def test_auxiliary(self):
obs_dim, action_dim, task_num, latent_dim = (2, ), (2, ), 2, 2
env = GarageEnv(DummyBoxEnv(obs_dim=obs_dim, action_dim=action_dim))
embedding_spec = InOutSpec(
input_space=akro.Box(low=np.zeros(task_num),
high=np.ones(task_num)),
output_space=akro.Box(low=np.zeros(latent_dim),
high=np.ones(latent_dim)))
encoder = GaussianMLPEncoder(embedding_spec)
policy = GaussianMLPTaskEmbeddingPolicy(env_spec=env.spec,
encoder=encoder)
obs_input = tf.compat.v1.placeholder(tf.float32, shape=(None, None, 2))
task_input = tf.compat.v1.placeholder(tf.float32,
shape=(None, None, 2))
policy.build(obs_input, task_input)
assert policy.distribution.loc.get_shape().as_list(
)[-1] == env.action_space.flat_dim
assert policy.encoder == encoder
assert policy.latent_space.flat_dim == latent_dim
assert policy.task_space.flat_dim == task_num
assert (policy.augmented_observation_space.flat_dim ==
env.observation_space.flat_dim + task_num)
assert policy.encoder_distribution.loc.get_shape().as_list(
)[-1] == latent_dim
def test_get_action(self, obs_dim, task_num, latent_dim, action_dim):
env = GarageEnv(DummyBoxEnv(obs_dim=obs_dim, action_dim=action_dim))
embedding_spec = InOutSpec(
input_space=akro.Box(low=np.zeros(task_num),
high=np.ones(task_num)),
output_space=akro.Box(low=np.zeros(latent_dim),
high=np.ones(latent_dim)))
encoder = GaussianMLPEncoder(embedding_spec)
policy = GaussianMLPTaskEmbeddingPolicy(env_spec=env.spec,
encoder=encoder)
env.reset()
obs, _, _, _ = env.step(1)
latent = np.random.random((latent_dim, ))
task = np.zeros(task_num)
task[0] = 1
action1, _ = policy.get_action_given_latent(obs, latent)
action2, _ = policy.get_action_given_task(obs, task)
action3, _ = policy.get_action(np.concatenate([obs.flatten(), task]))
assert env.action_space.contains(action1)
assert env.action_space.contains(action2)
assert env.action_space.contains(action3)
obses, latents, tasks = [obs] * 3, [latent] * 3, [task] * 3
aug_obses = [np.concatenate([obs.flatten(), task])] * 3
action1n, _ = policy.get_actions_given_latents(obses, latents)
action2n, _ = policy.get_actions_given_tasks(obses, tasks)
action3n, _ = policy.get_actions(aug_obses)
for action in chain(action1n, action2n, action3n):
assert env.action_space.contains(action)
def test_get_latent(self):
obs_dim, action_dim, task_num, latent_dim = (2, ), (2, ), 5, 2
env = GarageEnv(DummyBoxEnv(obs_dim=obs_dim, action_dim=action_dim))
embedding_spec = InOutSpec(
input_space=akro.Box(low=np.zeros(task_num),
high=np.ones(task_num)),
output_space=akro.Box(low=np.zeros(latent_dim),
high=np.ones(latent_dim)))
encoder = GaussianMLPEncoder(embedding_spec)
policy = GaussianMLPTaskEmbeddingPolicy(env_spec=env.spec,
encoder=encoder)
task_id = 3
task_onehot = np.zeros(task_num)
task_onehot[task_id] = 1
latent, latent_info = policy.get_latent(task_onehot)
assert latent.shape == (latent_dim, )
assert latent_info['mean'].shape == (latent_dim, )
assert latent_info['log_std'].shape == (latent_dim, )
def test_get_vars(self):
obs_dim, action_dim, task_num, latent_dim = (2, ), (2, ), 5, 2
env = GarageEnv(DummyBoxEnv(obs_dim=obs_dim, action_dim=action_dim))
embedding_spec = InOutSpec(
input_space=akro.Box(low=np.zeros(task_num),
high=np.ones(task_num)),
output_space=akro.Box(low=np.zeros(latent_dim),
high=np.ones(latent_dim)))
encoder = GaussianMLPEncoder(embedding_spec, hidden_sizes=[32, 32, 32])
policy = GaussianMLPTaskEmbeddingPolicy(env_spec=env.spec,
encoder=encoder,
hidden_sizes=[32, 32, 32])
