Add eval test

Emerge-Lab · Jan 28, 2024 · d59c271 · d59c271
1 parent 84f1fec
commit d59c271
Showing 1 changed file with 33 additions and 7 deletions.
diff --git a/evaluation/policy_evaluation.py b/evaluation/policy_evaluation.py
@@ -69,6 +69,7 @@ def evaluate_policy(
 
     # Run
     for _ in tqdm(range(num_episodes)):
+
         if traffic_files is not None:
             # Reset to a new scene
             obs_dict = env.reset(
@@ -77,7 +78,7 @@ def evaluate_policy(
             )
 
         else:
-            obs_dict = env.reset()
+            obs_dict = env.reset(use_av_only=use_av_only)
 
         agent_ids = list(obs_dict.keys())
         dead_agent_ids = []
@@ -176,7 +177,7 @@ def evaluate_policy(
                     off_road[agend_idx] += last_info_dicts[agent_id]["veh_edge_collision"] * 1
                     goal_achieved[agend_idx] += last_info_dicts[agent_id]["goal_achieved"] * 1
 
-                    logging.info(f"Goal achieved: {last_info_dicts[agent_id]['goal_achieved']}")
+                    logging.debug(f"Goal achieved: {last_info_dicts[agent_id]['goal_achieved']}")
 
                 # Get scene info
                 if scene_path_mapping is not None:
@@ -239,17 +240,42 @@ def evaluate_policy(
 
 
 if __name__ == "__main__":
+
+    # Logging
+    logger = logging.getLogger()
+    logger.setLevel('INFO')
+
+    # Load environment config
     env_config = load_config("env_config")
 
-    # Set data path to new scenes
+    # Set data path to NEW scenes (with is_av flag)
     env_config.data_path = "data_new/train_no_tl"
 
+    # Get all scene files
     train_file_paths = glob.glob(f"{env_config.data_path}" + "/tfrecord*")
     files = sorted([os.path.basename(file) for file in train_file_paths])
 
-    logging.info(f"Using {len(files)} scenes")
+    # Step through scene using expert-teleports
+    logging.info(f'Evaluating policy using EXPERT-TELEPORT mode {len(files)} new scenes...\n')
 
     df_expert_replay = evaluate_policy(
+        env_config=env_config,
+        controlled_agents=500,
+        data_path=env_config.data_path,
+        traffic_files=files,
+        mode="expert-replay",
+        select_from_k_scenes=1000, # Use all scenes
+        num_episodes=100,
+        use_av_only=True,
+    )
+
+    logging.info(f'--- Results: EXPERT-TELEPORT ---')
+    print(df_expert_replay[["goal_rate", "off_road", "veh_veh_collision"]].mean())
+
+
+    logging.info(f'Evaluating policy using EXPERT-TRAJECTORY ACTIONS {len(files)} new scenes...\n')
+
+    df_expert_traj = evaluate_policy(
         env_config=env_config,
         controlled_agents=500,
         data_path=env_config.data_path,
@@ -259,8 +285,8 @@ def evaluate_policy(
         num_episodes=100,
         use_av_only=True,
     )
+
+    logging.info(f'--- Results: EXPERT-TRAJECTORY ACTIONS ---')
+    print(df_expert_traj[["goal_rate", "off_road", "veh_veh_collision"]].mean())
 
-    print(df_expert_replay[["goal_rate", "off_road", "veh_veh_collision"]].mean())
 
-    # with open("invalid_train", "wb") as fp:   #Pickling
-    #     pickle.dump(inval_scenes, fp)