def sac_setup(env, trainer, args):
policy = TanhGaussianMLPPolicy(
env_spec=env.spec,
hidden_sizes=[args.hidden_dim] * args.depth,
hidden_nonlinearity=nn.ReLU,
output_nonlinearity=None,
min_std=np.exp(-20.),
max_std=np.exp(2.),
)
qf1 = ContinuousMLPQFunction(env_spec=env.spec,
hidden_sizes=[args.hidden_dim] * args.depth,
hidden_nonlinearity=F.relu)
qf2 = ContinuousMLPQFunction(env_spec=env.spec,
hidden_sizes=[args.hidden_dim] * args.depth,
hidden_nonlinearity=F.relu)
replay_buffer = PathBuffer(capacity_in_transitions=int(args.buffer_size))
sac = SAC(env_spec=env.spec,
policy=policy,
qf1=qf1,
qf2=qf2,
**convert_kwargs(args, SAC))
trainer.setup(algo=sac, env=env, sampler_cls=LocalSampler)
return sac
def sac_half_cheetah_batch(ctxt=None, seed=1):
"""Set up environment and algorithm and run the task.
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.
"""
deterministic.set_seed(seed)
trainer = Trainer(snapshot_config=ctxt)
env = normalize(GymEnv('HalfCheetah-v2'))
policy = TanhGaussianMLPPolicy(
env_spec=env.spec,
hidden_sizes=[256, 256],
hidden_nonlinearity=nn.ReLU,
output_nonlinearity=None,
min_std=np.exp(-20.),
max_std=np.exp(2.),
)
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))
sampler = LocalSampler(agents=policy,
envs=env,
max_episode_length=env.spec.max_episode_length,
worker_class=FragmentWorker)
sac = SAC(env_spec=env.spec,
policy=policy,
qf1=qf1,
qf2=qf2,
sampler=sampler,
gradient_steps_per_itr=1000,
max_episode_length_eval=1000,
replay_buffer=replay_buffer,
min_buffer_size=1e4,
target_update_tau=5e-3,
discount=0.99,
buffer_batch_size=256,
reward_scale=1.,
steps_per_epoch=1)
if torch.cuda.is_available():
set_gpu_mode(True)
else:
set_gpu_mode(False)
sac.to()
trainer.setup(algo=sac, env=env)
trainer.train(n_epochs=1000, batch_size=1000)
def test_fixed_alpha():
"""Test if using fixed_alpha ensures that alpha is non differentiable."""
env_names = ['InvertedDoublePendulum-v2', 'InvertedDoublePendulum-v2']
task_envs = [GymEnv(name, max_episode_length=100) for name in env_names]
env = MultiEnvWrapper(task_envs, sample_strategy=round_robin_strategy)
test_envs = MultiEnvWrapper(task_envs,
sample_strategy=round_robin_strategy)
deterministic.set_seed(0)
trainer = Trainer(snapshot_config=snapshot_config)
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), )
num_tasks = 2
buffer_batch_size = 128
sampler = LocalSampler(agents=policy,
envs=env,
max_episode_length=env.spec.max_episode_length,
worker_class=FragmentWorker)
mtsac = MTSAC(policy=policy,
qf1=qf1,
qf2=qf2,
sampler=sampler,
gradient_steps_per_itr=100,
eval_env=[test_envs],
env_spec=env.spec,
num_tasks=num_tasks,
steps_per_epoch=1,
replay_buffer=replay_buffer,
min_buffer_size=1e3,
target_update_tau=5e-3,
discount=0.99,
buffer_batch_size=buffer_batch_size,
fixed_alpha=np.exp(0.5))
if torch.cuda.is_available():
set_gpu_mode(True)
else:
set_gpu_mode(False)
mtsac.to()
assert torch.allclose(torch.Tensor([0.5] * num_tasks),
mtsac._log_alpha.to('cpu'))
trainer.setup(mtsac, env)
trainer.train(n_epochs=1, batch_size=128, plot=False)
assert torch.allclose(torch.Tensor([0.5] * num_tasks),
mtsac._log_alpha.to('cpu'))
assert not mtsac._use_automatic_entropy_tuning
def test_to():
"""Test the torch function that moves modules to GPU.
Test that the policy and qfunctions are moved to gpu if gpu is
available.
