to_driver=False,
include_webui=False)
deterministic.set_seed(seed)
env = GarageEnv(env_name='InvertedDoublePendulum-v2')
runner = LocalRunner(ctxt)
policy = GaussianMLPPolicy(env.spec,
hidden_sizes=[32, 32],
hidden_nonlinearity=torch.tanh,
output_nonlinearity=None)
value_function = GaussianMLPValueFunction(env_spec=env.spec,
hidden_sizes=(32, 32),
hidden_nonlinearity=torch.tanh,
output_nonlinearity=None)
algo = TRPO(env_spec=env.spec,
policy=policy,
value_function=value_function,
max_path_length=100,
discount=0.99,
center_adv=False)
runner.setup(algo, env, sampler_cls=RaySampler)
runner.train(n_epochs=100, batch_size=1024)
def test_dqn_cartpole_grad_clip(self):
"""Test DQN with CartPole environment."""
deterministic.set_seed(100)
with LocalTFRunner(snapshot_config, sess=self.sess) as runner:
n_epochs = 10
steps_per_epoch = 10
sampler_batch_size = 500
num_timesteps = n_epochs * steps_per_epoch * sampler_batch_size
env = GarageEnv(gym.make('CartPole-v0'))
replay_buffer = PathBuffer(capacity_in_transitions=int(1e4))
qf = DiscreteMLPQFunction(env_spec=env.spec, hidden_sizes=(64, 64))
policy = DiscreteQfDerivedPolicy(env_spec=env.spec, qf=qf)
epilson_greedy_policy = EpsilonGreedyPolicy(
env_spec=env.spec,
policy=policy,
total_timesteps=num_timesteps,
max_epsilon=1.0,
min_epsilon=0.02,
decay_ratio=0.1)
algo = DQN(env_spec=env.spec,
policy=policy,
qf=qf,
exploration_policy=epilson_greedy_policy,
replay_buffer=replay_buffer,
max_path_length=100,
qf_lr=1e-4,
discount=1.0,
min_buffer_size=int(1e3),
double_q=False,
n_train_steps=500,
grad_norm_clipping=5.0,
steps_per_epoch=steps_per_epoch,
target_network_update_freq=1,
buffer_batch_size=32)
runner.setup(algo, env)
last_avg_ret = runner.train(n_epochs=n_epochs,
batch_size=sampler_batch_size)
assert last_avg_ret > 9
env.close()
def trpo_garage_pytorch(ctxt, env_id, seed):
"""Create garage PyTorch TRPO model and training.
Args:
ctxt (garage.experiment.ExperimentContext): The experiment
configuration used by Trainer to create the
snapshotter.
env_id (str): Environment id of the task.
seed (int): Random positive integer for the trial.
"""
deterministic.set_seed(seed)
trainer = Trainer(ctxt)
env = normalize(GymEnv(env_id))
policy = PyTorch_GMP(env.spec,
hidden_sizes=hyper_parameters['hidden_sizes'],
hidden_nonlinearity=torch.tanh,
output_nonlinearity=None)
value_function = GaussianMLPValueFunction(env_spec=env.spec,
hidden_sizes=(32, 32),
hidden_nonlinearity=torch.tanh,
output_nonlinearity=None)
sampler = LocalSampler(agents=policy,
envs=env,
max_episode_length=env.spec.max_episode_length)
algo = PyTorch_TRPO(env_spec=env.spec,
policy=policy,
value_function=value_function,
sampler=sampler,
discount=hyper_parameters['discount'],
gae_lambda=hyper_parameters['gae_lambda'])
trainer.setup(algo, env)
trainer.train(n_epochs=hyper_parameters['n_epochs'],
batch_size=hyper_parameters['batch_size'])
def ppo_memorize_digits(ctxt=None, seed=1, batch_size=4000):
"""Train PPO on MemorizeDigits-v0 environment.
Args:
ctxt (garage.experiment.ExperimentContext): The experiment
configuration used by LocalRunner to create the snapshotter.
seed (int): Used to seed the random number generator to produce
determinism.
batch_size (int): Number of timesteps to use in each training step.