vars1 = sorted(policy.get_trainable_vars(), key=lambda v: v.name)
vars2 = sorted(policy.get_global_vars(), key=lambda v: v.name)
assert vars1 == vars2
# Two network. Each with 4 layers * (1 weight + 1 bias) + 1 log_std
assert len(vars1) == 2 * (4 * 2 + 1)
obs = np.random.random(obs_dim)
latent = np.random.random((latent_dim, ))
for var in vars1:
var.assign(np.ones(var.shape))
assert np.any(policy.get_action_given_latent(obs, latent) != 0)
for var in vars1:
var.assign(np.zeros(var.shape))
assert not np.all(policy.get_action_given_latent(obs, latent) == 0)
def test_encoder_dist_info(self):
obs_dim, action_dim, task_num, latent_dim = (2, ), (2, ), 5, 2
env = TfEnv(DummyBoxEnv(obs_dim=obs_dim, action_dim=action_dim))
with mock.patch(
'garage.tf.embeddings.'
'gaussian_mlp_encoder.GaussianMLPModel',
new=SimpleGaussianMLPModel):
old_build = SimpleGaussianMLPModel._build
def float32_build(this, obs_input, name):
mean, log_std, std, dist = old_build(this, obs_input, name)
return mean, tf.cast(log_std, tf.float32), std, dist
SimpleGaussianMLPModel._build = float32_build
embedding_spec = InOutSpec(
input_space=akro.Box(low=np.zeros(task_num),
high=np.ones(task_num)),
output_space=akro.Box(low=np.zeros(latent_dim),
high=np.ones(latent_dim)))
encoder = GaussianMLPEncoder(embedding_spec)
policy = GaussianMLPTaskEmbeddingPolicy(env_spec=env.spec,
encoder=encoder)
assert policy.encoder_distribution.dim == latent_dim
inp_ph = tf.compat.v1.placeholder(tf.float32, shape=(None, 5))
dist_sym = policy.encoder_dist_info_sym(inp_ph)
dist = self.sess.run(dist_sym,
feed_dict={inp_ph: [np.random.random(5)]})
expected_mean = np.full(latent_dim, 0.5)
expected_log_std = np.full(latent_dim, np.log(0.5))
assert np.allclose(dist['mean'], expected_mean)
assert np.allclose(dist['log_std'], expected_log_std)
SimpleGaussianMLPModel._dtype = np.float32
def test_get_latent(self):
obs_dim, action_dim, task_num, latent_dim = (2, ), (2, ), 5, 2
env = TfEnv(DummyBoxEnv(obs_dim=obs_dim, action_dim=action_dim))
with mock.patch(
'garage.tf.policies.'
'gaussian_mlp_task_embedding_policy.GaussianMLPModel',
new=SimpleGaussianMLPModel):
embedding_spec = InOutSpec(
input_space=akro.Box(low=np.zeros(task_num),
high=np.ones(task_num)),
output_space=akro.Box(low=np.zeros(latent_dim),
high=np.ones(latent_dim)))
encoder = GaussianMLPEncoder(embedding_spec)
policy = GaussianMLPTaskEmbeddingPolicy(env_spec=env.spec,
encoder=encoder)
task_id = 3
task_onehot = np.zeros(task_num)
task_onehot[task_id] = 1
latent, latent_info = policy.get_latent(task_onehot)
assert latent.shape == (latent_dim, )
assert latent_info['mean'].shape == (latent_dim, )
assert latent_info['log_std'].shape == (latent_dim, )
def test_pickling(self):
obs_dim, action_dim, task_num, latent_dim = (2, ), (2, ), 5, 2
env = GarageEnv(DummyBoxEnv(obs_dim=obs_dim, action_dim=action_dim))
embedding_spec = InOutSpec(
input_space=akro.Box(low=np.zeros(task_num),
high=np.ones(task_num)),
output_space=akro.Box(low=np.zeros(latent_dim),
high=np.ones(latent_dim)))
encoder = GaussianMLPEncoder(embedding_spec)
policy = GaussianMLPTaskEmbeddingPolicy(env_spec=env.spec,
encoder=encoder)
pickled = pickle.dumps(policy)
with tf.compat.v1.variable_scope('resumed'):
unpickled = pickle.loads(pickled)
assert hasattr(unpickled, '_f_dist_obs_latent')
assert hasattr(unpickled, '_f_dist_obs_task')
def test_auxiliary(self):
obs_dim, action_dim, task_num, latent_dim = (2, ), (2, ), 2, 2
env = TfEnv(DummyBoxEnv(obs_dim=obs_dim, action_dim=action_dim))
with mock.patch(
'garage.tf.policies.'
