def test_is_pickleable(self):
env = GarageEnv(DummyDiscreteEnv(obs_dim=(1, ), action_dim=1))
policy = CategoricalGRUPolicy(env_spec=env.spec,
state_include_action=False)
obs = env.reset()
policy._gru_cell.weights[0].load(
tf.ones_like(policy._gru_cell.weights[0]).eval())
state_input = tf.compat.v1.placeholder(tf.float32,
shape=(None, None,
policy.input_dim))
dist_sym = policy.build(state_input, name='dist_sym').dist
output1 = self.sess.run(
[dist_sym.probs],
feed_dict={state_input: [[obs.flatten()], [obs.flatten()]]})
p = pickle.dumps(policy)
with tf.compat.v1.Session(graph=tf.Graph()) as sess:
policy_pickled = pickle.loads(p)
state_input = tf.compat.v1.placeholder(tf.float32,
shape=(None, None,
policy.input_dim))
dist_sym = policy_pickled.build(state_input, name='dist_sym').dist
# yapf: disable
output2 = sess.run(
[dist_sym.probs],
feed_dict={
state_input: [[obs.flatten()], [obs.flatten()]]
})
# yapf: enable
assert np.array_equal(output1, output2)
def test_categorial_gru_policy(self):
categorical_gru_policy = CategoricalGRUPolicy(
env_spec=self.env, hidden_dim=1, state_include_action=False)
categorical_gru_policy.reset()
obs = self.env.observation_space.high
assert categorical_gru_policy.get_action(obs)
def test_clone(self):
env = GarageEnv(DummyDiscreteEnv(obs_dim=(10, ), action_dim=4))
policy = CategoricalGRUPolicy(env_spec=env.spec)
policy_clone = policy.clone('CategoricalGRUPolicyClone')
assert policy.env_spec == policy_clone.env_spec
for cloned_param, param in zip(policy_clone.parameters.values(),
policy.parameters.values()):
assert np.array_equal(cloned_param, param)
def test_categorial_gru_policy(self):
categorical_gru_policy = CategoricalGRUPolicy(env_spec=self.env,
hidden_dim=1)
self.sess.run(tf.global_variables_initializer())
categorical_gru_policy.reset()
obs = self.env.observation_space.high
assert categorical_gru_policy.get_action(obs)
def test_is_pickleable(self):
env = TfEnv(DummyDiscreteEnv(obs_dim=(1, ), action_dim=1))
with mock.patch(('garage.tf.policies.'
'categorical_gru_policy.GRUModel'),
new=SimpleGRUModel):
policy = CategoricalGRUPolicy(env_spec=env.spec,
state_include_action=False)
env.reset()
obs = env.reset()
with tf.compat.v1.variable_scope('CategoricalGRUPolicy/prob_network',
reuse=True):
return_var = tf.compat.v1.get_variable('return_var')
# assign it to all one
return_var.load(tf.ones_like(return_var).eval())
output1 = self.sess.run(
policy.model.outputs[0],
feed_dict={policy.model.input: [[obs.flatten()], [obs.flatten()]]})
p = pickle.dumps(policy)
with tf.compat.v1.Session(graph=tf.Graph()) as sess:
policy_pickled = pickle.loads(p)
# yapf: disable
output2 = sess.run(
policy_pickled.model.outputs[0],
feed_dict={
policy_pickled.model.input: [[obs.flatten()],
[obs.flatten()]]
})
# yapf: enable
assert np.array_equal(output1, output2)
def test_invalid_env(self):
env = TfEnv(DummyBoxEnv())
with mock.patch(('garage.tf.policies.'