"""
env_names = ['CartPole-v0', 'CartPole-v1']
task_envs = [GarageEnv(env_name=name) for name in env_names]
env = MultiEnvWrapper(task_envs, sample_strategy=round_robin_strategy)
deterministic.set_seed(0)
policy = TanhGaussianMLPPolicy(
env_spec=env.spec,
hidden_sizes=[1, 1],
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=[1, 1],
hidden_nonlinearity=F.relu)
qf2 = ContinuousMLPQFunction(env_spec=env.spec,
hidden_sizes=[1, 1],
hidden_nonlinearity=F.relu)
replay_buffer = PathBuffer(capacity_in_transitions=int(1e6), )
num_tasks = 2
buffer_batch_size = 2
mtsac = MTSAC(policy=policy,
qf1=qf1,
qf2=qf2,
gradient_steps_per_itr=150,
max_path_length=150,
eval_env=env,
env_spec=env.spec,
num_tasks=num_tasks,
steps_per_epoch=5,
replay_buffer=replay_buffer,
min_buffer_size=1e3,
target_update_tau=5e-3,
discount=0.99,
buffer_batch_size=buffer_batch_size)
set_gpu_mode(torch.cuda.is_available())
mtsac.to()
device = global_device()
for param in mtsac._qf1.parameters():
assert param.device == device
for param in mtsac._qf2.parameters():
assert param.device == device
for param in mtsac._qf2.parameters():
assert param.device == device
for param in mtsac.policy.parameters():
assert param.device == device
assert mtsac._log_alpha.device == device
def torch_sac_half_cheetah(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.
"""
deterministic.set_seed(seed)
runner = LocalRunner(snapshot_config=ctxt)
env = GarageEnv(normalize(gym.make('HalfCheetah-v2')))
policy = TanhGaussianMLPPolicy(
env_spec=env.spec,
hidden_sizes=[256, 256],
hidden_nonlinearity=nn.ReLU,
output_nonlinearity=None,
min_std=np.exp(-20.),
max_std=np.exp(2.),
)
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 = SimpleReplayBuffer(env_spec=env.spec,
size_in_transitions=int(1e6),
time_horizon=1)
sac = SAC(env_spec=env.spec,
policy=policy,
qf1=qf1,
qf2=qf2,
gradient_steps_per_itr=1000,
max_path_length=500,
use_automatic_entropy_tuning=True,
replay_buffer=replay_buffer,
min_buffer_size=1e4,
target_update_tau=5e-3,
discount=0.99,
buffer_batch_size=256,
reward_scale=1.,
steps_per_epoch=1)
if torch.cuda.is_available():
tu.set_gpu_mode(True)
else:
tu.set_gpu_mode(False)
sac.to()
runner.setup(algo=sac, env=env, sampler_cls=LocalSampler)
runner.train(n_epochs=1000, batch_size=1000)
def test_sac_inverted_double_pendulum():
"""Test Sac performance on inverted pendulum."""
# 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), )
trainer = Trainer(snapshot_config=snapshot_config)
sampler = LocalSampler(agents=policy,
envs=env,
max_episode_length=env.spec.max_episode_length,
worker_class=FragmentWorker)
sac = SAC(env_spec=env.spec,
policy=policy,
qf1=qf1,
qf2=qf2,
sampler=sampler,
gradient_steps_per_itr=100,
replay_buffer=replay_buffer,
min_buffer_size=1e3,
target_update_tau=5e-3,
discount=0.99,
buffer_batch_size=64,
reward_scale=1.,
steps_per_epoch=2)
trainer.setup(sac, env)
if torch.cuda.is_available():
set_gpu_mode(True)
else:
set_gpu_mode(False)
sac.to()
ret = trainer.train(n_epochs=12, batch_size=200, plot=False)
# check that automatic entropy tuning is used
assert sac._use_automatic_entropy_tuning
# assert that there was a gradient properly connected to alpha
# this doesn't verify that the path from the temperature objective is
# correct.
assert not torch.allclose(torch.Tensor([1.]), sac._log_alpha.to('cpu'))
# check that policy is learning beyond predecided threshold
assert ret > 80
def test_sac_to():
"""Test moving Sac between CPU and GPU."""
