[RLlib] Attention Net prep PR #2: Smaller cleanups. (#12449)

rllib/evaluation/collectors/sample_collector.py

@@ -31,7 +31,8 @@ class _SampleCollector(metaclass=ABCMeta):
 
     @abstractmethod
     def add_init_obs(self, episode: MultiAgentEpisode, agent_id: AgentID,
-                     policy_id: PolicyID, init_obs: TensorType) -> None:
+                     policy_id: PolicyID, t: int,
+                     init_obs: TensorType) -> None:
         """Adds an initial obs (after reset) to this collector.
 
         Since the very first observation in an environment is collected w/o
@@ -48,6 +49,8 @@ def add_init_obs(self, episode: MultiAgentEpisode, agent_id: AgentID,
                 values for.
             env_id (EnvID): The environment index (in a vectorized setup).
             policy_id (PolicyID): Unique id for policy controlling the agent.
+            t (int): The time step (episode length - 1). The initial obs has
+                ts=-1(!), then an action/reward/next-obs at t=0, etc..
             init_obs (TensorType): Initial observation (after env.reset()).
 
         Examples:
@@ -172,9 +175,10 @@ def postprocess_episode(self,
                 MultiAgentBatch. Used for batch_mode=`complete_episodes`.
 
         Returns:
-            Any: An ID that can be used in `build_multi_agent_batch` to
-                retrieve the samples that have been postprocessed as a
-                ready-built MultiAgentBatch.
+            Optional[MultiAgentBatch]: If `build` is True, the
+                SampleBatch or MultiAgentBatch built from `episode` (either
+                just from that episde or from the `_PolicyCollectorGroup`
+                in the `episode.batch_builder` property).
         """
         raise NotImplementedError
 

rllib/evaluation/collectors/simple_list_collector.py

@@ -52,17 +52,19 @@ def __init__(self, shift_before: int = 0):
         # each time a (non-initial!) observation is added.
         self.count = 0
 
-    def add_init_obs(self, episode_id: EpisodeID, agent_id: AgentID,
-                     env_id: EnvID, init_obs: TensorType,
+    def add_init_obs(self, episode_id: EpisodeID, agent_index: int,
+                     env_id: EnvID, t: int, init_obs: TensorType,
                      view_requirements: Dict[str, ViewRequirement]) -> None:
         """Adds an initial observation (after reset) to the Agent's trajectory.
 
         Args:
             episode_id (EpisodeID): Unique ID for the episode we are adding the
                 initial observation for.
-            agent_id (AgentID): Unique ID for the agent we are adding the
-                initial observation for.
+            agent_index (int): Unique int index (starting from 0) for the agent
+                within its episode.
             env_id (EnvID): The environment index (in a vectorized setup).
+            t (int): The time step (episode length - 1). The initial obs has
+                ts=-1(!), then an action/reward/next-obs at t=0, etc..
             init_obs (TensorType): The initial observation tensor (after
             `env.reset()`).
             view_requirements (Dict[str, ViewRequirements])
@@ -72,10 +74,15 @@ def add_init_obs(self, episode_id: EpisodeID, agent_id: AgentID,
                 single_row={
                     SampleBatch.OBS: init_obs,
                     SampleBatch.EPS_ID: episode_id,
-                    SampleBatch.AGENT_INDEX: agent_id,
+                    SampleBatch.AGENT_INDEX: agent_index,
                     "env_id": env_id,
+                    "t": t,
                 })
         self.buffers[SampleBatch.OBS].append(init_obs)
+        self.buffers[SampleBatch.EPS_ID].append(episode_id)
+        self.buffers[SampleBatch.AGENT_INDEX].append(agent_index)
+        self.buffers["env_id"].append(env_id)
+        self.buffers["t"].append(t)
 
     def add_action_reward_next_obs(self, values: Dict[str, TensorType]) -> \
             None:
@@ -133,7 +140,7 @@ def build(self, view_requirements: Dict[str, ViewRequirement]) -> \
                 continue
             # OBS are already shifted by -1 (the initial obs starts one ts
             # before all other data columns).
-            shift = view_req.shift - \
+            shift = view_req.data_rel_pos - \
                 (1 if data_col == SampleBatch.OBS else 0)
             if data_col not in np_data:
                 np_data[data_col] = to_float_np_array(self.buffers[data_col])
@@ -187,7 +194,10 @@ def _build_buffers(self, single_row: Dict[str, TensorType]) -> None:
         for col, data in single_row.items():
             if col in self.buffers:
                 continue
-            shift = self.shift_before - (1 if col == SampleBatch.OBS else 0)
+            shift = self.shift_before - (1 if col in [
+                SampleBatch.OBS, SampleBatch.EPS_ID, SampleBatch.AGENT_INDEX,
+                "env_id", "t"
+            ] else 0)
             # Python primitive or dict (e.g. INFOs).
             if isinstance(data, (int, float, bool, str, dict)):
                 self.buffers[col] = [0 for _ in range(shift)]
@@ -360,7 +370,7 @@ def episode_step(self, episode_id: EpisodeID) -> None:
 
