def test_update_envs_env_update():
max_path_length = 16
env = TfEnv(PointEnv())
policy = FixedPolicy(env.spec,
scripted_actions=[
env.action_space.sample()
for _ in range(max_path_length)
])
tasks = SetTaskSampler(PointEnv)
n_workers = 8
workers = WorkerFactory(seed=100,
max_path_length=max_path_length,
n_workers=n_workers)
sampler = MultiprocessingSampler.from_worker_factory(workers, policy, env)
rollouts = sampler.obtain_samples(0,
161,
np.asarray(policy.get_param_values()),
env_update=tasks.sample(n_workers))
mean_rewards = []
goals = []
for rollout in rollouts.split():
mean_rewards.append(rollout.rewards.mean())
goals.append(rollout.env_infos['task'][0]['goal'])
assert np.var(mean_rewards) > 0
assert np.var(goals) > 0
with pytest.raises(ValueError):
sampler.obtain_samples(0,
10,
np.asarray(policy.get_param_values()),
env_update=tasks.sample(n_workers + 1))
sampler.shutdown_worker()
env.close()
def test_cem_cartpole(self):
"""Test CEM with Cartpole-v1 environment."""
with LocalTFRunner(snapshot_config) as runner:
env = TfEnv(env_name='CartPole-v1')
policy = CategoricalMLPPolicy(name='policy',
env_spec=env.spec,
hidden_sizes=(32, 32))
baseline = LinearFeatureBaseline(env_spec=env.spec)
n_samples = 10
algo = CEM(env_spec=env.spec,
policy=policy,
baseline=baseline,
best_frac=0.1,
max_path_length=100,
n_samples=n_samples)
runner.setup(algo, env, sampler_cls=OnPolicyVectorizedSampler)
rtn = runner.train(n_epochs=10,
batch_size=2048,
n_epoch_cycles=n_samples)
assert rtn > 40
env.close()
def test_obtain_exact_trajectories():
max_path_length = 15
n_workers = 8
env = TfEnv(PointEnv())
per_worker_actions = [env.action_space.sample() for _ in range(n_workers)]
policies = [
FixedPolicy(env.spec, [action] * max_path_length)
for action in per_worker_actions
]
workers = WorkerFactory(seed=100,
max_path_length=max_path_length,
n_workers=n_workers)
sampler = MultiprocessingSampler.from_worker_factory(workers,
policies,
envs=env)
n_traj_per_worker = 3
rollouts = sampler.obtain_exact_trajectories(n_traj_per_worker,
agent_update=policies)
# At least one action per trajectory.
assert sum(rollouts.lengths) >= n_workers * n_traj_per_worker
# All of the trajectories.
assert len(rollouts.lengths) == n_workers * n_traj_per_worker
worker = -1
for count, rollout in enumerate(rollouts.split()):
if count % n_traj_per_worker == 0:
worker += 1
assert (rollout.actions == per_worker_actions[worker]).all()
sampler.shutdown_worker()
env.close()
def test_init_with_crashed_worker():
max_path_length = 16
env = TfEnv(PointEnv())
policy = FixedPolicy(env.spec,
scripted_actions=[
env.action_space.sample()
for _ in range(max_path_length)
])
tasks = SetTaskSampler(lambda: TfEnv(PointEnv()))
n_workers = 2
workers = WorkerFactory(seed=100,
max_path_length=max_path_length,
n_workers=n_workers)
class CrashingPolicy:
def reset(self, **kwargs):
raise Exception('Intentional subprocess crash')
bad_policy = CrashingPolicy()
# This causes worker 2 to crash.
sampler = MultiprocessingSampler.from_worker_factory(
workers, [policy, bad_policy], envs=tasks.sample(n_workers))
rollouts = sampler.obtain_samples(0, 160, None)
assert sum(rollouts.lengths) >= 160
sampler.shutdown_worker()
env.close()
def test_ppo_pendulum_recurrent(self):
"""Test PPO with Pendulum environment and recurrent policy."""
