Optionally save videos of policy evaluations

eval_experiments.py

@@ -6,7 +6,7 @@ import json
 
 activations = [torch.nn.Tanh, torch.nn.LeakyReLU]
 
-def main(method:str='expert', folder:str=None, locations=[(0,0)], skip_running=False):
+def main(method:str='expert', folder:str=None, locations=[(0,0)], skip_running=False, save_videos:bool=False, videos_folder:str='videos'):
     
     exclude_keys_from_policy_kwargs = {'learning_rate', 'learning_rate_decay', 'clip_ratio', 'iterations_per_epoch', 'option'}
     policy_kwargs = {}
@@ -50,7 +50,8 @@ def main(method:str='expert', folder:str=None, locations=[(0,0)], skip_running=F
                                 policy_file=policy_file, 
                                 policy_kwargs=policy_kwargs,
                                 env=env, 
-                                env_kwargs=env_kwargs)
+                                env_kwargs=env_kwargs,
+                                videos_folder=None if not save_videos else videos_folder)
         outfolder = os.path.dirname(outbase)
     else:
         locstr = 'loc_'+'_'.join([f'r{ro}t{tr}' for (ro,tr) in locations])

generate_videos.sh

@@ -0,0 +1,18 @@
+# can add --skip_running if you've already run the saved policies through the test environments and have appropriate
+# metrics in the out folder. Doing so will generate average metrics quickly.
+
+# Experiment A
+python -m eval_experiments 
+python -m eval_experiments --method idm --save_videos
+python -m eval_experiments --method bc --folder='test_policies/bc/expA' --save_videos
+python -m eval_experiments --method gail --folder='test_policies/gail/expA' --save_videos
+python -m eval_experiments --method hail --folder='test_policies/hail/expA' --save_videos
+python -m eval_experiments --method shail --folder='test_policies/shail/expA' --save_videos
+
+# Experiment B
+python -m eval_experiments --locations='[(0,4)]' 
+python -m eval_experiments --method idm --locations='[(0,4)]' --save_videos
+python -m eval_experiments --method bc --folder='test_policies/bc/expB' --locations='[(0,4)]' --save_videos
+python -m eval_experiments --method gail --folder='test_policies/gail/expB' --locations='[(0,4)]' --save_videos
+python -m eval_experiments --method hail --folder='test_policies/hail/expB' --locations='[(0,4)]' --save_videos
+python -m eval_experiments --method shail --folder='test_policies/shail/expB' --locations='[(0,4)]' --save_videos

src/__init__.py

@@ -1,2 +1 @@
-from src.data.expert_data import generate_expert_data, load_expert_data
+
-from src.data.data_utils import InteractionDatasetSingleAgent

src/eval_main.py

@@ -177,7 +177,8 @@ def evaluate_policy(locations:List[Tuple[int,int]],
     env_kwargs:dict,
     method: str,
     policy_file: str,
-    policy_kwargs:dict) -> List[Dict[str,list]]:
+    policy_kwargs:dict,
+    videos_folder: Optional[str] = None) -> List[Dict[str,list]]:
     """
     Evaluate policy on an incrementing agent environment at all locations. 
     Return metrics for that policy
@@ -230,7 +231,13 @@ def evaluate_policy(locations:List[Tuple[int,int]],
         policy = load_policy(method, policy_file, policy_kwargs, eval_env)
 
         # run policy on environment
-        policy_metrics[i] = evaluator.evaluate(policy)
+        policy_videos_folder = None
+        if videos_folder is not None:
+            policy_videos_folder = os.path.join(videos_folder, method, f'loc{iround}', f'track{track}')
+            os.makedirs(policy_videos_folder, exist_ok=True)
+        policy_metrics[i] = evaluator.evaluate(
+            policy, videos_folder=policy_videos_folder
+        )
     
     return policy_metrics
         
@@ -364,7 +371,8 @@ def eval_main(
     policy_kwargs: dict={},
     env: str='NRasterizedRouteIncrementingAgent', 
     env_kwargs: dict={},
-    seed: int=0):
+    seed: int=0,
+    videos_folder: Optional[str]=None):
     """
     Test a particular model at different testing locations/tracks and compute average metrics 
     over all files. 
@@ -410,7 +418,7 @@ def eval_main(
     else:
 
         # evaluate it on the given roundabouts
-        policy_metrics = evaluate_policy(locations, env, env_kwargs, method, policy_file, policy_kwargs)
+        policy_metrics = evaluate_policy(locations, env, env_kwargs, method, policy_file, policy_kwargs, videos_folder=videos_folder)
         smetrics = summary_metrics(policy_metrics)
         save_metrics(smetrics, outbase+'_summary.pkl')
         cmetrics = comparison_metrics(policy_metrics, expert_metrics, outbase=outbase)

src/evaluation/evaluation.py

@@ -9,6 +9,7 @@ from tqdm import tqdm
 from src.util.wrappers import IntersimpleTimeLimit
 from src.options.envs import OptionsEnv
 from src.safe_options.options import SafeOptionsEnv
+from src.evaluation.vec_env import CallbackWhenDoneVecEnv
 
 class IntersimpleEvaluation:
     """
@@ -80,7 +81,7 @@ class IntersimpleEvaluation:
         with open(filestr, 'wb') as f:
             pickle.dump(self._metrics, f)     
 