     @override(_SampleCollector)
     def add_init_obs(self, episode: MultiAgentEpisode, agent_id: AgentID,
-                     env_id: EnvID, policy_id: PolicyID,
+                     env_id: EnvID, policy_id: PolicyID, t: int,
                      init_obs: TensorType) -> None:
         # Make sure our mappings are up to date.
         agent_key = (episode.episode_id, agent_id)
@@ -378,8 +388,9 @@ def add_init_obs(self, episode: MultiAgentEpisode, agent_id: AgentID,
         self.agent_collectors[agent_key] = _AgentCollector()
         self.agent_collectors[agent_key].add_init_obs(
             episode_id=episode.episode_id,
-            agent_id=agent_id,
+            agent_index=episode._agent_index(agent_id),
             env_id=env_id,
+            t=t,
             init_obs=init_obs,
             view_requirements=view_reqs)
 
@@ -429,7 +440,7 @@ def get_inference_input_dict(self, policy_id: PolicyID) -> \
             # Create the batch of data from the different buffers.
             data_col = view_req.data_col or view_col
             time_indices = \
-                view_req.shift - (
+                view_req.data_rel_pos - (
                     1 if data_col in [SampleBatch.OBS, "t", "env_id",
                                       SampleBatch.EPS_ID,
                                       SampleBatch.AGENT_INDEX] else 0)

rllib/evaluation/rollout_worker.py

@@ -272,9 +272,11 @@ def __init__(
             output_creator (Callable[[IOContext], OutputWriter]): Function that
                 returns an OutputWriter object for saving generated
                 experiences.
-            remote_worker_envs (bool): If using num_envs > 1, whether to create
-                those new envs in remote processes instead of in the current
-                process. This adds overheads, but can make sense if your envs
+            remote_worker_envs (bool): If using num_envs_per_worker > 1,
+                whether to create those new envs in remote processes instead of
+                in the current process. This adds overheads, but can make sense
+                if your envs are expensive to step/reset (e.g., for StarCraft).
+                Use this cautiously, overheads are significant!
             remote_env_batch_wait_ms (float): Timeout that remote workers
                 are waiting when polling environments. 0 (continue when at
                 least one env is ready) is a reasonable default, but optimal

rllib/evaluation/sampler.py

@@ -1040,7 +1040,8 @@ def _process_observations_w_trajectory_view_api(
             # Record transition info if applicable.
             if last_observation is None:
                 _sample_collector.add_init_obs(episode, agent_id, env_id,
-                                               policy_id, filtered_obs)
+                                               policy_id, episode.length - 1,
+                                               filtered_obs)
             else:
                 # Add actions, rewards, next-obs to collectors.
                 values_dict = {
@@ -1158,7 +1159,8 @@ def _process_observations_w_trajectory_view_api(
 
                     # Add initial obs to buffer.
                     _sample_collector.add_init_obs(
-                        new_episode, agent_id, env_id, policy_id, filtered_obs)
+                        new_episode, agent_id, env_id, policy_id,
+                        new_episode.length - 1, filtered_obs)
 
                     item = PolicyEvalData(
                         env_id, agent_id, filtered_obs,

rllib/evaluation/tests/test_trajectory_view_api.py

@@ -59,7 +59,7 @@ def test_traj_view_normal_case(self):
                     assert view_req_policy[key].data_col is None
                 else:
                     assert view_req_policy[key].data_col == SampleBatch.OBS
-                    assert view_req_policy[key].shift == 1
+                    assert view_req_policy[key].data_rel_pos == 1
             rollout_worker = trainer.workers.local_worker()
             sample_batch = rollout_worker.sample()
             expected_count = \
@@ -99,18 +99,18 @@ def test_traj_view_lstm_prev_actions_and_rewards(self):
 
                 if key == SampleBatch.PREV_ACTIONS:
                     assert view_req_policy[key].data_col == SampleBatch.ACTIONS
-                    assert view_req_policy[key].shift == -1
+                    assert view_req_policy[key].data_rel_pos == -1
                 elif key == SampleBatch.PREV_REWARDS:
                     assert view_req_policy[key].data_col == SampleBatch.REWARDS
-                    assert view_req_policy[key].shift == -1
+                    assert view_req_policy[key].data_rel_pos == -1
                 elif key not in [
                         SampleBatch.NEXT_OBS, SampleBatch.PREV_ACTIONS,
                         SampleBatch.PREV_REWARDS
                 ]:
                     assert view_req_policy[key].data_col is None
                 else:
                     assert view_req_policy[key].data_col == SampleBatch.OBS
-                    assert view_req_policy[key].shift == 1
+                    assert view_req_policy[key].data_rel_pos == 1
             trainer.stop()
 