logger.reset()
env = TfEnv(normalize(gym.make("InvertedDoublePendulum-v2")))
policy = GaussianLSTMPolicy(env_spec=env.spec, )
baseline = GaussianMLPBaseline(
env_spec=env.spec,
regressor_args=dict(hidden_sizes=(32, 32)),
)
algo = PPO(
env=env,
policy=policy,
baseline=baseline,
batch_size=2048,
max_path_length=100,
n_itr=10,
discount=0.99,
lr_clip_range=0.01,
optimizer_args=dict(batch_size=32, max_epochs=10),
plot=False,
)
last_avg_ret = algo.train(sess=self.sess)
assert last_avg_ret > 40
env.close()
def test_ddpg_pendulum(self):
"""Test DDPG with Pendulum environment."""
with LocalRunner(self.sess) as runner:
env = TfEnv(gym.make('InvertedDoublePendulum-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,
target_update_tau=1e-2,
n_train_steps=50,
discount=0.9,
min_buffer_size=int(1e4),
exploration_strategy=action_noise,
)
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 test_npo_pendulum(self):
"""Test NPO with Pendulum environment."""
with LocalRunner(self.sess) as runner:
logger.reset()
env = TfEnv(normalize(gym.make("InvertedDoublePendulum-v2")))
policy = GaussianMLPPolicy(
env_spec=env.spec,
hidden_sizes=(64, 64),
hidden_nonlinearity=tf.nn.tanh,
output_nonlinearity=None,
)
baseline = GaussianMLPBaseline(
env_spec=env.spec,
regressor_args=dict(hidden_sizes=(32, 32)),
)
algo = NPO(
env=env,
policy=policy,
baseline=baseline,
max_path_length=100,
discount=0.99,
gae_lambda=0.98,
policy_ent_coeff=0.0)
runner.setup(algo, env)
last_avg_ret = runner.train(n_epochs=10, batch_size=2048)
assert last_avg_ret > 20
env.close()
def test_ppo_pendulum_with_model(self):
"""Test PPO with model, with Pendulum environment."""
with LocalRunner(self.sess) as runner:
env = TfEnv(normalize(gym.make("InvertedDoublePendulum-v2")))
policy = GaussianMLPPolicyWithModel(
env_spec=env.spec,
hidden_sizes=(64, 64),
hidden_nonlinearity=tf.nn.tanh,
output_nonlinearity=None,
)
baseline = GaussianMLPBaselineWithModel(
env_spec=env.spec,
regressor_args=dict(hidden_sizes=(32, 32)),
)
algo = PPO(
env_spec=env.spec,
policy=policy,
baseline=baseline,
max_path_length=100,
discount=0.99,
lr_clip_range=0.01,
optimizer_args=dict(batch_size=32, max_epochs=10),
)
runner.setup(algo, env)
last_avg_ret = runner.train(n_epochs=10, batch_size=2048)
assert last_avg_ret > 30
env.close()
def test_ppo_pendulum_recurrent_continuous_baseline(self):
"""Test PPO with Pendulum environment and recurrent policy."""
with LocalTFRunner(snapshot_config) as runner:
env = TfEnv(normalize(gym.make('InvertedDoublePendulum-v2')))
policy = GaussianLSTMPolicy(env_spec=env.spec, )
baseline = ContinuousMLPBaseline(
env_spec=env.spec,
regressor_args=dict(hidden_sizes=(32, 32)),
)
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,
optimizer_args=dict(
batch_size=32,
max_epochs=10,
),
stop_entropy_gradient=True,
entropy_method='max',
policy_ent_coeff=0.02,
center_adv=False,
)
runner.setup(algo, env)
last_avg_ret = runner.train(n_epochs=10, batch_size=2048)
assert last_avg_ret > 100
env.close()
def test_trpo_cnn_cubecrash(self):
with LocalTFRunner(snapshot_config, sess=self.sess) as runner:
env = TfEnv(normalize(gym.make('CubeCrash-v0')))
policy = CategoricalCNNPolicy(env_spec=env.spec,
conv_filters=(32, 64),
conv_filter_sizes=(8, 4),
conv_strides=(4, 2),
conv_pad='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=0.98,
max_kl_step=0.01,
policy_ent_coeff=0.0,
flatten_input=False)
runner.setup(algo, env)
last_avg_ret = runner.train(n_epochs=10, batch_size=2048)
assert last_avg_ret > -0.9
env.close()
def test_cma_es_cartpole(self):
"""Test CMAES with Cartpole-v1 environment."""