'categorical_gru_policy.GRUModel'),
new=SimpleGRUModel):
with pytest.raises(ValueError):
CategoricalGRUPolicy(env_spec=env.spec)
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)
sampler = LocalSampler(
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,
optimizer_args=dict(hvp_approach=FiniteDifferenceHVP(
base_eps=1e-5)))
trainer.setup(algo, env)
last_avg_ret = trainer.train(n_epochs=10, batch_size=2048)
assert last_avg_ret > 40
env.close()
def train_gru_trpo(ctxt=None):
set_seed(seed)
with TFTrainer(snapshot_config=ctxt) as trainer:
env = MyGymEnv(gym_env, max_episode_length=100)
policy = CategoricalGRUPolicy(name='policy',
env_spec=env.spec,
state_include_action=False)
baseline = LinearFeatureBaseline(env_spec=env.spec)
sampler = LocalSampler(
agents=policy,
envs=env,
max_episode_length=env.spec.max_episode_length,
worker_class=FragmentWorker,
)
self.algo = LoggedTRPO(
env=env,
env_spec=env.spec,
policy=policy,
baseline=baseline,
sampler=sampler,
discount=0.99,
max_kl_step=0.01,
optimizer_args=dict(hvp_approach=FiniteDifferenceHVP(
base_eps=1e-5)))
trainer.setup(self.algo, env)
trainer.train(n_epochs=n_eps, batch_size=4000)
return self.algo.rew_chkpts
def test_is_pickleable(self):
env = GarageEnv(DummyDiscreteEnv(obs_dim=(1, ), action_dim=1))
policy = CategoricalGRUPolicy(env_spec=env.spec,
state_include_action=False)
obs = env.reset()
policy.model._gru_cell.weights[0].load(
tf.ones_like(policy.model._gru_cell.weights[0]).eval())
output1 = self.sess.run(
[policy.distribution.probs],
feed_dict={policy.model.input: [[obs.flatten()], [obs.flatten()]]})
p = pickle.dumps(policy)
with tf.compat.v1.Session(graph=tf.Graph()) as sess:
policy_pickled = pickle.loads(p)
# yapf: disable
output2 = sess.run(
[policy_pickled.distribution.probs],
feed_dict={
policy_pickled.model.input: [[obs.flatten()],
[obs.flatten()]]
})
# yapf: enable
assert np.array_equal(output1, output2)
def train_ppo(ctxt=None):
set_seed(seed)
with TFTrainer(ctxt) as trainer:
env = MyGymEnv(gym_env, max_episode_length=100)
policy = CategoricalGRUPolicy(name='policy',
env_spec=env.spec,
state_include_action=False)
baseline = LinearFeatureBaseline(env_spec=env.spec)
sampler = LocalSampler(
agents=policy,
envs=env,
max_episode_length=env.spec.max_episode_length,
worker_class=FragmentWorker,
is_tf_worker=True,
)
self.algo = LoggedPPO(
env=env,
env_spec=env.spec,
policy=policy,
baseline=baseline,
sampler=sampler,
discount=0.99,
center_adv=False,
optimizer_args=dict(max_optimization_epochs=8))
trainer.setup(self.algo, env)
trainer.train(n_epochs=n_eps, batch_size=4000)
return self.algo.rew_chkpts
def run_garage(env, seed, log_dir):
"""Create garage model and training.
Replace the ppo with the algorithm you want to run.
Args:
env (gym.Env): Environment of the task.
seed (int): Random seed for the trial.
log_dir (str): Log dir path.
Returns:
str: Path to output csv file
"""
deterministic.set_seed(seed)
config = tf.compat.v1.ConfigProto(allow_soft_placement=True,
intra_op_parallelism_threads=12,
inter_op_parallelism_threads=12)
sess = tf.compat.v1.Session(config=config)
with LocalTFRunner(snapshot_config, sess=sess, max_cpus=12) as runner:
env = TfEnv(normalize(env))
policy = CategoricalGRUPolicy(
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_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),
),
)
# 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, sampler_args=dict(n_envs=12))
runner.train(n_epochs=488, batch_size=2048)
dowel_logger.remove_all()
return tabular_log_file
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_build_state_not_include_action(self, obs_dim, action_dim,
hidden_dim):
env = GarageEnv(
DummyDiscreteEnv(obs_dim=obs_dim, action_dim=action_dim))
policy = CategoricalGRUPolicy(env_spec=env.spec,
hidden_dim=hidden_dim,
state_include_action=False)
policy.reset(do_resets=None)
obs = env.reset()
state_input = tf.compat.v1.placeholder(tf.float32,
shape=(None, None,
policy.input_dim))
dist_sym = policy.build(state_input, name='dist_sym').dist
output1 = self.sess.run(
[policy.distribution.probs],
feed_dict={policy.model.input: [[obs.flatten()], [obs.flatten()]]})
output2 = self.sess.run(
[dist_sym.probs],
feed_dict={state_input: [[obs.flatten()], [obs.flatten()]]})
assert np.array_equal(output1, output2)
def test_get_action(self, mock_rand, obs_dim, action_dim, hidden_dim):
mock_rand.return_value = 0
env = TfEnv(DummyDiscreteEnv(obs_dim=obs_dim, action_dim=action_dim))
with mock.patch(('garage.tf.policies.'