'gaussian_mlp_task_embedding_policy.GaussianMLPModel',
new=SimpleGaussianMLPModel):
embedding_spec = InOutSpec(
input_space=akro.Box(low=np.zeros(task_num),
high=np.ones(task_num)),
output_space=akro.Box(low=np.zeros(latent_dim),
high=np.ones(latent_dim)))
encoder = GaussianMLPEncoder(embedding_spec)
policy = GaussianMLPTaskEmbeddingPolicy(env_spec=env.spec,
encoder=encoder)
assert policy.distribution.dim == env.action_space.flat_dim
assert policy.encoder == encoder
assert policy.latent_space.flat_dim == latent_dim
assert policy.task_space.flat_dim == task_num
assert (policy.augmented_observation_space.flat_dim ==
env.observation_space.flat_dim + task_num)
assert policy.encoder_distribution.dim == latent_dim
def te_ppo_mt1_push(ctxt, seed, n_epochs, batch_size_per_task):
"""Train Task Embedding PPO with PointEnv.
Args:
ctxt (ExperimentContext): The experiment configuration used by
:class:`~Trainer` to create the :class:`~Snapshotter`.
seed (int): Used to seed the random number generator to produce
determinism.
n_epochs (int): Total number of epochs for training.
batch_size_per_task (int): Batch size of samples for each task.
"""
set_seed(seed)
n_tasks = 50
mt1 = metaworld.MT1('push-v1')
task_sampler = MetaWorldTaskSampler(mt1,
'train',
lambda env, _: normalize(env),
add_env_onehot=False)
envs = [env_up() for env_up in task_sampler.sample(n_tasks)]
env = MultiEnvWrapper(envs,
sample_strategy=round_robin_strategy,
mode='vanilla')
latent_length = 2
inference_window = 6
batch_size = batch_size_per_task * n_tasks
policy_ent_coeff = 2e-2
encoder_ent_coeff = 2e-4
inference_ce_coeff = 5e-2
embedding_init_std = 0.1
embedding_max_std = 0.2
embedding_min_std = 1e-6
policy_init_std = 1.0
policy_max_std = None
policy_min_std = None
with TFTrainer(snapshot_config=ctxt) as trainer:
task_embed_spec = TEPPO.get_encoder_spec(env.task_space,
latent_dim=latent_length)
task_encoder = GaussianMLPEncoder(
name='embedding',
embedding_spec=task_embed_spec,
hidden_sizes=(20, 20),
std_share_network=True,
init_std=embedding_init_std,
max_std=embedding_max_std,
output_nonlinearity=tf.nn.tanh,
min_std=embedding_min_std,
)
traj_embed_spec = TEPPO.get_infer_spec(
env.spec,
latent_dim=latent_length,
inference_window_size=inference_window)
inference = GaussianMLPEncoder(
name='inference',
embedding_spec=traj_embed_spec,
hidden_sizes=(20, 10),
std_share_network=True,
init_std=2.0,
output_nonlinearity=tf.nn.tanh,
min_std=embedding_min_std,
)
policy = GaussianMLPTaskEmbeddingPolicy(
name='policy',
env_spec=env.spec,
encoder=task_encoder,
hidden_sizes=(32, 16),
std_share_network=True,
max_std=policy_max_std,
init_std=policy_init_std,
min_std=policy_min_std,
)
baseline = LinearMultiFeatureBaseline(
env_spec=env.spec, features=['observations', 'tasks', 'latents'])
algo = TEPPO(env_spec=env.spec,
policy=policy,
baseline=baseline,
inference=inference,
discount=0.99,
lr_clip_range=0.2,
policy_ent_coeff=policy_ent_coeff,
encoder_ent_coeff=encoder_ent_coeff,
inference_ce_coeff=inference_ce_coeff,
use_softplus_entropy=True,
optimizer_args=dict(
batch_size=32,
max_optimization_epochs=10,
learning_rate=1e-3,
),
inference_optimizer_args=dict(
batch_size=32,
max_optimization_epochs=10,
),
center_adv=True,
stop_ce_gradient=True)
trainer.setup(algo,
env,
sampler_cls=LocalSampler,
sampler_args=None,
worker_class=TaskEmbeddingWorker)
trainer.train(n_epochs=n_epochs, batch_size=batch_size, plot=False)
def setup_method(self):
super().setup_method()
def circle(r, n):
"""Generate n points on a circle of radius r.
