def test_is_pickleable(self, obs_dim, action_dim):
"""Test if ContinuousMLPPolicy is pickleable"""
env = GymEnv(DummyBoxEnv(obs_dim=obs_dim, action_dim=action_dim))
policy = ContinuousMLPPolicy(env_spec=env.spec)
state_input = tf.compat.v1.placeholder(tf.float32,
shape=(None, np.prod(obs_dim)))
outputs = policy.build(state_input, name='policy')
env.reset()
obs = env.step(1).observation
with tf.compat.v1.variable_scope('ContinuousMLPPolicy', reuse=True):
bias = tf.compat.v1.get_variable('mlp/hidden_0/bias')
# assign it to all one
bias.load(tf.ones_like(bias).eval())
output1 = self.sess.run([outputs],
feed_dict={state_input: [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,
np.prod(obs_dim)))
outputs = policy_pickled.build(state_input, name='policy')
output2 = sess.run([outputs],
feed_dict={state_input: [obs.flatten()]})
assert np.array_equal(output1, output2)
def test_build(self, obs_dim, action_dim):
"""Test build method"""
env = GymEnv(DummyBoxEnv(obs_dim=obs_dim, action_dim=action_dim))
policy = ContinuousMLPPolicy(env_spec=env.spec)
env.reset()
obs = env.step(1).observation
obs_dim = env.spec.observation_space.flat_dim
state_input = tf.compat.v1.placeholder(tf.float32,
shape=(None, obs_dim))
action_sym = policy.build(state_input, name='action_sym')
action = self.sess.run(action_sym,
feed_dict={state_input: [obs.flatten()]})
action = policy.action_space.unflatten(action)
assert env.action_space.contains(action)