'categorical_gru_policy.GRUModel'),
new=SimpleGRUModel):
policy = CategoricalGRUPolicy(env_spec=env.spec,
state_include_action=False)
policy.reset()
obs = env.reset()
expected_prob = np.full(action_dim, 0.5)
action, agent_info = policy.get_action(obs)
assert env.action_space.contains(action)
assert action == 0
assert np.array_equal(agent_info['prob'], expected_prob)
actions, agent_infos = policy.get_actions([obs])
for action, prob in zip(actions, agent_infos['prob']):
assert env.action_space.contains(action)
assert action == 0
assert np.array_equal(prob, expected_prob)
def test_dist_info_sym(self, obs_dim, action_dim, hidden_dim):
env = TfEnv(DummyDiscreteEnv(obs_dim=obs_dim, action_dim=action_dim))
obs_ph = tf.compat.v1.placeholder(
tf.float32, shape=(None, None, env.observation_space.flat_dim))
with mock.patch(('garage.tf.policies.'
'categorical_gru_policy.GRUModel'),
new=SimpleGRUModel):
policy = CategoricalGRUPolicy(env_spec=env.spec,
state_include_action=False)
policy.reset()
obs = env.reset()
dist_sym = policy.dist_info_sym(obs_var=obs_ph,
state_info_vars=None,
name='p2_sym')
dist = self.sess.run(
dist_sym, feed_dict={obs_ph: [[obs.flatten()], [obs.flatten()]]})
assert np.array_equal(dist['prob'], np.full((2, 1, action_dim), 0.5))
def categorical_gru_policy(ctxt, env_id, seed):
"""Create Categorical CNN Policy on TF-PPO.
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 = CategoricalGRUPolicy(
env_spec=env.spec,
hidden_dim=32,
hidden_nonlinearity=tf.nn.tanh,
)
baseline = LinearFeatureBaseline(env_spec=env.spec)
sampler = LocalSampler(agents=policy,
envs=env,
max_episode_length=env.spec.max_episode_length)
algo = PPO(
env_spec=env.spec,
policy=policy,
baseline=baseline,
sampler=sampler,
discount=0.99,
gae_lambda=0.95,
lr_clip_range=0.2,
policy_ent_coeff=0.0,
optimizer_args=dict(
batch_size=32,
max_optimization_epochs=10,
learning_rate=1e-3,
),
)
trainer.setup(algo, env)
trainer.train(n_epochs=488, batch_size=2048)
def test_get_action(self, obs_dim, action_dim, hidden_dim):
env = TfEnv(DummyDiscreteEnv(obs_dim=obs_dim, action_dim=action_dim))
obs_var = tf.compat.v1.placeholder(
tf.float32,
shape=[None, None, env.observation_space.flat_dim],
name='obs')
policy = CategoricalGRUPolicy(env_spec=env.spec,
hidden_dim=hidden_dim,
state_include_action=False)
policy.build(obs_var)
policy.reset(do_resets=None)
obs = env.reset()
action, _ = policy.get_action(obs.flatten())
assert env.action_space.contains(action)
actions, _ = policy.get_actions([obs.flatten()])
for action in actions:
assert env.action_space.contains(action)
def test_policies(self):
"""Test the policies initialization."""
