def actions_for(self,
observations,
n_action_samples=1,
reuse=tf.AUTO_REUSE):
n_state_samples = tf.shape(observations)[0]
if n_action_samples > 1:
observations = observations[:, None, :]
latent_shape = (n_state_samples, n_action_samples,
self._action_dim)
else:
latent_shape = (n_state_samples, self._action_dim)
latents = tf.random_normal(latent_shape)
with tf.variable_scope('policy', reuse=reuse):
raw_actions = feedforward_net(
observations,
latents,
layer_sizes=self._layer_sizes,
activation_fn=tf.nn.relu,
output_nonlinearity=None)
return tf.tanh(raw_actions) if self._squash else raw_actions
single_latent = np.random.normal(size=(1, action_dim))
raw_action = sess.run(raw_actions, feed_dict={observation_ph: observation,
latent_ph: single_latent})
tan_action = sess.run(tan_actions, feed_dict={observation_ph: observation,
latent_ph: single_latent})
return tan_action if squash else raw_action
# forward computation graph of stochastic policy network
with tf.variable_scope('policy', reuse=False):
observation_ph = tf.placeholder(tf.float32, shape=[None, observation_dim])
latent_ph = tf.placeholder(tf.float32, shape=[None, action_dim])
raw_actions = feedforward_net((observation_ph, latent_ph), layer_sizes=layer_size,
activation_fn=tf.nn.relu, output_nonlinearity=None)
tan_actions = tf.tanh(raw_actions)
saver = tf.train.Saver()
with tf.Session() as sess:
# restore the trained model
ckpt = tf.train.get_checkpoint_state(SHARED_PARAMS['model_save_path'])
# saver.restore(sess, ckpt.model_checkpoint_path)
saver.restore(sess, SHARED_PARAMS['model_save_path']+'model-465.ckpt')
# TODO: roll-out with the trained policy
observation = np.expand_dims(env.reset(), axis=0) # s_0, shape[1, 18]
for t in range(ENV_PARAMS['max_path_length']):