Args:
r (float): Radius of the circle.
n (int): Number of points to generate.
Yields:
tuple(float, float): Coordinate of a point.
"""
for t in np.arange(0, 2 * np.pi, 2 * np.pi / n):
yield r * np.sin(t), r * np.cos(t)
N = 4
goals = circle(3.0, N)
tasks = {
str(i + 1): {
'args': [],
'kwargs': {
'goal': g,
'never_done': False,
'done_bonus': 0.0,
}
}
for i, g in enumerate(goals)
}
latent_length = 1
inference_window = 2
self.batch_size = 100 * len(tasks)
self.policy_ent_coeff = 2e-2
self.encoder_ent_coeff = 2.2e-3
self.inference_ce_coeff = 5e-2
self.max_path_length = 100
embedding_init_std = 1.0
embedding_max_std = 2.0
embedding_min_std = 0.38
policy_init_std = 1.0
policy_max_std = None
policy_min_std = None
task_names = sorted(tasks.keys())
task_args = [tasks[t]['args'] for t in task_names]
task_kwargs = [tasks[t]['kwargs'] for t in task_names]
task_envs = [
GarageEnv(PointEnv(*t_args, **t_kwargs))
for t_args, t_kwargs in zip(task_args, task_kwargs)
]
self.env = env = MultiEnvWrapper(task_envs,
round_robin_strategy,
mode='vanilla')
latent_lb = np.zeros(latent_length, )
latent_ub = np.ones(latent_length, )
latent_space = akro.Box(latent_lb, latent_ub)
obs_lb, obs_ub = env.observation_space.bounds
obs_lb_flat = env.observation_space.flatten(obs_lb)
obs_ub_flat = env.observation_space.flatten(obs_ub)
traj_lb = np.stack([obs_lb_flat] * inference_window)
traj_ub = np.stack([obs_ub_flat] * inference_window)
traj_space = akro.Box(traj_lb, traj_ub)
task_embed_spec = InOutSpec(env.task_space, latent_space)
traj_embed_spec = InOutSpec(traj_space, latent_space)
self.inference = GaussianMLPEncoder(
name='inference',
embedding_spec=traj_embed_spec,
hidden_sizes=[20, 10],
std_share_network=True,
init_std=2.0,
output_nonlinearity=tf.nn.tanh,
min_std=embedding_min_std,
)
task_encoder = GaussianMLPEncoder(
name='embedding',
embedding_spec=task_embed_spec,
hidden_sizes=[20, 20],
std_share_network=True,
init_std=embedding_init_std,
max_std=embedding_max_std,
output_nonlinearity=tf.nn.tanh,
min_std=embedding_min_std,
)
self.policy = GaussianMLPTaskEmbeddingPolicy(
name='policy',
env_spec=env.spec,
encoder=task_encoder,
hidden_sizes=[32, 16],
std_share_network=True,
max_std=policy_max_std,
init_std=policy_init_std,
min_std=policy_min_std,
)
self.baseline = LinearMultiFeatureBaseline(
env_spec=env.spec, features=['observations', 'tasks', 'latents'])
def te_ppo_mt50(ctxt, seed, n_epochs, batch_size_per_task):
"""Train Task Embedding PPO with PointEnv.
