da6b040e254710f0df907d838db51f4f2cc9bf93,social_rl/multiagent_tfagents/multiagent_ppo.py,MultiagentPPO,get_single_agent_specs,#MultiagentPPO#Any#Any#,217

Before Change

  def get_single_agent_specs(self, time_step_spec, action_spec):
    Get single agent version of environment specs to feed to baby agents.
    single_obs_spec = collections.OrderedDict()
    for k in time_step_spec.observation.keys():
      if k == "direction":
        shape = [1]
      elif k == "image":
        // Remove agent dimension
        shape = time_step_spec.observation[k].shape[1:]
      else:
        // Additional control fields like "reward", and "done" should not be sent
        // to individual agents
        continue
      single_obs_spec[k] = tensor_spec.BoundedTensorSpec(
          shape=shape, name=time_step_spec.observation[k].name,
          minimum=time_step_spec.observation[k].minimum,
          maximum=time_step_spec.observation[k].maximum,
          dtype=time_step_spec.observation[k].dtype)
    single_reward_spec = tensor_spec.TensorSpec(
        shape=(), dtype=time_step_spec.reward.dtype, name="reward")
    single_time_step_spec = ts_lib.TimeStep(
        time_step_spec.step_type, single_reward_spec,

After Change

          maximum=spec.maximum,
          dtype=spec.dtype)

    single_obs_spec = tf.nest.map_structure(make_single_agent_spec,
                                            time_step_spec.observation)
    single_reward_spec = tensor_spec.TensorSpec(
        shape=(), dtype=time_step_spec.reward.dtype, name="reward")
    single_time_step_spec = ts_lib.TimeStep(time_step_spec.step_type,
                                            single_reward_spec,

Instances