with LocalRunner() as runner:
env = TfEnv(env_name="CartPole-v1")
policy = CategoricalMLPPolicy(
name="policy", env_spec=env.spec, hidden_sizes=(32, 32))
baseline = LinearFeatureBaseline(env_spec=env.spec)
runner.initialize_tf_vars()
n_samples = 20
algo = CMAES(
env_spec=env.spec,
policy=policy,
baseline=baseline,
max_path_length=100,
n_samples=n_samples)
runner.setup(algo, env, sampler_cls=OnPolicyVectorizedSampler)
runner.train(n_epochs=1, batch_size=1000, n_epoch_cycles=n_samples)
# No assertion on return because CMAES is not stable.
env.close()
def test_dm_control_tf_policy(self):
task = ALL_TASKS[0]
with LocalRunner(sess=self.sess) as runner:
env = TfEnv(DmControlEnv.from_suite(*task))
policy = GaussianMLPPolicy(
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=5,
discount=0.99,
max_kl_step=0.01,
)
runner.setup(algo, env)
runner.train(n_epochs=1, batch_size=10)
env.close()
def test_baseline(self):
"""Test the baseline initialization."""
box_env = TfEnv(DummyBoxEnv())
deterministic_mlp_baseline = ContinuousMLPBaseline(env_spec=box_env)
gaussian_mlp_baseline = GaussianMLPBaseline(env_spec=box_env)
self.sess.run(tf.compat.v1.global_variables_initializer())
deterministic_mlp_baseline.get_param_values()
gaussian_mlp_baseline.get_param_values()
box_env.close()
def test_dqn_cartpole_pickle(self):
"""Test DQN with CartPole environment."""
with LocalRunner(self.sess) as runner:
n_epochs = 10
n_epoch_cycles = 10
sampler_batch_size = 500
num_timesteps = n_epochs * n_epoch_cycles * sampler_batch_size
env = TfEnv(gym.make('CartPole-v0'))
replay_buffer = SimpleReplayBuffer(
env_spec=env.spec,
size_in_transitions=int(1e4),
time_horizon=1)
qf = DiscreteMLPQFunction(env_spec=env.spec, hidden_sizes=(64, 64))
policy = DiscreteQfDerivedPolicy(env_spec=env.spec, qf=qf)
epilson_greedy_strategy = EpsilonGreedyStrategy(
env_spec=env.spec,
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_strategy=epilson_greedy_strategy,
replay_buffer=replay_buffer,
qf_lr=1e-4,
discount=1.0,
min_buffer_size=int(1e3),
double_q=False,
n_train_steps=500,
grad_norm_clipping=5.0,
n_epoch_cycles=n_epoch_cycles,
target_network_update_freq=1,
buffer_batch_size=32)
runner.setup(algo, env)
with tf.variable_scope(
'DiscreteMLPQFunction/MLPModel/mlp/hidden_0', reuse=True):
bias = tf.get_variable('bias')
# assign it to all one
old_bias = tf.ones_like(bias).eval()
bias.load(old_bias)
h = pickle.dumps(algo)
with tf.Session(graph=tf.Graph()):
pickle.loads(h)
with tf.variable_scope(
'DiscreteMLPQFunction/MLPModel/mlp/hidden_0',
reuse=True):
new_bias = tf.get_variable('bias')
new_bias = new_bias.eval()
assert np.array_equal(old_bias, new_bias)
env.close()
class TestContinuousPolicies(TfGraphTestCase):
def setup_method(self):
super().setup_method()
self.env = TfEnv(DummyBoxEnv())
self.obs_var = tf.compat.v1.placeholder(
tf.float32,
shape=[None, None, self.env.observation_space.flat_dim],
name='obs')
def teardown_method(self):
self.env.close()
super().teardown_method()
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_gaussian_gru_policy(self):
gaussian_gru_policy = GaussianGRUPolicy(env_spec=self.env,
hidden_dim=1,