"""
set_seed(seed)
with LocalTFRunner(ctxt) as runner:
env = TfEnv(normalize(gym.make('MemorizeDigits-v0')), is_image=True)
policy = CategoricalCNNPolicy(env_spec=env.spec,
conv_filters=(32, 64, 64),
conv_filter_sizes=(5, 3, 2),
conv_strides=(4, 2, 1),
conv_pad='VALID',
hidden_sizes=(256, ))
baseline = GaussianCNNBaseline(env_spec=env.spec,
regressor_args=dict(
num_filters=(32, 64, 64),
filter_dims=(5, 3, 2),
strides=(4, 2, 1),
padding='VALID',
hidden_sizes=(256, ),
use_trust_region=True))
algo = PPO(env_spec=env.spec,
policy=policy,
baseline=baseline,
max_path_length=100,
discount=0.99,
max_kl_step=0.01,
flatten_input=False)
runner.setup(algo, env)
runner.train(n_epochs=1000, batch_size=batch_size)
def trpo_swimmer_ray_sampler(ctxt=None, seed=1):
"""tf_trpo_swimmer.
Args:
ctxt (garage.experiment.ExperimentContext): The experiment
configuration used by Trainer to create the snapshotter.
seed (int): Used to seed the random number generator to produce
determinism.
"""
# Since this is an example, we are running ray in a reduced state.
# One can comment this line out in order to run ray at full capacity
ray.init(_memory=52428800,
object_store_memory=78643200,
ignore_reinit_error=True,
log_to_driver=False,
include_dashboard=False)
with TFTrainer(snapshot_config=ctxt) as trainer:
set_seed(seed)
env = GymEnv('Swimmer-v2')
policy = GaussianMLPPolicy(env_spec=env.spec, hidden_sizes=(32, 32))
baseline = LinearFeatureBaseline(env_spec=env.spec)
sampler = RaySampler(agents=policy,
envs=env,
max_episode_length=env.spec.max_episode_length,
is_tf_worker=True)
algo = TRPO(env_spec=env.spec,
policy=policy,
baseline=baseline,
sampler=sampler,
discount=0.99,
max_kl_step=0.01)
trainer.setup(algo, env)
trainer.train(n_epochs=40, batch_size=4000)
def test_td3_inverted_double_pendulum(self):
deterministic.set_seed(0)
n_epochs = 10
steps_per_epoch = 20
sampler_batch_size = 100
num_timesteps = n_epochs * steps_per_epoch * sampler_batch_size
trainer = Trainer(snapshot_config=snapshot_config)
env = normalize(
GymEnv('InvertedDoublePendulum-v2', max_episode_length=100))
policy = DeterministicMLPPolicy(env_spec=env.spec,
hidden_sizes=[64, 64],
hidden_nonlinearity=F.relu,
output_nonlinearity=None)
exploration_policy = AddGaussianNoise(env.spec,
policy,
total_timesteps=num_timesteps,
max_sigma=0.1,
min_sigma=0.1)
qf1 = ContinuousMLPQFunction(env_spec=env.spec,
hidden_sizes=[256, 256],
hidden_nonlinearity=F.relu)
qf2 = ContinuousMLPQFunction(env_spec=env.spec,
hidden_sizes=[256, 256],
hidden_nonlinearity=F.relu)
replay_buffer = PathBuffer(capacity_in_transitions=int(1e6))
td3 = TD3(env_spec=env.spec,
policy=policy,
qf1=qf1,
qf2=qf2,
replay_buffer=replay_buffer,
exploration_policy=exploration_policy,
steps_per_epoch=steps_per_epoch,
grad_steps_per_env_step=1,
num_evaluation_episodes=1,
discount=0.99)
prefer_gpu()
td3.to()
trainer.setup(td3, env, sampler_cls=LocalSampler)
trainer.train(n_epochs=n_epochs, batch_size=sampler_batch_size)
def ppo_garage_pytorch(ctxt, env_id, seed):
"""Create garage PyTorch PPO model and training.