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.
n_epochs (int): Total number of epochs for training.
batch_size_per_task (int): Batch size of samples for each task.
"""
set_seed(seed)
tasks = MT50.get_train_tasks().all_task_names
envs = [
normalize(GymEnv(MT50.from_task(task), max_episode_length=150))
for task in tasks
]
env = MultiEnvWrapper(envs,
sample_strategy=round_robin_strategy,
mode='del-onehot')
latent_length = 6
inference_window = 6
batch_size = batch_size_per_task * len(tasks)
policy_ent_coeff = 2e-2
encoder_ent_coeff = 2e-4
inference_ce_coeff = 5e-2
embedding_init_std = 0.1
embedding_max_std = 0.2
embedding_min_std = 1e-6
policy_init_std = 1.0
policy_max_std = None
policy_min_std = None
with LocalTFRunner(snapshot_config=ctxt) as runner:
task_embed_spec = TEPPO.get_encoder_spec(env.task_space,
latent_dim=latent_length)
task_encoder = GaussianMLPEncoder(
name='embedding',
embedding_spec=task_embed_spec,
hidden_sizes=(20, 20),
std_share_network=True,
init_std=embedding_init_std,
max_std=embedding_max_std,
output_nonlinearity=tf.nn.tanh,
min_std=embedding_min_std,
)
traj_embed_spec = TEPPO.get_infer_spec(
env.spec,
latent_dim=latent_length,
inference_window_size=inference_window)
inference = GaussianMLPEncoder(
name='inference',
embedding_spec=traj_embed_spec,
hidden_sizes=(20, 10),
std_share_network=True,
init_std=2.0,
output_nonlinearity=tf.nn.tanh,
min_std=embedding_min_std,
)
policy = GaussianMLPTaskEmbeddingPolicy(
name='policy',
env_spec=env.spec,
encoder=task_encoder,
hidden_sizes=(32, 16),
std_share_network=True,
max_std=policy_max_std,
init_std=policy_init_std,
min_std=policy_min_std,
)
baseline = LinearMultiFeatureBaseline(
env_spec=env.spec, features=['observations', 'tasks', 'latents'])
algo = TEPPO(env_spec=env.spec,
policy=policy,
baseline=baseline,
inference=inference,
discount=0.99,
lr_clip_range=0.2,
policy_ent_coeff=policy_ent_coeff,
encoder_ent_coeff=encoder_ent_coeff,
inference_ce_coeff=inference_ce_coeff,
use_softplus_entropy=True,
optimizer_args=dict(
batch_size=32,
max_optimization_epochs=10,
learning_rate=1e-3,
),
inference_optimizer_args=dict(
batch_size=32,
max_optimization_epochs=10,
),
center_adv=True,
stop_ce_gradient=True)
runner.setup(algo,
env,
sampler_cls=LocalSampler,
sampler_args=None,
worker_class=TaskEmbeddingWorker)
runner.train(n_epochs=n_epochs, batch_size=batch_size, plot=False)
def te_ppo_pointenv(ctxt, seed, n_epochs, batch_size_per_task):
"""Train Task Embedding PPO with PointEnv.
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.
n_epochs (int): Total number of epochs for training.
batch_size_per_task (int): Batch size of samples for each task.
