[RLlib] Bring back BC and Marwil learning tests. (#21574)

release/rllib_tests/learning_tests/hard_learning_tests.yaml

@@ -136,37 +136,35 @@ appo-pongnoframeskip-v4:
 #        observation_filter: NoFilter
 #        report_length: 3
 
-# TODO: (sven) Fix all BC-dependent learning tests for cont. actions.
+ bc-halfcheetahbulletenv-v0:
-#  These seem quite hard to learn from the SAC-recorded HalfCheetahBulletEnv.
+    env: HalfCheetahBulletEnv-v0
-# bc-halfcheetahbulletenv-v0:
+    run: BC
-#    env: HalfCheetahBulletEnv-v0
+    pass_criteria:
-#    run: BC
+        evaluation/episode_reward_mean: 400.0
-#    pass_criteria:
+        timesteps_total: 10000000
-#        episode_reward_mean: 400.0
+    stop:
-#        timesteps_total: 10000000
+        time_total_s: 3600
-#    stop:
+    config:
-#        time_total_s: 3600
+        # Use input produced by expert SAC algo.
-#    config:
+        input: ["~/halfcheetah_expert_sac.zip"]
-#        # Use input produced by expert SAC algo.
+        actions_in_input_normalized: true
-#        input: ["~/halfcheetah_expert_sac.zip"]
-#        actions_in_input_normalized: true
 
-#        num_gpus: 1
+        num_gpus: 1
 
-#        model:
+        model:
-#            fcnet_activation: relu
+            fcnet_activation: relu
-#            fcnet_hiddens: [256, 256, 256]
+            fcnet_hiddens: [256, 256, 256]
 
-#        evaluation_num_workers: 1
+        evaluation_num_workers: 1
-#        evaluation_interval: 3
+        evaluation_interval: 3
-#        evaluation_config:
+        evaluation_config:
-#            input: sampler
+            input: sampler
 
 cql-halfcheetahbulletenv-v0:
     env: HalfCheetahBulletEnv-v0
     run: CQL
     pass_criteria:
-        episode_reward_mean: 400.0
+        evaluation/episode_reward_mean: 400.0
         timesteps_total: 10000000
     stop:
         time_total_s: 3600
@@ -363,31 +361,31 @@ impala-breakoutnoframeskip-v4:
         ]
         num_gpus: 1
 
-# marwil-halfcheetahbulletenv-v0:
+ marwil-halfcheetahbulletenv-v0:
-#    env: HalfCheetahBulletEnv-v0
+    env: HalfCheetahBulletEnv-v0
-#    run: MARWIL
+    run: MARWIL
-#    pass_criteria:
+    pass_criteria:
-#        episode_reward_mean: 400.0
+        evaluation/episode_reward_mean: 400.0
-#        timesteps_total: 10000000
+        timesteps_total: 10000000
-#    stop:
+    stop:
-#        time_total_s: 3600
+        time_total_s: 3600
-#    config:
+    config:
-#        # Use input produced by expert SAC algo.
+        # Use input produced by expert SAC algo.
-#        input: ["~/halfcheetah_expert_sac.zip"]
+        input: ["~/halfcheetah_expert_sac.zip"]
-#        actions_in_input_normalized: true
+        actions_in_input_normalized: true
-#        # Switch off input evaluation (data does not contain action probs).
+        # Switch off input evaluation (data does not contain action probs).
-#        input_evaluation: []
+        input_evaluation: []
 
-#        num_gpus: 1
+        num_gpus: 1
 
-#        model:
+        model:
-#            fcnet_activation: relu
+            fcnet_activation: relu
-#            fcnet_hiddens: [256, 256, 256]
+            fcnet_hiddens: [256, 256, 256]
 
