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)
def test_get_action(self, obs_dim, action_dim):
env = MetaRLEnv(DummyBoxEnv(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 = GaussianMLPPolicy(env_spec=env.spec)
policy.build(obs_var)
env.reset()
obs, _, _, _ = env.step(1)
action, _ = policy.get_action(obs.flatten())
assert env.action_space.contains(action)
actions, _ = policy.get_actions(
[obs.flatten(), obs.flatten(),
obs.flatten()])
for action in actions:
assert env.action_space.contains(action)
def test_is_pickleable(self, obs_dim, action_dim):
env = MetaRLEnv(DummyBoxEnv(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 = GaussianMLPPolicy(env_spec=env.spec)
policy.build(obs_var)
obs = env.reset()
with tf.compat.v1.variable_scope('GaussianMLPPolicy/GaussianMLPModel',
reuse=True):
bias = tf.compat.v1.get_variable(
'dist_params/mean_network/hidden_0/bias')
# assign it to all one
bias.load(tf.ones_like(bias).eval())
output1 = self.sess.run(
[policy.distribution.loc,
policy.distribution.stddev()],
feed_dict={policy.model.input: [[obs.flatten()]]})
p = pickle.dumps(policy)
with tf.compat.v1.Session(graph=tf.Graph()) as sess:
obs_var = tf.compat.v1.placeholder(
tf.float32,
shape=[None, None, env.observation_space.flat_dim],
name='obs')
policy_pickled = pickle.loads(p)
policy_pickled.build(obs_var)
output2 = sess.run(
[
policy_pickled.distribution.loc,
policy_pickled.distribution.stddev()
],
feed_dict={policy_pickled.model.input: [[obs.flatten()]]})
assert np.array_equal(output1, output2)