     def test_traj_view_simple_performance(self):

rllib/examples/policy/episode_env_aware_policy.py

@@ -28,22 +28,24 @@ class _fake_model:
             "t": ViewRequirement(),
             SampleBatch.OBS: ViewRequirement(),
             SampleBatch.PREV_ACTIONS: ViewRequirement(
-                SampleBatch.ACTIONS, space=self.action_space, shift=-1),
+                SampleBatch.ACTIONS, space=self.action_space, data_rel_pos=-1),
             SampleBatch.PREV_REWARDS: ViewRequirement(
-                SampleBatch.REWARDS, shift=-1),
+                SampleBatch.REWARDS, data_rel_pos=-1),
         }
         for i in range(2):
             self.model.inference_view_requirements["state_in_{}".format(i)] = \
                 ViewRequirement(
-                    "state_out_{}".format(i), shift=-1, space=self.state_space)
+                    "state_out_{}".format(i),
+                    data_rel_pos=-1,
+                    space=self.state_space)
             self.model.inference_view_requirements[
                 "state_out_{}".format(i)] = \
                 ViewRequirement(space=self.state_space)
 
         self.view_requirements = dict(
             **{
                 SampleBatch.NEXT_OBS: ViewRequirement(
-                    SampleBatch.OBS, shift=1),
+                    SampleBatch.OBS, data_rel_pos=1),
                 SampleBatch.ACTIONS: ViewRequirement(space=self.action_space),
                 SampleBatch.REWARDS: ViewRequirement(),
                 SampleBatch.DONES: ViewRequirement(),

rllib/examples/policy/rock_paper_scissors_dummies.py

@@ -16,7 +16,7 @@ def __init__(self, *args, **kwargs):
         self.view_requirements.update({
             "state_in_0": ViewRequirement(
                 "state_out_0",
-                shift=-1,
+                data_rel_pos=-1,
                 space=gym.spaces.Box(0, 100, shape=(), dtype=np.int32))
         })
 

rllib/models/modelv2.py

@@ -61,7 +61,8 @@ def __init__(self, obs_space: gym.spaces.Space,
         self.time_major = self.model_config.get("_time_major")
         # Basic view requirement for all models: Use the observation as input.
         self.inference_view_requirements = {
-            SampleBatch.OBS: ViewRequirement(shift=0, space=self.obs_space),
+            SampleBatch.OBS: ViewRequirement(
+                data_rel_pos=0, space=self.obs_space),
         }
 
     # TODO: (sven): Get rid of `get_initial_state` once Trajectory

rllib/models/tf/recurrent_net.py

@@ -178,10 +178,10 @@ def __init__(self, obs_space: gym.spaces.Space,
         if model_config["lstm_use_prev_action"]:
             self.inference_view_requirements[SampleBatch.PREV_ACTIONS] = \
                 ViewRequirement(SampleBatch.ACTIONS, space=self.action_space,
-                                shift=-1)
+                                data_rel_pos=-1)
         if model_config["lstm_use_prev_reward"]:
             self.inference_view_requirements[SampleBatch.PREV_REWARDS] = \
-                ViewRequirement(SampleBatch.REWARDS, shift=-1)
+                ViewRequirement(SampleBatch.REWARDS, data_rel_pos=-1)
 
     @override(RecurrentNetwork)
     def forward(self, input_dict: Dict[str, TensorType],

rllib/models/torch/recurrent_net.py

@@ -159,10 +159,10 @@ def __init__(self, obs_space: gym.spaces.Space,
         if model_config["lstm_use_prev_action"]:
             self.inference_view_requirements[SampleBatch.PREV_ACTIONS] = \
                 ViewRequirement(SampleBatch.ACTIONS, space=self.action_space,
-                                shift=-1)
+                                data_rel_pos=-1)
         if model_config["lstm_use_prev_reward"]:
             self.inference_view_requirements[SampleBatch.PREV_REWARDS] = \
-                ViewRequirement(SampleBatch.REWARDS, shift=-1)
+                ViewRequirement(SampleBatch.REWARDS, data_rel_pos=-1)
 