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), )
trainer = Trainer(snapshot_config=snapshot_config)
sampler = LocalSampler(agents=policy,
envs=env,
max_episode_length=env.spec.max_episode_length,
worker_class=FragmentWorker)
sac = SAC(env_spec=env.spec,
policy=policy,
qf1=qf1,
qf2=qf2,
sampler=sampler,
gradient_steps_per_itr=100,
replay_buffer=replay_buffer,
min_buffer_size=1e3,
target_update_tau=5e-3,
discount=0.99,
buffer_batch_size=64,
reward_scale=1.,
steps_per_epoch=2)
trainer.setup(sac, env)
if torch.cuda.is_available():
set_gpu_mode(True)
else:
set_gpu_mode(False)
sac.to()
trainer.setup(algo=sac, env=env)
trainer.train(n_epochs=1, batch_size=100)
log_alpha = torch.clone(sac._log_alpha).cpu()
set_gpu_mode(False)
sac.to()
assert torch.allclose(log_alpha, sac._log_alpha)
def test_mtsac_get_log_alpha_incorrect_num_tasks(monkeypatch):
"""Check that if the num_tasks passed does not match the number of tasks
in the environment, then the algorithm should raise an exception.
MTSAC uses disentangled alphas, meaning that
"""
env_names = ['CartPole-v0', 'CartPole-v1']
task_envs = [GymEnv(name, max_episode_length=150) for name in env_names]
env = MultiEnvWrapper(task_envs, sample_strategy=round_robin_strategy)
deterministic.set_seed(0)
policy = TanhGaussianMLPPolicy(
env_spec=env.spec,
hidden_sizes=[1, 1],
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=[1, 1],
hidden_nonlinearity=F.relu)
qf2 = ContinuousMLPQFunction(env_spec=env.spec,
hidden_sizes=[1, 1],
hidden_nonlinearity=F.relu)
replay_buffer = PathBuffer(capacity_in_transitions=int(1e6), )
buffer_batch_size = 2
mtsac = MTSAC(policy=policy,
qf1=qf1,
qf2=qf2,
sampler=None,
gradient_steps_per_itr=150,
eval_env=[env],
env_spec=env.spec,
num_tasks=4,
steps_per_epoch=5,
replay_buffer=replay_buffer,
min_buffer_size=1e3,
target_update_tau=5e-3,
discount=0.99,
buffer_batch_size=buffer_batch_size)
monkeypatch.setattr(mtsac, '_log_alpha', torch.Tensor([1., 2.]))
error_string = ('The number of tasks in the environment does '
'not match self._num_tasks. Are you sure that you passed '
'The correct number of tasks?')
obs = torch.Tensor([env.reset()[0]] * buffer_batch_size)
with pytest.raises(ValueError, match=error_string):
mtsac._get_log_alpha(dict(observation=obs))
def test_mtsac_get_log_alpha(monkeypatch):
"""Check that the private function _get_log_alpha functions correctly.
MTSAC uses disentangled alphas, meaning that
"""
env_names = ['CartPole-v0', 'CartPole-v1']
task_envs = [GarageEnv(env_name=name) for name in env_names]
env = MultiEnvWrapper(task_envs, sample_strategy=round_robin_strategy)
deterministic.set_seed(0)
policy = TanhGaussianMLPPolicy(
env_spec=env.spec,
hidden_sizes=[1, 1],
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=[1, 1],
hidden_nonlinearity=F.relu)
qf2 = ContinuousMLPQFunction(env_spec=env.spec,
hidden_sizes=[1, 1],
hidden_nonlinearity=F.relu)
replay_buffer = PathBuffer(capacity_in_transitions=int(1e6), )
num_tasks = 2
buffer_batch_size = 2
mtsac = MTSAC(policy=policy,
qf1=qf1,
qf2=qf2,
gradient_steps_per_itr=150,
max_path_length=150,
eval_env=env,
env_spec=env.spec,
num_tasks=num_tasks,
steps_per_epoch=5,
replay_buffer=replay_buffer,
min_buffer_size=1e3,
target_update_tau=5e-3,
discount=0.99,
buffer_batch_size=buffer_batch_size)
monkeypatch.setattr(mtsac, '_log_alpha', torch.Tensor([1., 2.]))
for i, _ in enumerate(env_names):
obs = torch.Tensor([env.reset()] * buffer_batch_size)
log_alpha = mtsac._get_log_alpha(dict(observation=obs))
assert (log_alpha == torch.Tensor([i + 1, i + 1])).all().item()
assert log_alpha.size() == torch.Size([mtsac._buffer_batch_size])
def test_mtsac_inverted_double_pendulum():
"""Performance regression test of MTSAC on 2 InvDoublePendulum envs."""