box_env = TfEnv(DummyBoxEnv())
discrete_env = TfEnv(DummyDiscreteEnv())
categorical_gru_policy = CategoricalGRUPolicy(env_spec=discrete_env,
hidden_dim=1)
categorical_lstm_policy = CategoricalLSTMPolicy(env_spec=discrete_env,
hidden_dim=1)
categorical_mlp_policy = CategoricalMLPPolicy(env_spec=discrete_env,
hidden_sizes=(1, ))
continuous_mlp_policy = ContinuousMLPPolicy(env_spec=box_env,
hidden_sizes=(1, ))
deterministic_mlp_policy = DeterministicMLPPolicy(env_spec=box_env,
hidden_sizes=(1, ))
gaussian_gru_policy = GaussianGRUPolicy(env_spec=box_env, hidden_dim=1)
gaussian_lstm_policy = GaussianLSTMPolicy(env_spec=box_env,
hidden_dim=1)
gaussian_mlp_policy = GaussianMLPPolicy(env_spec=box_env,
hidden_sizes=(1, ))
def test_get_action(self, obs_dim, action_dim, hidden_dim):
env = GarageEnv(
DummyDiscreteEnv(obs_dim=obs_dim, action_dim=action_dim))
policy = CategoricalGRUPolicy(env_spec=env.spec,
hidden_dim=hidden_dim,
state_include_action=False)
policy.reset(do_resets=None)
obs = env.reset()
action, _ = policy.get_action(obs.flatten())
assert env.action_space.contains(action)
actions, _ = policy.get_actions([obs.flatten()])
for action in actions:
assert env.action_space.contains(action)
def test_trpo_gru_cartpole(self):
with LocalTFRunner(snapshot_config, sess=self.sess) as runner:
env = TfEnv(normalize(gym.make('CartPole-v1')))
policy = CategoricalGRUPolicy(name='policy', env_spec=env.spec)
baseline = LinearFeatureBaseline(env_spec=env.spec)
algo = TRPO(env_spec=env.spec,
policy=policy,
baseline=baseline,
max_path_length=100,
discount=0.99,
max_kl_step=0.01,
optimizer_args=dict(hvp_approach=FiniteDifferenceHvp(
base_eps=1e-5)))
runner.setup(algo, env)
last_avg_ret = runner.train(n_epochs=10, batch_size=2048)
assert last_avg_ret > 80
env.close()
def test_build_state_include_action(self, obs_dim, action_dim, hidden_dim):
env = GarageEnv(
DummyDiscreteEnv(obs_dim=obs_dim, action_dim=action_dim))
policy = CategoricalGRUPolicy(env_spec=env.spec,
hidden_dim=hidden_dim,
state_include_action=True)
policy.reset(do_resets=None)
obs = env.reset()
state_input = tf.compat.v1.placeholder(tf.float32,
shape=(None, None,
policy.input_dim))
dist_sym = policy.build(state_input, name='dist_sym').dist
dist_sym2 = policy.build(state_input, name='dist_sym2').dist
concat_obs = np.concatenate([obs.flatten(), np.zeros(action_dim)])
output1 = self.sess.run(
[dist_sym.probs],
feed_dict={state_input: [[concat_obs], [concat_obs]]})
output2 = self.sess.run(
[dist_sym2.probs],
feed_dict={state_input: [[concat_obs], [concat_obs]]})
assert np.array_equal(output1, output2)
def test_dist_info_sym_wrong_input(self):
env = TfEnv(DummyDiscreteEnv(obs_dim=(1, ), action_dim=1))
obs_ph = tf.compat.v1.placeholder(
tf.float32, shape=(None, None, env.observation_space.flat_dim))
with mock.patch(('garage.tf.policies.'
'categorical_gru_policy.GRUModel'),
new=SimpleGRUModel):
policy = CategoricalGRUPolicy(env_spec=env.spec,
state_include_action=True)
policy.reset()
obs = env.reset()
policy.dist_info_sym(
obs_var=obs_ph,
state_info_vars={'prev_action': np.zeros((3, 1, 1))},
name='p2_sym')
# observation batch size = 2 but prev_action batch size = 3
with pytest.raises(tf.errors.InvalidArgumentError):
self.sess.run(
policy.model.networks['p2_sym'].input,
feed_dict={obs_ph: [[obs.flatten()], [obs.flatten()]]})
def test_invalid_env(self):
env = GarageEnv(DummyBoxEnv())
with pytest.raises(ValueError):
CategoricalGRUPolicy(env_spec=env.spec)
def test_clone(self):
env = GarageEnv(DummyDiscreteEnv(obs_dim=(10, ), action_dim=4))
policy = CategoricalGRUPolicy(env_spec=env.spec)
policy_clone = policy.clone('CategoricalGRUPolicyClone')