state_include_action=False)
self.sess.run(tf.compat.v1.global_variables_initializer())
gaussian_gru_policy.build(self.obs_var)
gaussian_gru_policy.reset()
obs = self.env.observation_space.high
assert gaussian_gru_policy.get_action(obs)
def test_gaussian_lstm_policy(self):
gaussian_lstm_policy = GaussianLSTMPolicy(env_spec=self.env,
hidden_dim=1,
state_include_action=False)
self.sess.run(tf.compat.v1.global_variables_initializer())
gaussian_lstm_policy.build(self.obs_var)
gaussian_lstm_policy.reset()
obs = self.env.observation_space.high
assert gaussian_lstm_policy.get_action(obs)
def test_gaussian_mlp_policy(self):
gaussian_mlp_policy = GaussianMLPPolicy(env_spec=self.env,
hidden_sizes=(1, ))
self.sess.run(tf.compat.v1.global_variables_initializer())
gaussian_mlp_policy.build(self.obs_var)
obs = self.env.observation_space.high
assert gaussian_mlp_policy.get_action(obs)
class TestContinuousPolicies(TfGraphTestCase):
def setUp(self):
super().setUp()
self.env = TfEnv(DummyBoxEnv())
def tearDown(self):
self.env.close()
super().tearDown()
def test_continuous_mlp_policy(self):
continuous_mlp_policy = ContinuousMLPPolicy(env_spec=self.env,
hidden_sizes=(1, ))
self.sess.run(tf.global_variables_initializer())
obs = self.env.observation_space.high
assert continuous_mlp_policy.get_action(obs)
def test_deterministic_mlp_policy(self):
deterministic_mlp_policy = DeterministicMLPPolicy(env_spec=self.env,
hidden_sizes=(1, ))
self.sess.run(tf.global_variables_initializer())
obs = self.env.observation_space.high
assert deterministic_mlp_policy.get_action(obs)
def test_gaussian_gru_policy(self):
gaussian_gru_policy = GaussianGRUPolicy(env_spec=self.env,
hidden_dim=1)
self.sess.run(tf.global_variables_initializer())
gaussian_gru_policy.reset()
obs = self.env.observation_space.high
assert gaussian_gru_policy.get_action(obs)
def test_gaussian_lstm_policy(self):
gaussian_lstm_policy = GaussianLSTMPolicy(env_spec=self.env,
hidden_dim=1)
self.sess.run(tf.global_variables_initializer())
gaussian_lstm_policy.reset()
obs = self.env.observation_space.high
assert gaussian_lstm_policy.get_action(obs)
def test_gaussian_mlp_policy(self):
gaussian_mlp_policy = GaussianMLPPolicy(env_spec=self.env,
hidden_sizes=(1, ))
self.sess.run(tf.global_variables_initializer())
obs = self.env.observation_space.high
assert gaussian_mlp_policy.get_action(obs)
class TestGrayscale(unittest.TestCase):
def setUp(self):
self.env = TfEnv(DummyDiscretePixelEnv(random=False))
self.env_g = TfEnv(Grayscale(DummyDiscretePixelEnv(random=False)))
def tearDown(self):
self.env.close()
self.env_g.close()
def test_gray_scale_invalid_environment_type(self):
with self.assertRaises(ValueError):
self.env.observation_space = Discrete(64)
Grayscale(self.env)
def test_gray_scale_invalid_environment_shape(self):
with self.assertRaises(ValueError):
self.env.observation_space = Box(
low=0, high=255, shape=(4, ), dtype=np.uint8)
Grayscale(self.env)
def test_grayscale_observation_space(self):
assert self.env_g.observation_space.shape == (
self.env.observation_space.shape[:-1])
def test_grayscale_reset(self):
"""
RGB to grayscale conversion using scikit-image.