env_names = ['InvertedDoublePendulum-v2', 'InvertedDoublePendulum-v2']
task_envs = [GymEnv(name, max_episode_length=100) for name in env_names]
env = MultiEnvWrapper(task_envs, sample_strategy=round_robin_strategy)
test_envs = MultiEnvWrapper(task_envs,
sample_strategy=round_robin_strategy)
deterministic.set_seed(0)
trainer = Trainer(snapshot_config=snapshot_config)
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), )
num_tasks = 2
buffer_batch_size = 128
sampler = LocalSampler(agents=policy,
envs=env,
max_episode_length=env.spec.max_episode_length,
worker_class=FragmentWorker)
mtsac = MTSAC(policy=policy,
qf1=qf1,
qf2=qf2,
sampler=sampler,
gradient_steps_per_itr=100,
eval_env=[test_envs],
env_spec=env.spec,
num_tasks=num_tasks,
steps_per_epoch=5,
replay_buffer=replay_buffer,
min_buffer_size=1e3,
target_update_tau=5e-3,
discount=0.99,
buffer_batch_size=buffer_batch_size)
trainer.setup(mtsac, env)
ret = trainer.train(n_epochs=8, batch_size=128, plot=False)
assert ret > 0
def create_qf_net(env_spec, net_params):
net_type = net_params["net_type"]
assert net_type in {"MLP", "Dendrite_MLP"}
if net_type == "MLP":
net = ContinuousMLPQFunction(
env_spec=env_spec,
hidden_sizes=net_params["qf_hidden_sizes"],
hidden_nonlinearity=create_nonlinearity(net_params["qf_hidden_nonlinearity"]),
output_nonlinearity=create_nonlinearity(net_params["qf_output_nonlinearity"]),
)
elif net_type == "Dendrite_MLP":
dendritic_layer_class = create_dendritic_layer(net_params["dendritic_layer_class"])
net = ContinuousDendriteMLPQFunction(
env_spec=env_spec,
hidden_sizes=net_params["hidden_sizes"],
num_segments=net_params["num_segments"],
dim_context=net_params["dim_context"],
kw=net_params["kw"],
kw_percent_on=net_params["kw_percent_on"],
context_percent_on=net_params["context_percent_on"],
weight_sparsity=net_params["weight_sparsity"],
weight_init=net_params["weight_init"],
dendrite_init=net_params["dendrite_init"],
dendritic_layer_class=dendritic_layer_class,
output_nonlinearity=net_params["output_nonlinearity"],
preprocess_module_type=net_params["preprocess_module_type"],
preprocess_output_dim=net_params["preprocess_output_dim"],
representation_module_type=net_params["representation_module_type"],
representation_module_dims=net_params["representation_module_dims"],
)
else:
raise NotImplementedError
return net
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))
sampler = LocalSampler(agents=exploration_policy,
envs=env,
max_episode_length=env.spec.max_episode_length,
worker_class=FragmentWorker)
td3 = TD3(env_spec=env.spec,
policy=policy,
qf1=qf1,
qf2=qf2,
replay_buffer=replay_buffer,
sampler=sampler,
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 test_sac_inverted_pendulum():
"""Test Sac performance on inverted pendulum."""
# pylint: disable=unexpected-keyword-arg
env = GarageEnv(normalize(gym.make('InvertedDoublePendulum-v2')))
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 = SimpleReplayBuffer(env_spec=env.spec,
size_in_transitions=int(1e6),
time_horizon=1)
runner = LocalRunner(snapshot_config=snapshot_config)
sac = SAC(env_spec=env.spec,
policy=policy,
qf1=qf1,
qf2=qf2,
gradient_steps_per_itr=100,
max_path_length=100,
use_automatic_entropy_tuning=True,
replay_buffer=replay_buffer,
min_buffer_size=1e3,
target_update_tau=5e-3,
discount=0.99,
buffer_batch_size=64,
reward_scale=1.,
steps_per_epoch=2)
runner.setup(sac, env, sampler_cls=LocalSampler)
if torch.cuda.is_available():
tu.set_gpu_mode(True)
else:
tu.set_gpu_mode(False)
sac.to()
ret = runner.train(n_epochs=12, batch_size=200, plot=False)
assert ret > 85
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), )
trainer = Trainer(snapshot_config=snapshot_config)
sampler = LocalSampler(agents=policy,
envs=env,
max_episode_length=env.spec.max_episode_length,
worker_class=FragmentWorker)
sac = SAC(env_spec=env.spec,
policy=policy,
qf1=qf1,
qf2=qf2,
sampler=sampler,
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))
trainer.setup(sac, env)
sac.to()
trainer.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 load_pearl(env_name="CartPole-v0"):
"""Return an instance of the PEARL algorithm.
