def test_is_pickleable(self): env = GarageEnv(DummyDiscreteEnv(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 = CategoricalLSTMPolicy(env_spec=env.spec, state_include_action=False) policy.build(obs_var) policy.reset() obs = env.reset() policy.model._lstm_cell.weights[0].load( tf.ones_like(policy.model._lstm_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) obs_var = tf.compat.v1.placeholder( tf.float32, shape=[None, None, env.observation_space.flat_dim], name='obs') policy_pickled.build(obs_var) output2 = sess.run([policy_pickled.distribution.probs], feed_dict={ policy_pickled.model.input: [[obs.flatten()], [obs.flatten()]] }) # noqa: E126 assert np.array_equal(output1, output2)
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_is_pickleable(self): env = GarageEnv(DummyDiscreteEnv(obs_dim=(1, ), action_dim=1)) policy = CategoricalLSTMPolicy(env_spec=env.spec, state_include_action=False) policy.reset() 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 policy._lstm_cell.weights[0].load( tf.ones_like(policy._lstm_cell.weights[0]).eval()) 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_pickled.input_dim)) dist_sym = policy_pickled.build(state_input, name='dist_sym').dist output2 = sess.run( [dist_sym.probs], feed_dict={state_input: [[obs.flatten()], [obs.flatten()]]}) # noqa: E126 assert np.array_equal(output1, output2)
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 = CategoricalLSTMPolicy(env_spec=env.spec, hidden_dim=hidden_dim, state_include_action=False) policy.build(obs_var) policy.reset() 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_build_state_not_include_action(self, obs_dim, action_dim, hidden_dim): env = GymEnv(DummyDiscreteEnv(obs_dim=obs_dim, action_dim=action_dim)) policy = CategoricalLSTMPolicy(env_spec=env.spec, hidden_dim=hidden_dim, state_include_action=False) 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 output1 = self.sess.run( [dist_sym.probs], feed_dict={state_input: [[obs.flatten()], [obs.flatten()]]}) output2 = self.sess.run( [dist_sym2.probs], feed_dict={state_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 = GarageEnv( DummyDiscreteEnv(obs_dim=obs_dim, action_dim=action_dim)) policy = CategoricalLSTMPolicy(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 concat_obs = np.concatenate([obs.flatten(), np.zeros(action_dim)]) output1 = self.sess.run( [policy.distribution.probs], feed_dict={policy.model.input: [[concat_obs], [concat_obs]]}) output2 = self.sess.run( [dist_sym.probs], feed_dict={state_input: [[concat_obs], [concat_obs]]}) assert np.array_equal(output1, output2)