def _infer(self, **kwargs):
feed_dict = {
self.step_obs_ph: kwargs['obs_t'],
}
sess = tf.get_default_session()
ops = [self.action, self.log_policy, self.value]
return ActionOutput(*sess.run(ops, feed_dict=feed_dict))
def _infer(self, **kwargs):
sess = tf.get_default_session()
feed_dict = {
self.obs_t_ph: np.array([kwargs['obs_t']])
}
ops = [self.action, self.log_prob, self.value]
return ActionOutput(*sess.run(ops, feed_dict=feed_dict))
def _infer(self, **kwargs):
feed_dict = {
self.obs_t_ph: np.array([kwargs['obs_t']])
}
sess = tf.get_default_session()
ops = [self.action, self.value]
action, value = sess.run(ops, feed_dict=feed_dict)
return ActionOutput(action=action[0], log_prob=None, value=value[0])
def make_output(num_actions=4, batch_size=1, batch=False):
if batch:
action = np.random.random((batch_size, num_actions))
log_prob = np.random.random((batch_size,))
value = np.random.random((batch_size,))
else:
action = np.random.random((num_actions,))
log_prob = np.random.random()
value = np.random.random()
return ActionOutput(action, log_prob, value)
def test_properties(self):
output = ActionOutput('action', 'log_prob', 'value')
self.assertEqual(output.action, 'action')
self.assertEqual(output.value, 'value')
self.assertEqual(output.log_prob, 'log_prob')