Args:
ctxt (garage.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)
runner = LocalRunner(ctxt)
env = TfEnv(normalize(gym.make(env_id)))
policy = PyTorch_GMP(env.spec,
hidden_sizes=(32, 32),
hidden_nonlinearity=torch.tanh,
output_nonlinearity=None)
value_functions = LinearFeatureBaseline(env_spec=env.spec)
algo = PyTorch_PPO(env_spec=env.spec,
policy=policy,
value_function=value_functions,
optimizer=torch.optim.Adam,
policy_lr=3e-4,
max_path_length=hyper_parameters['max_path_length'],
discount=0.99,
gae_lambda=0.95,
center_adv=True,
lr_clip_range=0.2,
minibatch_size=128,
max_optimization_epochs=10)
runner.setup(algo, env)
runner.train(n_epochs=hyper_parameters['n_epochs'],
batch_size=hyper_parameters['batch_size'])
def test_state_dict_to():
"""Test state_dict_to"""
set_seed(42)
# Using tensor instead of Tensor so it can be declared on GPU
# pylint: disable=not-callable
expected = collections.OrderedDict([
('_module._layers.0.linear.weight',
tensor([[
0.13957974, -0.2693157, -0.19351028, 0.09471931, -0.43573233,
0.03590716, -0.4272097, -0.13935488, -0.35843086, -0.25814268,
0.03060348
],
[
0.20623916, -0.1914061, 0.46729338, -0.5437773,
-0.50449526, -0.55039907, 0.0141218, -0.02489783,
0.26499796, -0.03836302, 0.7235093
]],
device='cuda:0')),
('_module._layers.0.linear.bias', tensor([0., 0.], device='cuda:0')),
('_module._layers.1.linear.weight',
tensor([[-0.7181905, -0.6284401], [0.10591025, -0.14771031]],
device='cuda:0')),
('_module._layers.1.linear.bias', tensor([0., 0.], device='cuda:0')),
('_module._output_layers.0.linear.weight',
tensor([[-0.29133463, 0.58353233]], device='cuda:0')),
('_module._output_layers.0.linear.bias', tensor([0.], device='cuda:0'))
])
# pylint: enable=not-callable
env = normalize(GymEnv('InvertedDoublePendulum-v2'))
policy = DeterministicMLPPolicy(env_spec=env.spec,
hidden_sizes=[2, 2],
hidden_nonlinearity=F.relu,
output_nonlinearity=torch.tanh)
moved_state_dict = state_dict_to(policy.state_dict(), 'cuda')
assert np.all([
torch.allclose(expected[key], moved_state_dict[key])
for key in expected.keys()
])
assert np.all(
[moved_state_dict[key].is_cuda for key in moved_state_dict.keys()])
def trpo_cartpole_batch_sampler(ctxt=None,
seed=1,
batch_size=4000,
max_path_length=100):
"""Train TRPO with CartPole-v1 environment.
Args:
ctxt (garage.experiment.ExperimentContext): The experiment
configuration used by LocalRunner to create the snapshotter.
seed (int): Used to seed the random number generator to produce
determinism.
batch_size (int): Number of timesteps to use in each training step.
max_path_length (int): Number of timesteps to truncate paths to.
"""
set_seed(seed)
n_envs = batch_size // max_path_length
with LocalTFRunner(ctxt, max_cpus=n_envs) as runner:
env = GarageEnv(env_name='CartPole-v1')
policy = CategoricalMLPPolicy(name='policy',
env_spec=env.spec,
hidden_sizes=(32, 32))
baseline = LinearFeatureBaseline(env_spec=env.spec)
algo = TRPO(env_spec=env.spec,
policy=policy,
baseline=baseline,
max_path_length=max_path_length,
discount=0.99,
max_kl_step=0.01)
runner.setup(algo=algo,
env=env,
sampler_cls=BatchSampler,
sampler_args={'n_envs': n_envs})
runner.train(n_epochs=100, batch_size=4000, plot=False)
def trpo_cubecrash(ctxt=None, seed=1, max_episode_length=5, batch_size=4000):
"""Train TRPO with CubeCrash-v0 environment.
Args:
ctxt (garage.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_episode_length (int): Maximum length of a single episode.
batch_size (int): Number of timesteps to use in each training step.