Weights used for conversion:
Y = 0.2125 R + 0.7154 G + 0.0721 B
Reference:
http://scikit-image.org/docs/dev/api/skimage.color.html#skimage.color.rgb2grey
"""
gray_scale_output = np.round(
np.dot(self.env.reset()[:, :, :3],
[0.2125, 0.7154, 0.0721])).astype(np.uint8)
np.testing.assert_array_almost_equal(gray_scale_output,
self.env_g.reset())
def test_grayscale_step(self):
self.env.reset()
self.env_g.reset()
obs, _, _, _ = self.env.step(1)
obs_g, _, _, _ = self.env_g.step(1)
gray_scale_output = np.round(
np.dot(obs[:, :, :3], [0.2125, 0.7154, 0.0721])).astype(np.uint8)
np.testing.assert_array_almost_equal(gray_scale_output, obs_g)
class TestStackFrames(unittest.TestCase):
def setUp(self):
self.n_frames = 4
self.env = TfEnv(DummyDiscrete2DEnv(random=False))
self.env_s = TfEnv(
StackFrames(
DummyDiscrete2DEnv(random=False), n_frames=self.n_frames))
self.width, self.height = self.env.observation_space.shape
def tearDown(self):
self.env.close()
self.env_s.close()
def test_stack_frames_invalid_environment_type(self):
with self.assertRaises(ValueError):
self.env.observation_space = Discrete(64)
StackFrames(self.env, n_frames=4)
def test_stack_frames_invalid_environment_shape(self):
with self.assertRaises(ValueError):
self.env.observation_space = Box(
low=0, high=255, shape=(4, ), dtype=np.uint8)
StackFrames(self.env, n_frames=4)
def test_stack_frames_output_observation_space(self):
assert self.env_s.observation_space.shape == (self.width, self.height,
self.n_frames)
def test_stack_frames_for_reset(self):
frame_stack = self.env.reset()
for i in range(self.n_frames - 1):
frame_stack = np.dstack((frame_stack, self.env.reset()))
np.testing.assert_array_equal(self.env_s.reset(), frame_stack)
def test_stack_frames_for_step(self):
self.env.reset()
self.env_s.reset()
frame_stack = np.empty((self.width, self.height, self.n_frames))
for i in range(10):
frame_stack = frame_stack[:, :, 1:]
obs, _, _, _ = self.env.step(1)
frame_stack = np.dstack((frame_stack, obs))
obs_stack, _, _, _ = self.env_s.step(1)
np.testing.assert_array_equal(obs_stack, frame_stack)
def test_dqn_cartpole_grad_clip(self):
"""Test DQN with CartPole environment."""
with LocalRunner(self.sess) as runner:
n_epochs = 10
n_epoch_cycles = 10
sampler_batch_size = 500
num_timesteps = n_epochs * n_epoch_cycles * sampler_batch_size
env = TfEnv(gym.make('CartPole-v0'))
replay_buffer = SimpleReplayBuffer(
env_spec=env.spec,
size_in_transitions=int(1e4),
time_horizon=1)
qf = DiscreteMLPQFunction(env_spec=env.spec, hidden_sizes=(64, 64))
policy = DiscreteQfDerivedPolicy(env_spec=env.spec, qf=qf)
epilson_greedy_strategy = EpsilonGreedyStrategy(
env_spec=env.spec,
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_strategy=epilson_greedy_strategy,
replay_buffer=replay_buffer,
qf_lr=1e-4,
discount=1.0,
min_buffer_size=int(1e3),
double_q=False,
n_train_steps=500,
grad_norm_clipping=5.0,
n_epoch_cycles=n_epoch_cycles,
target_network_update_freq=1,
buffer_batch_size=32)
runner.setup(algo, env)
last_avg_ret = runner.train(
n_epochs=n_epochs,
n_epoch_cycles=n_epoch_cycles,
batch_size=sampler_batch_size)
assert last_avg_ret > 20
env.close()
class TestRaySamplerTF():
"""
Uses mock policy for 4x4 gridworldenv
'4x4': [
'SFFF',
'FHFH',
'FFFH',
'HFFG'
]
0: left
1: down
2: right
3: up
-1: no move
'S' : starting point
'F' or '.': free space
'W' or 'x': wall
'H' or 'o': hole (terminates episode)
'G' : goal
[2,2,1,0,3,1,1,1,2,2,1,1,1,2,2,1]
"""
def setup_method(self):
ray.init(local_mode=True, ignore_reinit_error=True)
self.env = TfEnv(GridWorldEnv(desc='4x4'))
self.policy = ScriptedPolicy(
scripted_actions=[2, 2, 1, 0, 3, 1, 1, 1, 2, 2, 1, 1, 1, 2, 2, 1])
self.algo = Mock(env_spec=self.env.spec,
policy=self.policy,
max_path_length=16)
def teardown_method(self):
self.env.close()
def test_ray_batch_sampler(self):
workers = WorkerFactory(seed=100,
max_path_length=self.algo.max_path_length)
sampler1 = RaySamplerTF(workers,
self.policy,
self.env,
num_processors=1)
sampler1.start_worker()
sampler1.shutdown_worker()
def test_init_with_env_updates():
max_path_length = 16
env = TfEnv(PointEnv())
policy = FixedPolicy(env.spec,