assert policy.env_spec == policy_clone.env_spec
def test_state_info_specs_with_state_include_action(self):
env = GarageEnv(DummyDiscreteEnv(obs_dim=(10, ), action_dim=4))
policy = CategoricalGRUPolicy(env_spec=env.spec,
state_include_action=True)
assert policy.state_info_specs == [('prev_action', (4, ))]
def test_state_info_specs(self):
env = GarageEnv(DummyDiscreteEnv(obs_dim=(10, ), action_dim=4))
policy = CategoricalGRUPolicy(env_spec=env.spec,
state_include_action=False)
assert policy.state_info_specs == []
class TestCategoricalGRUPolicyWithModelTransit(TfGraphTestCase):
def setUp(self):
super().setUp()
env = TfEnv(DummyDiscreteEnv(obs_dim=(1, ), action_dim=1))
self.default_initializer = tf.constant_initializer(1)
self.default_hidden_nonlinearity = tf.nn.tanh
self.default_recurrent_nonlinearity = tf.nn.sigmoid
self.default_output_nonlinearity = None
self.time_step = 1
self.policy1 = CategoricalGRUPolicy(
env_spec=env.spec,
hidden_dim=4,
hidden_nonlinearity=self.default_hidden_nonlinearity,
recurrent_nonlinearity=self.default_recurrent_nonlinearity,
recurrent_w_x_init=self.default_initializer,
recurrent_w_h_init=self.default_initializer,
output_nonlinearity=self.default_output_nonlinearity,
output_w_init=self.default_initializer,
state_include_action=True,
name='P1')
self.policy2 = CategoricalGRUPolicy(
env_spec=env.spec,
hidden_dim=4,
hidden_nonlinearity=self.default_hidden_nonlinearity,
recurrent_nonlinearity=self.default_recurrent_nonlinearity,
recurrent_w_x_init=self.default_initializer,
recurrent_w_h_init=self.default_initializer,
output_nonlinearity=self.default_output_nonlinearity,
output_w_init=tf.constant_initializer(2),
state_include_action=True,
name='P2')
self.sess.run(tf.global_variables_initializer())
self.policy3 = CategoricalGRUPolicyWithModel(
env_spec=env.spec,
hidden_dim=4,
hidden_nonlinearity=self.default_hidden_nonlinearity,
hidden_w_init=self.default_initializer,
recurrent_nonlinearity=self.default_recurrent_nonlinearity,
recurrent_w_init=self.default_initializer,
output_nonlinearity=self.default_output_nonlinearity,
output_w_init=self.default_initializer,
state_include_action=True,
name='P3')
self.policy4 = CategoricalGRUPolicyWithModel(
env_spec=env.spec,
hidden_dim=4,
hidden_nonlinearity=self.default_hidden_nonlinearity,
hidden_w_init=self.default_initializer,
recurrent_nonlinearity=self.default_recurrent_nonlinearity,
recurrent_w_init=self.default_initializer,
output_nonlinearity=self.default_output_nonlinearity,
output_w_init=tf.constant_initializer(2),
state_include_action=True,
name='P4')
self.policy1.reset()
self.policy2.reset()
self.policy3.reset()
self.policy4.reset()
self.obs = [env.reset()]
self.obs = np.concatenate([self.obs for _ in range(self.time_step)],
axis=0)
self.obs_ph = tf.placeholder(tf.float32,
shape=(None, None,
env.observation_space.flat_dim))
self.action_ph = tf.placeholder(tf.float32,
shape=(None, None,
env.action_space.flat_dim))
self.dist1_sym = self.policy1.dist_info_sym(
obs_var=self.obs_ph,
state_info_vars={'prev_action': np.zeros((2, self.time_step, 1))},
name='p1_sym')
self.dist2_sym = self.policy2.dist_info_sym(
obs_var=self.obs_ph,
state_info_vars={'prev_action': np.zeros((2, self.time_step, 1))},
name='p2_sym')
self.dist3_sym = self.policy3.dist_info_sym(
obs_var=self.obs_ph,
state_info_vars={'prev_action': np.zeros((2, self.time_step, 1))},
name='p3_sym')
self.dist4_sym = self.policy4.dist_info_sym(
obs_var=self.obs_ph,
state_info_vars={'prev_action': np.zeros((2, self.time_step, 1))},
name='p4_sym')
def test_dist_info_sym_output(self):
# batch size = 2
dist1 = self.sess.run(self.dist1_sym,
