def test_get_action_state_include_action(self, obs_dim, action_dim,
hidden_dim):
env = TfEnv(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 + np.prod(action_dim)
],
name='obs')
policy = GaussianGRUPolicy(env_spec=env.spec,
hidden_dim=hidden_dim,
state_include_action=True)
policy.build(obs_var)
policy.reset()
obs = env.reset()
action, _ = policy.get_action(obs.flatten())
assert env.action_space.contains(action)
policy.reset()
actions, _ = policy.get_actions([obs.flatten()])
for action in actions:
assert env.action_space.contains(action)
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_is_pickleable(self):
env = TfEnv(DummyBoxEnv(obs_dim=(1, ), action_dim=(1, )))
obs_var = tf.compat.v1.placeholder(
tf.float32,
shape=[None, None, env.observation_space.flat_dim],
name='obs')
policy = GaussianGRUPolicy(env_spec=env.spec,
state_include_action=False)
policy.build(obs_var)
env.reset()
obs = env.reset()
with tf.compat.v1.variable_scope('GaussianGRUPolicy/GaussianGRUModel',
reuse=True):
param = tf.compat.v1.get_variable(
'dist_params/log_std_param/parameter')
# assign it to all one
param.load(tf.ones_like(param).eval())
output1 = self.sess.run(
[policy.distribution.loc,
policy.distribution.stddev()],
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)
obs_var = tf.compat.v1.placeholder(
tf.float32,
shape=[None, None, env.observation_space.flat_dim],
name='obs')
policy_pickled.build(obs_var)
# yapf: disable
output2 = sess.run(
[
policy_pickled.distribution.loc,
policy_pickled.distribution.stddev()
],
feed_dict={
policy_pickled.model.input: [[obs.flatten()],
[obs.flatten()]]
})
assert np.array_equal(output1, output2)
def test_build_state_include_action(self, obs_dim, action_dim, hidden_dim):
env = GymEnv(DummyBoxEnv(obs_dim=obs_dim, action_dim=action_dim))
policy = GaussianGRUPolicy(env_spec=env.spec,
hidden_dim=hidden_dim,
state_include_action=True)
policy.reset(do_resets=None)
obs = env.reset()[0]
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.loc],
feed_dict={state_input: [[concat_obs], [concat_obs]]})
output2 = self.sess.run(
[dist_sym2.loc],
feed_dict={state_input: [[concat_obs], [concat_obs]]})
assert np.array_equal(output1, output2)
def test_build_state_not_include_action(self, obs_dim, action_dim,
hidden_dim):
env = GarageEnv(DummyBoxEnv(obs_dim=obs_dim, action_dim=action_dim))
policy = GaussianGRUPolicy(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.loc],
feed_dict={policy.model.input: [[obs.flatten()], [obs.flatten()]]})
output2 = self.sess.run(
[dist_sym.loc],
feed_dict={state_input: [[obs.flatten()], [obs.flatten()]]})
assert np.array_equal(output1, output2)
def test_is_pickleable(self):
env = GymEnv(DummyBoxEnv(obs_dim=(1, ), action_dim=(1, )))
policy = GaussianGRUPolicy(env_spec=env.spec,
state_include_action=False)
env.reset()
obs = env.reset()[0]
with tf.compat.v1.variable_scope('GaussianGRUPolicy', reuse=True):
param = tf.compat.v1.get_variable(
'dist_params/log_std_param/parameter')
# assign it to all one
param.load(tf.ones_like(param).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.loc, dist_sym.stddev()],
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)
# yapf: disable
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
output2 = sess.run(
[
dist_sym.loc,
dist_sym.stddev()
],
feed_dict={
state_input: [[obs.flatten()], [obs.flatten()]]
})
assert np.array_equal(output1, output2)