scripted_actions=[
env.action_space.sample()
for _ in range(max_path_length)
])
tasks = SetTaskSampler(lambda: TfEnv(PointEnv()))
n_workers = 8
workers = WorkerFactory(seed=100,
max_path_length=max_path_length,
n_workers=n_workers)
sampler = MultiprocessingSampler.from_worker_factory(
workers, policy, envs=tasks.sample(n_workers))
rollouts = sampler.obtain_samples(0, 160, policy)
assert sum(rollouts.lengths) >= 160
sampler.shutdown_worker()
env.close()
class TestNormalizedGym(unittest.TestCase):
def setUp(self):
self.env = TfEnv(
normalize(
gym.make('Pendulum-v0'),
normalize_reward=True,
normalize_obs=True,
flatten_obs=True))
def tearDown(self):
self.env.close()
def test_does_not_modify_action(self):
a = self.env.action_space.sample()
a_copy = a
self.env.reset()
self.env.step(a)
self.assertEqual(a, a_copy)
def test_flatten(self):
for _ in range(10):
self.env.reset()
for _ in range(5):
self.env.render()
action = self.env.action_space.sample()
next_obs, _, done, _ = self.env.step(action)
self.assertEqual(next_obs.shape,
self.env.observation_space.low.shape)
if done:
break
def test_unflatten(self):
for _ in range(10):
self.env.reset()
for _ in range(5):
action = self.env.action_space.sample()
next_obs, _, done, _ = self.env.step(action)
self.assertEqual(
self.env.observation_space.flatten(next_obs).shape,
self.env.observation_space.flat_dim)
if done:
break
class TestDiscretePolicies(TfGraphTestCase):
def setup_method(self):
super().setup_method()
self.env = TfEnv(DummyDiscreteEnv())
self.obs_var = tf.compat.v1.placeholder(
tf.float32,
shape=[None, None, self.env.observation_space.flat_dim],
name='obs')
def teardown_method(self):
self.env.close()
super().teardown_method()
def test_categorial_gru_policy(self):
categorical_gru_policy = CategoricalGRUPolicy(
env_spec=self.env, hidden_dim=1, state_include_action=False)
self.sess.run(tf.compat.v1.global_variables_initializer())
categorical_gru_policy.build(self.obs_var)
categorical_gru_policy.reset()
obs = self.env.observation_space.high
assert categorical_gru_policy.get_action(obs)
def test_categorical_lstm_policy(self):
categorical_lstm_policy = CategoricalLSTMPolicy(
env_spec=self.env, hidden_dim=1, state_include_action=False)
self.sess.run(tf.compat.v1.global_variables_initializer())
categorical_lstm_policy.build(self.obs_var)
categorical_lstm_policy.reset()
obs = self.env.observation_space.high
assert categorical_lstm_policy.get_action(obs)
def test_categorial_mlp_policy(self):
categorical_mlp_policy = CategoricalMLPPolicy(env_spec=self.env,
hidden_sizes=(1, ))
self.sess.run(tf.compat.v1.global_variables_initializer())
categorical_mlp_policy.build(self.obs_var)
obs = self.env.observation_space.high
assert categorical_mlp_policy.get_action(obs)
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')))
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 = 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_policy=exploration_policy,
)
runner.setup(algo, env)
last_avg_ret = runner.train(n_epochs=10, batch_size=100)
assert last_avg_ret > 10
env.close()
class TestNormalizedGym:
def setup_method(self):
self.env = TfEnv(
normalize(gym.make('Pendulum-v0'),
normalize_reward=True,
normalize_obs=True,
flatten_obs=True))
def teardown_method(self):
self.env.close()
def test_does_not_modify_action(self):
a = self.env.action_space.sample()
a_copy = a
self.env.reset()
self.env.step(a)
assert a == a_copy
def test_flatten(self):
for _ in range(10):
self.env.reset()
for _ in range(5):
self.env.render()
action = self.env.action_space.sample()
next_obs, _, done, _ = self.env.step(action)
assert next_obs.shape == self.env.observation_space.low.shape
if done:
break
def test_unflatten(self):
for _ in range(10):
self.env.reset()
for _ in range(5):
action = self.env.action_space.sample()
next_obs, _, done, _ = self.env.step(action)
# yapf: disable
assert (self.env.observation_space.flatten(next_obs).shape
== self.env.observation_space.flat_dim)
# yapf: enable
if done:
break
def test_ppo_pendulum_recurrent_continuous_baseline(self):
"""Test PPO with Pendulum environment and recurrent policy."""