feed_dict={self.obs_ph: [self.obs, self.obs]})
dist2 = self.sess.run(self.dist2_sym,
feed_dict={self.obs_ph: [self.obs, self.obs]})
dist3 = self.sess.run(self.dist3_sym,
feed_dict={self.obs_ph: [self.obs, self.obs]})
dist4 = self.sess.run(self.dist4_sym,
feed_dict={self.obs_ph: [self.obs, self.obs]})
assert np.array_equal(dist1['prob'], dist3['prob'])
assert np.array_equal(dist2['prob'], dist4['prob'])
@mock.patch('numpy.random.rand')
def test_get_action(self, mock_rand):
mock_rand.return_value = 0
action1, agent_info1 = self.policy1.get_action(self.obs)
action2, agent_info2 = self.policy2.get_action(self.obs)
action3, agent_info3 = self.policy3.get_action(self.obs)
action4, agent_info4 = self.policy4.get_action(self.obs)
assert action1 == action3
assert action2 == action4
assert np.array_equal(agent_info1['prob'], agent_info3['prob'])
assert np.array_equal(agent_info2['prob'], agent_info4['prob'])
actions1, agent_infos1 = self.policy1.get_actions([self.obs])
actions2, agent_infos2 = self.policy2.get_actions([self.obs])
actions3, agent_infos3 = self.policy3.get_actions([self.obs])
actions4, agent_infos4 = self.policy4.get_actions([self.obs])
assert np.array_equal(actions1, actions3)
assert np.array_equal(actions2, actions4)
assert np.array_equal(agent_infos1['prob'], agent_infos3['prob'])
assert np.array_equal(agent_infos2['prob'], agent_infos4['prob'])
def test_kl_sym(self):
kl_diff_sym1 = self.policy1.distribution.kl_sym(
self.dist1_sym, self.dist2_sym)
objective1 = tf.reduce_mean(kl_diff_sym1)
kl_func = tensor_utils.compile_function([self.obs_ph], objective1)
kl1 = kl_func([self.obs, self.obs])
kl_diff_sym2 = self.policy3.distribution.kl_sym(
self.dist3_sym, self.dist4_sym)
objective2 = tf.reduce_mean(kl_diff_sym2)
kl_func = tensor_utils.compile_function([self.obs_ph], objective2)
kl2 = kl_func([self.obs, self.obs])
assert np.array_equal(kl1, kl2)
def test_log_likehihood_sym(self):
log_prob_sym1 = self.policy1.distribution.log_likelihood_sym(
self.action_ph, self.dist1_sym)
log_prob_func = tensor_utils.compile_function(
[self.obs_ph, self.action_ph], log_prob_sym1)
log_prob1 = log_prob_func([self.obs, self.obs],
np.ones((2, self.time_step, 1)))
log_prob_sym2 = self.policy3.distribution.log_likelihood_sym(
self.action_ph, self.dist3_sym)
log_prob_func2 = tensor_utils.compile_function(
[self.obs_ph, self.action_ph], log_prob_sym2)
log_prob2 = log_prob_func2([self.obs, self.obs],
np.ones((2, self.time_step, 1)))
assert np.array_equal(log_prob1, log_prob2)
log_prob_sym1 = self.policy2.distribution.log_likelihood_sym(
self.action_ph, self.dist2_sym)
log_prob_func = tensor_utils.compile_function(
[self.obs_ph, self.action_ph], log_prob_sym1)
log_prob1 = log_prob_func([self.obs, self.obs],
np.ones((2, self.time_step, 1)))
log_prob_sym2 = self.policy4.distribution.log_likelihood_sym(
self.action_ph, self.dist4_sym)
log_prob_func2 = tensor_utils.compile_function(
[self.obs_ph, self.action_ph], log_prob_sym2)
log_prob2 = log_prob_func2([self.obs, self.obs],
np.ones((2, self.time_step, 1)))
assert np.array_equal(log_prob1, log_prob2)
def test_policy_entropy_sym(self):
entropy_sym1 = self.policy1.distribution.entropy_sym(
self.dist1_sym, name='entropy_sym1')
entropy_func = tensor_utils.compile_function([self.obs_ph],
entropy_sym1)
entropy1 = entropy_func([self.obs, self.obs])
entropy_sym2 = self.policy3.distribution.entropy_sym(
self.dist3_sym, name='entropy_sym1')
entropy_func = tensor_utils.compile_function([self.obs_ph],
entropy_sym2)
entropy2 = entropy_func([self.obs, self.obs])
assert np.array_equal(entropy1, entropy2)
def test_likelihood_ratio_sym(self):
likelihood_ratio_sym1 = self.policy1.distribution.likelihood_ratio_sym(