"""
set_seed(seed)
with LocalTFRunner(ctxt) as runner:
env = normalize(
GymEnv('CubeCrash-v0', max_episode_length=max_episode_length))
policy = CategoricalCNNPolicy(env_spec=env.spec,
filters=((32, (8, 8)), (64, (4, 4))),
strides=(4, 2),
padding='VALID',
hidden_sizes=(32, 32))
baseline = GaussianCNNBaseline(env_spec=env.spec,
filters=((32, (8, 8)), (64, (4, 4))),
strides=(4, 2),
padding='VALID',
hidden_sizes=(32, 32),
use_trust_region=True)
algo = TRPO(env_spec=env.spec,
policy=policy,
baseline=baseline,
discount=0.99,
gae_lambda=0.95,
lr_clip_range=0.2,
policy_ent_coeff=0.0)
runner.setup(algo, env)
runner.train(n_epochs=100, batch_size=batch_size)
def test_trpo_gru_cartpole(self):
deterministic.set_seed(2)
with TFTrainer(snapshot_config, sess=self.sess) as trainer:
env = normalize(GymEnv('CartPole-v1', max_episode_length=100))
policy = CategoricalGRUPolicy(name='policy', env_spec=env.spec)
baseline = LinearFeatureBaseline(env_spec=env.spec)
algo = TRPO(env_spec=env.spec,
policy=policy,
baseline=baseline,
discount=0.99,
max_kl_step=0.01,
optimizer_args=dict(hvp_approach=FiniteDifferenceHVP(
base_eps=1e-5)))
trainer.setup(algo, env, sampler_cls=LocalSampler)
last_avg_ret = trainer.train(n_epochs=10, batch_size=2048)
assert last_avg_ret > 40
env.close()
def setup():
set_seed(24)
n_epochs = 11
steps_per_epoch = 10
sampler_batch_size = 512
num_timesteps = 100 * steps_per_epoch * sampler_batch_size
env = GymEnv('CartPole-v0')
replay_buffer = PathBuffer(capacity_in_transitions=int(1e6))
qf = DiscreteMLPQFunction(env_spec=env.spec, hidden_sizes=(8, 5))
policy = DiscreteQFArgmaxPolicy(env_spec=env.spec, qf=qf)
exploration_policy = EpsilonGreedyPolicy(env_spec=env.spec,
policy=policy,
total_timesteps=num_timesteps,
max_epsilon=1.0,
min_epsilon=0.01,
decay_ratio=0.4)
sampler = LocalSampler(agents=exploration_policy,
envs=env,
max_episode_length=env.spec.max_episode_length,
worker_class=FragmentWorker)
algo = DQN(env_spec=env.spec,
policy=policy,
qf=qf,
exploration_policy=exploration_policy,
replay_buffer=replay_buffer,
sampler=sampler,
steps_per_epoch=steps_per_epoch,
qf_lr=5e-5,
double_q=False,
discount=0.9,
min_buffer_size=int(1e4),
n_train_steps=500,
target_update_freq=30,
buffer_batch_size=64)
return algo, env, replay_buffer, n_epochs, sampler_batch_size
def test_meta_evaluator_n_traj():
set_seed(100)
tasks = SetTaskSampler(PointEnv)
max_path_length = 200
env = GarageEnv(PointEnv())
n_traj = 3
with tempfile.TemporaryDirectory() as log_dir_name:
runner = LocalRunner(
SnapshotConfig(snapshot_dir=log_dir_name,
snapshot_mode='last',
snapshot_gap=1))
algo = MockAlgo(env, max_path_length, n_traj)
runner.setup(algo, env)
meta_eval = MetaEvaluator(runner,
test_task_sampler=tasks,
max_path_length=max_path_length,
n_test_tasks=10,
n_exploration_traj=n_traj)
log_file = tempfile.NamedTemporaryFile()
csv_output = CsvOutput(log_file.name)
logger.add_output(csv_output)
meta_eval.evaluate(algo)
def test_dqn_cartpole_double_q(self):
"""Test DQN with CartPole environment."""
deterministic.set_seed(100)
with TFTrainer(snapshot_config, sess=self.sess) as trainer:
n_epochs = 10
steps_per_epoch = 10
sampler_batch_size = 500
num_timesteps = n_epochs * steps_per_epoch * sampler_batch_size
env = GymEnv('CartPole-v0')
replay_buffer = PathBuffer(capacity_in_transitions=int(1e4))
qf = DiscreteMLPQFunction(env_spec=env.spec, hidden_sizes=(64, 64))
policy = DiscreteQFArgmaxPolicy(env_spec=env.spec, qf=qf)
epilson_greedy_policy = EpsilonGreedyPolicy(
env_spec=env.spec,
policy=policy,
total_timesteps=num_timesteps,
max_epsilon=1.0,
min_epsilon=0.02,
decay_ratio=0.1)
algo = DQN(env_spec=env.spec,
policy=policy,
qf=qf,
exploration_policy=epilson_greedy_policy,
replay_buffer=replay_buffer,
qf_lr=1e-4,
discount=1.0,
min_buffer_size=int(1e3),
double_q=True,
n_train_steps=500,
steps_per_epoch=steps_per_epoch,
target_network_update_freq=1,
buffer_batch_size=32)
trainer.setup(algo, env)
last_avg_ret = trainer.train(n_epochs=n_epochs,
batch_size=sampler_batch_size)
assert last_avg_ret > 8.8
env.close()
def multi_env_ppo(ctxt=None, seed=1):
"""Train PPO on two Atari environments simultaneously.