NOTE: currently not working.
"""
num_train_tasks = 100
num_test_tasks = 30
latent_size = 5
net_size = 300
encoder_hidden_size = 200
encoder_hidden_sizes = (encoder_hidden_size, encoder_hidden_size,
encoder_hidden_size)
# Create multi-task environment and sample tasks.
env_start = GarageEnv(env_name=env_name)
env_sampler = SetTaskSampler(lambda: GarageEnv(normalize(env_start)))
env = env_sampler.sample(num_train_tasks)
test_env_sampler = SetTaskSampler(lambda: GarageEnv(normalize(env_start)))
# Instantiate networks.
augmented_env = PEARL.augment_env_spec(env[0](), latent_size)
qf = ContinuousMLPQFunction(env_spec=augmented_env,
hidden_sizes=[net_size, net_size, net_size])
vf_env = PEARL.get_env_spec(env[0](), latent_size, 'vf')
vf = ContinuousMLPQFunction(env_spec=vf_env,
hidden_sizes=[net_size, net_size, net_size])
inner_policy = TanhGaussianMLPPolicy(
env_spec=augmented_env, hidden_sizes=[net_size, net_size, net_size])
pearl = PEARL(env=env,
inner_policy=inner_policy,
qf=qf,
vf=vf,
num_train_tasks=num_train_tasks,
num_test_tasks=num_test_tasks,
latent_dim=latent_size,
encoder_hidden_sizes=encoder_hidden_sizes,
test_env_sampler=test_env_sampler)
return pearl
def test_pickling(self):
"""Test pickle and unpickle."""
net_size = 10
env_sampler = SetTaskSampler(PointEnv)
env = env_sampler.sample(5)
test_env_sampler = SetTaskSampler(PointEnv)
augmented_env = PEARL.augment_env_spec(env[0](), 5)
qf = ContinuousMLPQFunction(
env_spec=augmented_env,
hidden_sizes=[net_size, net_size, net_size])
vf_env = PEARL.get_env_spec(env[0](), 5, 'vf')
vf = ContinuousMLPQFunction(
env_spec=vf_env, hidden_sizes=[net_size, net_size, net_size])
inner_policy = TanhGaussianMLPPolicy(
env_spec=augmented_env,
hidden_sizes=[net_size, net_size, net_size])
pearl = PEARL(env=env,
inner_policy=inner_policy,
qf=qf,
vf=vf,
num_train_tasks=5,
num_test_tasks=5,
latent_dim=5,
encoder_hidden_sizes=[10, 10],
test_env_sampler=test_env_sampler)
# This line is just to improve coverage
pearl.to()
pickled = pickle.dumps(pearl)
unpickled = pickle.loads(pickled)
assert hasattr(unpickled, '_replay_buffers')
assert hasattr(unpickled, '_context_replay_buffers')
assert unpickled._is_resuming
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 ddpg_pendulum(ctxt=None, seed=1, lr=1e-4):
"""Train DDPG with InvertedDoublePendulum-v2 environment.
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.
lr (float): Learning rate for policy optimization.
"""
set_seed(seed)
trainer = Trainer(ctxt)
env = normalize(GymEnv('InvertedDoublePendulum-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 = PathBuffer(capacity_in_transitions=int(1e6))
policy_optimizer = (torch.optim.Adagrad, {'lr': lr, 'lr_decay': 0.99})
sampler = LocalSampler(agents=exploration_policy,
envs=env,
max_episode_length=env.spec.max_episode_length,
worker_class=FragmentWorker)
ddpg = 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,
policy_optimizer=policy_optimizer,
qf_optimizer=torch.optim.Adam)
trainer.setup(algo=ddpg, env=env)
trainer.train(n_epochs=500, batch_size=100)
def load_sac(env_name="MountainCarContinuous-v0"):
"""Return an instance of the SAC algorithm."""