with LocalRunner() as runner:
env = TfEnv(normalize(gym.make('InvertedDoublePendulum-v2')))
policy = GaussianLSTMPolicy(env_spec=env.spec, )
baseline = ContinuousMLPBaselineWithModel(
env_spec=env.spec,
regressor_args=dict(hidden_sizes=(32, 32)),
)
algo = PPO(env_spec=env.spec,
policy=policy,
baseline=baseline,
max_path_length=100,
discount=0.99,
lr_clip_range=0.01,
optimizer_args=dict(batch_size=32, max_epochs=10))
runner.setup(algo, env)
last_avg_ret = runner.train(n_epochs=10, batch_size=2048)
assert last_avg_ret > 30
env.close()
def test_baseline(self):
"""Test the baseline initialization."""
box_env = TfEnv(DummyBoxEnv())
deterministic_mlp_baseline = DeterministicMLPBaseline(env_spec=box_env)
gaussian_mlp_baseline = GaussianMLPBaseline(env_spec=box_env)
discrete_env = TfEnv(Resize(DummyDiscrete2DEnv(), width=64, height=64))
gaussian_conv_baseline = GaussianConvBaseline(
env_spec=discrete_env,
regressor_args=dict(conv_filters=[32, 32],
conv_filter_sizes=[1, 1],
conv_strides=[1, 1],
conv_pads=["VALID", "VALID"],
hidden_sizes=(32, 32)))
self.sess.run(tf.global_variables_initializer())
deterministic_mlp_baseline.get_param_values(trainable=True)
gaussian_mlp_baseline.get_param_values(trainable=True)
gaussian_conv_baseline.get_param_values(trainable=True)
box_env.close()
class TestRepeatAction(unittest.TestCase):
def setUp(self):
self.env = TfEnv(DummyDiscreteEnv(random=False))
self.env_r = TfEnv(
RepeatAction(DummyDiscreteEnv(random=False), n_frame_to_repeat=4))
def tearDown(self):
self.env.close()
self.env_r.close()
def test_repeat_action_reset(self):
np.testing.assert_array_equal(self.env.reset(), self.env_r.reset())
def test_repeat_action_step(self):
self.env.reset()
self.env_r.reset()
obs_repeat, _, _, _ = self.env_r.step(1)
for i in range(4):
obs, _, _, _ = self.env.step(1)
np.testing.assert_array_equal(obs, obs_repeat)
def test_erwr_cartpole(self):
"""Test ERWR with Cartpole-v1 environment."""
env = TfEnv(env_name="CartPole-v1")
policy = CategoricalMLPPolicy(name="policy",
env_spec=env.spec,
hidden_sizes=(32, 32))
baseline = LinearFeatureBaseline(env_spec=env.spec)
algo = ERWR(env=env,
policy=policy,
baseline=baseline,
batch_size=10000,
max_path_length=100,
n_itr=10,
discount=0.99)
last_avg_ret = algo.train(sess=self.sess)
assert last_avg_ret > 80
env.close()
def test_reps_cartpole(self):
"""Test REPS with gym Cartpole environment."""
with LocalRunner(sess=self.sess) as runner:
env = TfEnv(gym.make('CartPole-v0'))
policy = CategoricalMLPPolicy(env_spec=env.spec,
hidden_sizes=[32, 32])
baseline = LinearFeatureBaseline(env_spec=env.spec)
algo = REPS(env_spec=env.spec,
policy=policy,
baseline=baseline,
max_path_length=100,
discount=0.99)
runner.setup(algo, env)
last_avg_ret = runner.train(n_epochs=10, batch_size=4000)
assert last_avg_ret > 5
env.close()