Args:
ctxt (garage.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:
env1 = TfEnv(normalize(gym.make('Adventure-ram-v4')))
env2 = TfEnv(normalize(gym.make('Alien-ram-v4')))
env = MultiEnvWrapper([env1, env2])
policy = CategoricalMLPPolicy(
env_spec=env.spec,
hidden_nonlinearity=tf.nn.tanh,
)
baseline = LinearFeatureBaseline(env_spec=env.spec)
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)
runner.train(n_epochs=120, batch_size=2048, plot=False)
def run_garage_tf(env, seed, log_dir):
"""Create garage TensorFlow VPG model and training."""
deterministic.set_seed(seed)
with LocalTFRunner(snapshot_config) as runner:
env = TfEnv(normalize(env))
policy = TF_GMP(
env_spec=env.spec,
hidden_sizes=hyper_parameters['hidden_sizes'],
hidden_nonlinearity=tf.nn.tanh,
output_nonlinearity=None,
)
baseline = LinearFeatureBaseline(env_spec=env.spec)
algo = TF_VPG(env_spec=env.spec,
policy=policy,
baseline=baseline,
max_path_length=hyper_parameters['max_path_length'],
discount=hyper_parameters['discount'],
center_adv=hyper_parameters['center_adv'],
optimizer_args=dict(
tf_optimizer_args=dict(
learning_rate=hyper_parameters['learning_rate']),
verbose=True)) # yapf: disable
# 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(algo, env)
runner.train(n_epochs=hyper_parameters['n_epochs'], batch_size=2048)
dowel_logger.remove_all()
return tabular_log_file
def test_ppo_pendulum(self):
"""Test PPO with Pendulum environment."""
deterministic.set_seed(0)
episodes_per_task = 5
max_episode_length = self.env.spec.max_episode_length
task_sampler = SetTaskSampler(
HalfCheetahDirEnv,
wrapper=lambda env, _: normalize(GymEnv(
env, max_episode_length=max_episode_length),
expected_action_scale=10.))
meta_evaluator = MetaEvaluator(test_task_sampler=task_sampler,
n_test_tasks=1,
n_test_episodes=10)
sampler = LocalSampler(
agents=self.policy,
envs=self.env,
max_episode_length=self.env.spec.max_episode_length)
trainer = Trainer(snapshot_config)
algo = MAMLVPG(env=self.env,
policy=self.policy,
sampler=sampler,
task_sampler=self.task_sampler,
value_function=self.value_function,
meta_batch_size=5,
discount=0.99,
gae_lambda=1.,
inner_lr=0.1,
num_grad_updates=1,
meta_evaluator=meta_evaluator)
trainer.setup(algo, self.env)
last_avg_ret = trainer.train(n_epochs=10,
batch_size=episodes_per_task *
max_episode_length)
assert last_avg_ret > -5
def test_bc_point_deterministic(ray_local_session_fixture): # NOQA
del ray_local_session_fixture
assert ray.is_initialized()
deterministic.set_seed(100)
trainer = Trainer(snapshot_config)
goal = np.array([1., 1.])
env = PointEnv(goal=goal, max_episode_length=200)
expert = OptimalPolicy(env.spec, goal=goal)
policy = DeterministicMLPPolicy(env.spec, hidden_sizes=[8, 8])
batch_size = 600
sampler = LocalSampler(agents=expert,
envs=env,
max_episode_length=env.spec.max_episode_length)
algo = BC(env.spec,
policy,
batch_size=batch_size,
source=expert,
sampler=sampler,
policy_lr=1e-2,
loss='mse')
trainer.setup(algo, env)
run_bc(trainer, algo, batch_size)