"""
set_seed(seed)
tasks = TASKS
latent_length = 2
inference_window = 6
batch_size = batch_size_per_task * len(TASKS)
policy_ent_coeff = 1e-3
encoder_ent_coeff = 1e-3
inference_ce_coeff = 5e-2
embedding_init_std = 0.1
embedding_max_std = 0.2
embedding_min_std = 1e-6
policy_init_std = 1.0
policy_max_std = 2.0
policy_min_std = None
task_names = sorted(tasks.keys())
task_args = [tasks[t]['args'] for t in task_names]
task_kwargs = [tasks[t]['kwargs'] for t in task_names]
with TFTrainer(snapshot_config=ctxt) as trainer:
task_envs = [
PointEnv(*t_args, **t_kwargs, max_episode_length=100)
for t_args, t_kwargs in zip(task_args, task_kwargs)
]
env = MultiEnvWrapper(task_envs, round_robin_strategy, mode='vanilla')
task_embed_spec = TEPPO.get_encoder_spec(env.task_space,
latent_dim=latent_length)
task_encoder = GaussianMLPEncoder(
name='embedding',
embedding_spec=task_embed_spec,
hidden_sizes=(20, 20),
std_share_network=True,
init_std=embedding_init_std,
max_std=embedding_max_std,
output_nonlinearity=tf.nn.tanh,
std_output_nonlinearity=tf.nn.tanh,
min_std=embedding_min_std,
)
traj_embed_spec = TEPPO.get_infer_spec(
env.spec,
latent_dim=latent_length,
inference_window_size=inference_window)
inference = GaussianMLPEncoder(
name='inference',
embedding_spec=traj_embed_spec,
hidden_sizes=(20, 20),
std_share_network=True,
init_std=0.1,
output_nonlinearity=tf.nn.tanh,
std_output_nonlinearity=tf.nn.tanh,
min_std=embedding_min_std,
)
policy = GaussianMLPTaskEmbeddingPolicy(
name='policy',
env_spec=env.spec,
encoder=task_encoder,
hidden_sizes=(32, 16),
std_share_network=True,
max_std=policy_max_std,
init_std=policy_init_std,
min_std=policy_min_std,
)
baseline = LinearMultiFeatureBaseline(
env_spec=env.spec, features=['observations', 'tasks', 'latents'])
sampler = LocalSampler(agents=policy,
envs=env,
max_episode_length=env.spec.max_episode_length,
is_tf_worker=True,
worker_class=TaskEmbeddingWorker)
algo = TEPPO(env_spec=env.spec,
policy=policy,
baseline=baseline,
sampler=sampler,
inference=inference,
discount=0.99,
lr_clip_range=0.2,
policy_ent_coeff=policy_ent_coeff,
encoder_ent_coeff=encoder_ent_coeff,
inference_ce_coeff=inference_ce_coeff,
use_softplus_entropy=True,
optimizer_args=dict(
batch_size=32,
max_optimization_epochs=10,
learning_rate=1e-3,
),
inference_optimizer_args=dict(
batch_size=32,
max_optimization_epochs=10,
learning_rate=1e-3,
),
center_adv=True,
stop_ce_gradient=True)
trainer.setup(algo, env)
trainer.train(n_epochs=n_epochs, batch_size=batch_size, plot=False)
def test_get_action(self, mock_normal, obs_dim, task_num, latent_dim,
action_dim):
mock_normal.return_value = 0.5
env = TfEnv(DummyBoxEnv(obs_dim=obs_dim, action_dim=action_dim))
with mock.patch(
'garage.tf.policies.'
'gaussian_mlp_task_embedding_policy.GaussianMLPModel',
new=SimpleGaussianMLPModel):
embedding_spec = InOutSpec(
input_space=akro.Box(low=np.zeros(task_num),
high=np.ones(task_num)),
output_space=akro.Box(low=np.zeros(latent_dim),
high=np.ones(latent_dim)))
encoder = GaussianMLPEncoder(embedding_spec)
policy = GaussianMLPTaskEmbeddingPolicy(env_spec=env.spec,
encoder=encoder)
env.reset()
obs, _, _, _ = env.step(1)
latent = np.random.random((latent_dim, ))
task = np.zeros(task_num)
task[0] = 1
action1, prob1 = policy.get_action_given_latent(obs, latent)
action2, prob2 = policy.get_action_given_task(obs, task)
action3, prob3 = policy.get_action(
np.concatenate([obs.flatten(), task]))
expected_action = np.full(action_dim, 0.75)
expected_mean = np.full(action_dim, 0.5)