     @override(RecurrentNetwork)
     def forward(self, input_dict: Dict[str, TensorType],

rllib/policy/policy.py

@@ -564,13 +564,14 @@ def _get_default_view_requirements(self):
             SampleBatch.OBS: ViewRequirement(space=self.observation_space),
             SampleBatch.NEXT_OBS: ViewRequirement(
                 data_col=SampleBatch.OBS,
-                shift=1,
+                data_rel_pos=1,
                 space=self.observation_space),
             SampleBatch.ACTIONS: ViewRequirement(space=self.action_space),
             SampleBatch.REWARDS: ViewRequirement(),
             SampleBatch.DONES: ViewRequirement(),
             SampleBatch.INFOS: ViewRequirement(),
             SampleBatch.EPS_ID: ViewRequirement(),
+            SampleBatch.UNROLL_ID: ViewRequirement(),
             SampleBatch.AGENT_INDEX: ViewRequirement(),
             SampleBatch.UNROLL_ID: ViewRequirement(),
             "t": ViewRequirement(),
@@ -617,7 +618,7 @@ def _initialize_loss_from_dummy_batch(
         batch_for_postproc.count = self._dummy_batch.count
         postprocessed_batch = self.postprocess_trajectory(batch_for_postproc)
         if state_outs:
-            B = 4  # For RNNs, have B=2, T=[depends on sample_batch_size]
+            B = 4  # For RNNs, have B=4, T=[depends on sample_batch_size]
             # TODO: (sven) This hack will not work for attention net traj.
             #  view setup.
             i = 0
@@ -657,7 +658,8 @@ def _initialize_loss_from_dummy_batch(
                 # Tag those only needed for post-processing.
                 for key in batch_for_postproc.accessed_keys:
                     if key not in train_batch.accessed_keys and \
-                            key in self.view_requirements:
+                            key in self.view_requirements and \
+                            key not in self.model.inference_view_requirements:
                         self.view_requirements[key].used_for_training = False
                 # Remove those not needed at all (leave those that are needed
                 # by Sampler to properly execute sample collection).
@@ -680,18 +682,6 @@ def _initialize_loss_from_dummy_batch(
                                 "postprocessing function.".format(key))
                         else:
                             del self.view_requirements[key]
-            # Add those data_cols (again) that are missing and have
-            # dependencies by view_cols.
-            for key in list(self.view_requirements.keys()):
-                vr = self.view_requirements[key]
-                if vr.data_col is not None and \
-                        vr.data_col not in self.view_requirements:
-                    used_for_training = \
-                        vr.data_col in train_batch.accessed_keys
-                    self.view_requirements[vr.data_col] = \
-                        ViewRequirement(
-                            space=vr.space,
-                            used_for_training=used_for_training)
 
     def _get_dummy_batch_from_view_requirements(
             self, batch_size: int = 1) -> SampleBatch:
@@ -727,7 +717,7 @@ def _update_model_inference_view_requirements_from_init_state(self):
             model.inference_view_requirements["state_in_{}".format(i)] = \
                 ViewRequirement(
                     "state_out_{}".format(i),
-                    shift=-1,
+                    data_rel_pos=-1,
                     space=Box(-1.0, 1.0, shape=state.shape))
             model.inference_view_requirements["state_out_{}".format(i)] = \
                 ViewRequirement(space=Box(-1.0, 1.0, shape=state.shape))

rllib/policy/view_requirement.py

@@ -29,7 +29,7 @@ class ViewRequirement:
     def __init__(self,
                  data_col: Optional[str] = None,
                  space: gym.Space = None,
-                 shift: Union[int, List[int]] = 0,
+                 data_rel_pos: Union[int, List[int]] = 0,
                  used_for_training: bool = True):
         """Initializes a ViewRequirement object.
 
@@ -40,19 +40,20 @@ def __init__(self,
             space (gym.Space): The gym Space used in case we need to pad data
                 in inaccessible areas of the trajectory (t<0 or t>H).
                 Default: Simple box space, e.g. rewards.
-            shift (Union[int, List[int]]): Single shift value of list of
-                shift values to use relative to the underlying `data_col`.
+            data_rel_pos (Union[int, str, List[int]]): Single shift value or
+                list of relative positions to use (relative to the underlying
+                `data_col`).
                 Example: For a view column "prev_actions", you can set
-                `data_col="actions"` and `shift=-1`.
+                `data_col="actions"` and `data_rel_pos=-1`.
                 Example: For a view column "obs" in an Atari framestacking
                 fashion, you can set `data_col="obs"` and
-                `shift=[-3, -2, -1, 0]`.
+                `data_rel_pos=[-3, -2, -1, 0]`.
             used_for_training (bool): Whether the data will be used for
                 training. If False, the column will not be copied into the
                 final train batch.
         """
         self.data_col = data_col
         self.space = space or gym.spaces.Box(
             float("-inf"), float("inf"), shape=())
-        self.shift = shift
+        self.data_rel_pos = data_rel_pos
         self.used_for_training = used_for_training