env = GarageEnv(env_name=env_name)
policy = DeterministicMLPPolicy(name='policy',
env_spec=env.spec,
hidden_sizes=[64, 64])
qf1 = ContinuousMLPQFunction(env_spec=env.spec,
hidden_sizes=[64, 64],
hidden_nonlinearity=F.relu)
qf2 = ContinuousMLPQFunction(env_spec=env.spec,
hidden_sizes=[64, 64],
hidden_nonlinearity=F.relu)
replay_buffer = PathBuffer(capacity_in_transitions=int(1e6))
algo = SAC(env_spec=env.spec,
policy=policy,
qf1=qf1,
qf2=qf2,
gradient_steps_per_itr=1000,
max_path_length=500,
replay_buffer=replay_buffer)
return algo
def test_output_shape(self, batch_size, hidden_sizes):
env_spec = GymEnv(DummyBoxEnv()).spec
obs_dim = env_spec.observation_space.flat_dim
act_dim = env_spec.action_space.flat_dim
obs = torch.ones(batch_size, obs_dim, dtype=torch.float32)
act = torch.ones(batch_size, act_dim, dtype=torch.float32)
qf = ContinuousMLPQFunction(env_spec=env_spec,
hidden_nonlinearity=None,
hidden_sizes=hidden_sizes,
hidden_w_init=nn.init.ones_,
output_w_init=nn.init.ones_)
output = qf(obs, act)
assert output.shape == (batch_size, 1)
def ddpg_pendulum(ctxt=None, seed=1, lr=1e-4):
"""Train DDPG with InvertedDoublePendulum-v2 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.
lr (float): Learning rate for policy optimization.
"""
set_seed(seed)
runner = LocalRunner(ctxt)
env = GarageEnv(normalize(gym.make('InvertedDoublePendulum-v2')))
action_noise = OUStrategy(env.spec, sigma=0.2)
policy = DeterministicMLPPolicy(env_spec=env.spec,
hidden_sizes=[64, 64],
hidden_nonlinearity=F.relu,
output_nonlinearity=torch.tanh)
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)
policy_optimizer = (torch.optim.Adagrad, {'lr': lr, 'lr_decay': 0.99})
ddpg = 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_strategy=action_noise,
target_update_tau=1e-2,
discount=0.9,
policy_optimizer=policy_optimizer,
qf_optimizer=torch.optim.Adam)
runner.setup(algo=ddpg, env=env)
runner.train(n_epochs=500, batch_size=100)
def test_ddpg_pendulum(self):
"""Test DDPG with Pendulum environment.
This environment has a [-3, 3] action_space bound.
"""
deterministic.set_seed(0)
trainer = Trainer(snapshot_config)
env = normalize(GymEnv('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 = 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 > 10
env.close()
def run_task(snapshot_config, *_):
"""Set up environment and algorithm and run the task.
Args:
snapshot_config (garage.experiment.SnapshotConfig): The snapshot
configuration used by LocalRunner to create the snapshotter.
If None, it will create one with default settings.
_ : Unused parameters
"""
runner = LocalRunner(snapshot_config)
env = GarageEnv(normalize(gym.make('InvertedDoublePendulum-v2')))
action_noise = OUStrategy(env.spec, sigma=0.2)
policy = DeterministicMLPPolicy(env_spec=env.spec,
hidden_sizes=[64, 64],
hidden_nonlinearity=F.relu,
output_nonlinearity=torch.tanh)
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)
policy_optimizer = (torch.optim.Adagrad, {'lr': 1e-4, 'lr_decay': 0.99})
ddpg = 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_strategy=action_noise,
target_update_tau=1e-2,
discount=0.9,
policy_optimizer=policy_optimizer,
qf_optimizer=torch.optim.Adam)
runner.setup(algo=ddpg, env=env)
runner.train(n_epochs=500, batch_size=100)
def load_ddpg(env_name="MountainCarContinuous-v0"):
"""Return an instance of the DDPG algorithm.
Note: does this only work with continous?