expected_log_std = np.full(action_dim, np.log(0.5))
assert env.action_space.contains(action1)
assert np.array_equal(action1, expected_action)
assert np.array_equal(prob1['mean'], expected_mean)
assert np.array_equal(prob1['log_std'], expected_log_std)
assert env.action_space.contains(action2)
assert np.array_equal(action2, expected_action)
assert np.array_equal(prob2['mean'], expected_mean)
assert np.array_equal(prob2['log_std'], expected_log_std)
assert env.action_space.contains(action3)
assert np.array_equal(action3, expected_action)
assert np.array_equal(prob3['mean'], expected_mean)
assert np.array_equal(prob3['log_std'], expected_log_std)
obses, latents, tasks = [obs] * 3, [latent] * 3, [task] * 3
aug_obses = [np.concatenate([obs.flatten(), task])] * 3
action1n, prob1n = policy.get_actions_given_latents(obses, latents)
action2n, prob2n = policy.get_actions_given_tasks(obses, tasks)
action3n, prob3n = policy.get_actions(aug_obses)
for action, mean, log_std in chain(
zip(action1n, prob1n['mean'], prob1n['log_std']),
zip(action2n, prob2n['mean'], prob2n['log_std']),
zip(action3n, prob3n['mean'], prob3n['log_std'])):
assert env.action_space.contains(action)
assert np.array_equal(action, expected_action)
assert np.array_equal(mean, expected_mean)
assert np.array_equal(log_std, expected_log_std)
def test_dist_info(self, obs_dim, task_num, latent_dim, action_dim):
env = TfEnv(DummyBoxEnv(obs_dim=obs_dim, action_dim=action_dim))
with mock.patch(
'garage.tf.policies.'
'gaussian_mlp_task_embedding_policy.GaussianMLPModel',
new=SimpleGaussianMLPModel):
embedding_spec = InOutSpec(
input_space=akro.Box(low=np.zeros(task_num),
high=np.ones(task_num)),
output_space=akro.Box(low=np.zeros(latent_dim),
high=np.ones(latent_dim)))
encoder = GaussianMLPEncoder(embedding_spec)
policy = GaussianMLPTaskEmbeddingPolicy(env_spec=env.spec,
encoder=encoder)
env.reset()
obs, _, _, _ = env.step(1)
task = np.zeros(task_num)
task[0] = 1
aug_obs = np.concatenate([obs.flatten(), task])
latent = np.random.random(latent_dim)
obs_dim = env.spec.observation_space.flat_dim
obs_ph = tf.compat.v1.placeholder(tf.float32, shape=(None, obs_dim))
task_ph = tf.compat.v1.placeholder(tf.float32, shape=(None, task_num))
latent_ph = tf.compat.v1.placeholder(tf.float32,
shape=(None, latent_dim))
aug_obs_ph = tf.compat.v1.concat([obs_ph, task_ph], axis=1)
dist0_sym = policy.dist_info_sym(aug_obs_ph, name='p0_sym')
dist1_sym = policy.dist_info_sym_given_task(obs_ph,
task_ph,
name='p1_sym')
dist2_sym = policy.dist_info_sym_given_latent(obs_ph,
latent_ph,
name='p2_sym')
# flatten output
expected_mean = [np.full(np.prod(action_dim), 0.5)]
expected_log_std = [np.full(np.prod(action_dim), np.log(0.5))]
prob0 = self.sess.run(dist0_sym,
feed_dict={aug_obs_ph: [aug_obs.flatten()]})
prob1 = self.sess.run(dist1_sym,
feed_dict={
obs_ph: [obs.flatten()],
task_ph: [task]
})
prob2 = self.sess.run(dist2_sym,
feed_dict={
obs_ph: [obs.flatten()],
latent_ph: [latent]
})
prob3 = policy.dist_info(aug_obs)
assert np.array_equal(prob0['mean'].flatten(), expected_mean[0])
assert np.array_equal(prob0['log_std'].flatten(), expected_log_std[0])
assert np.array_equal(prob1['mean'], expected_mean)
assert np.array_equal(prob1['log_std'], expected_log_std)
assert np.array_equal(prob2['mean'], expected_mean)
assert np.array_equal(prob2['log_std'], expected_log_std)
assert np.array_equal(prob3['mean'].flatten(), expected_mean[0])
assert np.array_equal(prob3['log_std'].flatten(), expected_log_std[0])