self.action_ph,
self.dist1_sym,
self.dist2_sym,
name='li_ratio_sym1')
likelihood_ratio_func = tensor_utils.compile_function(
[self.action_ph, self.obs_ph], likelihood_ratio_sym1)
likelihood_ratio1 = likelihood_ratio_func(np.ones((2, 1, 1)),
[self.obs, self.obs])
likelihood_ratio_sym2 = self.policy3.distribution.likelihood_ratio_sym(
self.action_ph,
self.dist3_sym,
self.dist4_sym,
name='li_ratio_sym2')
likelihood_ratio_func = tensor_utils.compile_function(
[self.action_ph, self.obs_ph], likelihood_ratio_sym2)
likelihood_ratio2 = likelihood_ratio_func(np.ones((2, 1, 1)),
[self.obs, self.obs])
assert np.array_equal(likelihood_ratio1, likelihood_ratio2)
def setUp(self):
super().setUp()
env = TfEnv(DummyDiscreteEnv(obs_dim=(1, ), action_dim=1))
self.default_initializer = tf.constant_initializer(1)
self.default_hidden_nonlinearity = tf.nn.tanh
self.default_recurrent_nonlinearity = tf.nn.sigmoid
self.default_output_nonlinearity = None
self.time_step = 1
self.policy1 = CategoricalGRUPolicy(
env_spec=env.spec,
hidden_dim=4,
hidden_nonlinearity=self.default_hidden_nonlinearity,
recurrent_nonlinearity=self.default_recurrent_nonlinearity,
recurrent_w_x_init=self.default_initializer,
recurrent_w_h_init=self.default_initializer,
output_nonlinearity=self.default_output_nonlinearity,
output_w_init=self.default_initializer,
state_include_action=True,
name='P1')
self.policy2 = CategoricalGRUPolicy(
env_spec=env.spec,
hidden_dim=4,
hidden_nonlinearity=self.default_hidden_nonlinearity,
recurrent_nonlinearity=self.default_recurrent_nonlinearity,
recurrent_w_x_init=self.default_initializer,
recurrent_w_h_init=self.default_initializer,
output_nonlinearity=self.default_output_nonlinearity,
output_w_init=tf.constant_initializer(2),
state_include_action=True,
name='P2')
self.sess.run(tf.global_variables_initializer())
self.policy3 = CategoricalGRUPolicyWithModel(
env_spec=env.spec,
hidden_dim=4,
hidden_nonlinearity=self.default_hidden_nonlinearity,
hidden_w_init=self.default_initializer,
recurrent_nonlinearity=self.default_recurrent_nonlinearity,
recurrent_w_init=self.default_initializer,
output_nonlinearity=self.default_output_nonlinearity,
output_w_init=self.default_initializer,
state_include_action=True,
name='P3')
self.policy4 = CategoricalGRUPolicyWithModel(
env_spec=env.spec,
hidden_dim=4,
hidden_nonlinearity=self.default_hidden_nonlinearity,
hidden_w_init=self.default_initializer,
recurrent_nonlinearity=self.default_recurrent_nonlinearity,
recurrent_w_init=self.default_initializer,
output_nonlinearity=self.default_output_nonlinearity,
output_w_init=tf.constant_initializer(2),
state_include_action=True,
name='P4')
self.policy1.reset()
self.policy2.reset()
self.policy3.reset()
self.policy4.reset()
self.obs = [env.reset()]
self.obs = np.concatenate([self.obs for _ in range(self.time_step)],
axis=0)
self.obs_ph = tf.placeholder(tf.float32,
shape=(None, None,
env.observation_space.flat_dim))
self.action_ph = tf.placeholder(tf.float32,
shape=(None, None,
env.action_space.flat_dim))
self.dist1_sym = self.policy1.dist_info_sym(
obs_var=self.obs_ph,
state_info_vars={'prev_action': np.zeros((2, self.time_step, 1))},
name='p1_sym')
self.dist2_sym = self.policy2.dist_info_sym(
obs_var=self.obs_ph,
state_info_vars={'prev_action': np.zeros((2, self.time_step, 1))},
name='p2_sym')
self.dist3_sym = self.policy3.dist_info_sym(
obs_var=self.obs_ph,
state_info_vars={'prev_action': np.zeros((2, self.time_step, 1))},
name='p3_sym')
self.dist4_sym = self.policy4.dist_info_sym(
obs_var=self.obs_ph,
state_info_vars={'prev_action': np.zeros((2, self.time_step, 1))},
name='p4_sym')