"""
env = GarageEnv(env_name=env_name)
policy = DeterministicMLPPolicy(name='policy',
env_spec=env.spec,
hidden_sizes=[64, 64])
qf = ContinuousMLPQFunction(env_spec=env.spec,
hidden_sizes=[64, 64],
hidden_nonlinearity=F.relu)
replay_buffer = PathBuffer(capacity_in_transitions=int(1e6))
algo = DDPG(env_spec=env.spec,
policy=policy,
qf=qf,
replay_buffer=replay_buffer)
return algo
def test_forward(self, hidden_sizes):
env_spec = GymEnv(DummyBoxEnv()).spec
obs_dim = env_spec.observation_space.flat_dim
act_dim = env_spec.action_space.flat_dim
obs = torch.ones(obs_dim, dtype=torch.float32).unsqueeze(0)
act = torch.ones(act_dim, dtype=torch.float32).unsqueeze(0)
qf = ContinuousMLPQFunction(env_spec=env_spec,
hidden_nonlinearity=None,
hidden_sizes=hidden_sizes,
hidden_w_init=nn.init.ones_,
output_w_init=nn.init.ones_)
output = qf(obs, act)
expected_output = torch.full([1, 1],
fill_value=(obs_dim + act_dim) *
np.prod(hidden_sizes),
dtype=torch.float32)
assert torch.eq(output, expected_output)
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 test_is_pickleable(self, hidden_sizes):
env_spec = GymEnv(DummyBoxEnv()).spec
obs_dim = env_spec.observation_space.flat_dim
act_dim = env_spec.action_space.flat_dim
obs = torch.ones(obs_dim, dtype=torch.float32).unsqueeze(0)
act = torch.ones(act_dim, dtype=torch.float32).unsqueeze(0)
qf = ContinuousMLPQFunction(env_spec=env_spec,
hidden_nonlinearity=None,
hidden_sizes=hidden_sizes,
hidden_w_init=nn.init.ones_,
output_w_init=nn.init.ones_)
output1 = qf(obs, act)
p = pickle.dumps(qf)
qf_pickled = pickle.loads(p)
output2 = qf_pickled(obs, act)
assert torch.eq(output1, output2)
def test_ddpg_double_pendulum(self):
"""Test DDPG with Pendulum environment."""
runner = LocalRunner()
env = GarageEnv(gym.make('InvertedDoublePendulum-v2'))
action_noise = OUStrategy(env.spec, sigma=0.2)
policy = DeterministicMLPPolicy(env_spec=env.spec,
hidden_sizes=[64, 64],
hidden_nonlinearity=F.relu,
output_nonlinearity=torch.tanh)
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,
n_train_steps=50,
min_buffer_size=int(1e4),
exploration_strategy=action_noise,
target_update_tau=1e-2,
policy_lr=1e-4,
qf_lr=1e-3,
discount=0.9)
runner.setup(algo, env)
last_avg_ret = runner.train(n_epochs=10,
n_epoch_cycles=20,
batch_size=100)
assert last_avg_ret > 60
env.close()
def create_qf_net(env_spec, net_params):
if net_params.net_type == "MLP":
net = ContinuousMLPQFunction(
env_spec=env_spec,
hidden_sizes=net_params.hidden_sizes,
hidden_nonlinearity=create_nonlinearity(
net_params.qf_hidden_nonlinearity),
output_nonlinearity=create_nonlinearity(
net_params.output_nonlinearity),
)
elif net_params.net_type == "Dendrite_MLP":
dendritic_layer_class = create_dendritic_layer(
net_params.dendritic_layer_class)
net = ContinuousDendriteMLPQFunction(
env_spec=env_spec,
num_tasks=net_params.num_tasks,
input_data=net_params.input_data,
context_data=net_params.context_data,
hidden_sizes=net_params.hidden_sizes,
layers_modulated=net_params.layers_modulated,
num_segments=net_params.num_segments,
kw_percent_on=net_params.kw_percent_on,
context_percent_on=net_params.context_percent_on,
weight_sparsity=net_params.weight_sparsity,
weight_init=net_params.weight_init,
dendrite_weight_sparsity=net_params.dendrite_weight_sparsity,
dendrite_init=net_params.dendrite_init,
dendritic_layer_class=dendritic_layer_class,
output_nonlinearity=net_params.output_nonlinearity,
preprocess_module_type=net_params.preprocess_module_type,
preprocess_output_dim=net_params.preprocess_output_dim,
)
else:
raise NotImplementedError
return net
def test_ddpg_double_pendulum(self):
"""Test DDPG with Pendulum environment."""
deterministic.set_seed(0)
runner = LocalRunner(snapshot_config)
env = GarageEnv(gym.make('InvertedDoublePendulum-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 = PathBuffer(capacity_in_transitions=int(1e6))
algo = DDPG(env_spec=env.spec,
policy=policy,
qf=qf,
replay_buffer=replay_buffer,
max_path_length=100,
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 > 45
env.close()