-#        evaluation_num_workers: 1
+        evaluation_num_workers: 1
-#        evaluation_interval: 1
+        evaluation_interval: 1
-#        evaluation_config:
+        evaluation_config:
-#            input: sampler
+            input: sampler
 
 ppo-breakoutnoframeskip-v4:
     env: BreakoutNoFrameskip-v4

rllib/agents/tests/test_trainer.py

@@ -189,7 +189,7 @@ class TestTrainer(unittest.TestCase):
             # Eval results are not available at step 0.
             # But step 3 should still have it, even though no eval was
             # run during that step.
-            self.assertFalse("evaluation" in r0)
+            self.assertTrue("evaluation" in r0)
             self.assertTrue("evaluation" in r1)
             self.assertTrue("evaluation" in r2)
             self.assertTrue("evaluation" in r3)

rllib/agents/trainer.py

@@ -724,10 +724,19 @@ class Trainer(Trainable):
         self._episode_history = []
         self._episodes_to_be_collected = []
 
-        # Evaluation WorkerSet.
+        # Evaluation WorkerSet and metrics last returned by `self.evaluate()`.
         self.evaluation_workers: Optional[WorkerSet] = None
-        # Metrics most recently returned by `self.evaluate()`.
+        # Initialize common evaluation_metrics to nan, before they become
-        self.evaluation_metrics = {}
+        # available. We want to make sure the metrics are always present
+        # (although their values may be nan), so that Tune does not complain
+        # when we use these as stopping criteria.
+        self.evaluation_metrics = {
+            "evaluation": {
+                "episode_reward_max": np.nan,
+                "episode_reward_min": np.nan,
+                "episode_reward_mean": np.nan,
+            }
+        }
 
         super().__init__(config, logger_creator, remote_checkpoint_dir,
                          sync_function_tpl)

rllib/examples/parallel_evaluation_and_training.py

@@ -74,7 +74,7 @@ class AssertEvalCallback(DefaultCallbacks):
         # Make sure we always run exactly the given evaluation duration,
         # no matter what the other settings are (such as
         # `evaluation_num_workers` or `evaluation_parallel_to_training`).
-        if "evaluation" in result:
+        if "evaluation" in result and "hist_stats" in result["evaluation"]:
             hist_stats = result["evaluation"]["hist_stats"]
             # We count in episodes.
             if trainer.config["evaluation_duration_unit"] == "episodes":

rllib/utils/test_utils.py

@@ -607,6 +607,13 @@ def run_learning_tests_from_yaml(
 
     start_time = time.monotonic()
 
+    def should_check_eval(experiment):
+        # If we have evaluation workers, use their rewards.
+        # This is useful for offline learning tests, where
+        # we evaluate against an actual environment.
+        return experiment["config"].get("evaluation_interval",
+                                        None) is not None
+
     # Loop through all collected files and gather experiments.
     # Augment all by `torch` framework.
     for yaml_file in yaml_files:
@@ -637,11 +644,13 @@ def run_learning_tests_from_yaml(
                 # create its trainer and run a first iteration.
                 e["stop"]["time_total_s"] = 0
             else:
+                check_eval = should_check_eval(e)
+                episode_reward_key = ("episode_reward_mean" if not check_eval
+                                      else "evaluation/episode_reward_mean")
                 # We also stop early, once we reach the desired reward.
-                min_reward = e.get("pass_criteria",
+                min_reward = e.get("pass_criteria", {}).get(episode_reward_key)
-                                   {}).get("episode_reward_mean")
                 if min_reward is not None:
-                    e["stop"]["episode_reward_mean"] = min_reward
+                    e["stop"][episode_reward_key] = min_reward
 
             # Generate `checks` dict for all experiments
             # (tf, tf2 and/or torch).
@@ -723,11 +732,7 @@ def run_learning_tests_from_yaml(
                     trials_for_experiment.append(t)
             print(f" ... Trials: {trials_for_experiment}.")
 
-            # If we have evaluation workers, use their rewards.
+            check_eval = should_check_eval(experiments[experiment])
-            # This is useful for offline learning tests, where
-            # we evaluate against an actual environment.
-            check_eval = experiments[experiment]["config"].get(
-                "evaluation_interval", None) is not None
 
             # Error: Increase failure count and repeat.
             if any(t.status == "ERROR" for t in trials_for_experiment):