-    def evaluate(self, policy, filestr: Optional[str] = None) -> Dict[str, list]:
+    def evaluate(self, policy, filestr: Optional[str] = None, videos_folder: Optional[str] = None) -> Dict[str, list]:
         """
         Evaluate a policy on the incrementing agent evaluation environment
         
@@ -88,6 +89,8 @@ class IntersimpleEvaluation:
             policy (BaseClass.BaseAlgorithm): policy in which policy.predict(observation)[0] returns an action
             filestr (str): path-like string to dump metrics to or None
         """
+        self.videos_folder = videos_folder
+
         self.reset()
         if self.use_pbar:
             self.pbar = tqdm(total=self.n_episodes)
@@ -97,10 +100,11 @@ class IntersimpleEvaluation:
 
         evaluate_policy(
             policy, 
-            self.env,
+            self.env if self.videos_folder is None else CallbackWhenDoneVecEnv([lambda: self.env], self.done_callback),
             n_eval_episodes=self.n_episodes,
             callback=self.evaluate_options_policy_callback if self.is_options_env else self.evaluate_policy_callback,
-            return_episode_rewards=False
+            return_episode_rewards=False,
+            render=self.videos_folder is not None,
         )
         if self.use_pbar:
             self.pbar.close()
@@ -145,6 +149,13 @@ class IntersimpleEvaluation:
         
         if done and self.use_pbar:
             self.pbar.update(1)
+        
+    def done_callback(self, info):
+        if self.is_options_env:
+            info = info['ll']['infos'][0]
+        agent = info['agent']
+        filestr = os.path.join(self.videos_folder, f'agent{agent}')
+        self.env.close(filestr=filestr)
     
     def post_proc(self):
         """

src/evaluation/vec_env.py

@@ -0,0 +1,30 @@
+from stable_baselines3.common.vec_env import DummyVecEnv
+from stable_baselines3.common.vec_env.base_vec_env import VecEnvStepReturn
+from copy import deepcopy
+import numpy as np
+
+class CallbackWhenDoneVecEnv(DummyVecEnv):
+    """DummyVecEnv that calls `done_callback` before resetting the wrapped environment."""
+
+    def __init__(self, env_fns, done_callback):
+        assert len(env_fns) == 1 # for now
+        super().__init__(env_fns)
+        self.done_callback = done_callback
+
+    def step_wait(self) -> VecEnvStepReturn:
+        for env_idx in range(self.num_envs):
+            obs, self.buf_rews[env_idx], self.buf_dones[env_idx], self.buf_infos[env_idx] = self.envs[env_idx].step(
+                self.actions[env_idx]
+            )
+            if self.buf_dones[env_idx]:
+                # save final observation where user can get it, then reset
+                self.buf_infos[env_idx]["terminal_observation"] = obs
+
+                self.done_callback(deepcopy(self.buf_infos[env_idx]))
+                
+                obs = self.envs[env_idx].reset()
+            self._save_obs(env_idx, obs)
+        return (self._obs_from_buf(), np.copy(self.buf_rews), np.copy(self.buf_dones), deepcopy(self.buf_infos))
+    
+    def render(self, mode='post'):
+        super().render(mode)

src/options/envs.py

@@ -45,6 +45,7 @@ class OptionsEnv(Wrapper):
         self.options = options
         self.action_space = gym.spaces.Discrete(len(options))
         self.max_plan_length = max(t for _, t in options)
+        self.render_mode = None
     
     def plan(self, option):
         target_v, t = option
@@ -84,11 +85,14 @@ class OptionsEnv(Wrapper):
         
         n_steps = k + 1
         return observations, actions, rewards, env_done, plan_done, infos, n_steps
+    
+    def render(self, mode='post'):
+        self.render_mode = mode
 
-    def step(self, action, render_mode=None):
+    def step(self, action):
         a = int(action)
         assert a == action
-        ll_obs, ll_actions, ll_rewards, ll_env_done, ll_plan_done, ll_infos, ll_steps = self.execute_plan(self.last_obs, self.options[a], render_mode)
+        ll_obs, ll_actions, ll_rewards, ll_env_done, ll_plan_done, ll_infos, ll_steps = self.execute_plan(self.last_obs, self.options[a], self.render_mode)
         hl_obs = ll_obs[ll_steps]
         hl_reward = (ll_rewards * ~ll_plan_done).sum().item()
         hl_done = ll_env_done[ll_steps-1].item()

src/safe_options/options.py

@@ -225,8 +225,8 @@ class SafeOptionsEnv(OptionsEnv):
         }
         return obs
 
-    def step(self, action, render_mode=None):
+    def step(self, action):
-        obs, reward, done, info = super().step(action, render_mode)
+        obs, reward, done, info = super().step(action)
         obs = {
             'observation': obs,
             'safe_actions': self.safe_actions(),

src/util/wrappers.py

@@ -5,14 +5,23 @@ class Wrapper(gym.Wrapper):
     def __getattr__(self, name):
         return getattr(self.env, name)
 
+    def close(self, *args, **kwargs):
+        return self.env.close(*args, **kwargs)
+
 class TransformObservation(gym.wrappers.TransformObservation):
     def __getattr__(self, name):
         return getattr(self.env, name)
 
+    def close(self, *args, **kwargs):
+        return self.env.close(*args, **kwargs)
+
 class IntersimpleTimeLimit(gym.wrappers.TimeLimit):
     def __getattr__(self, name):
         return getattr(self.env, name)
 
+    def close(self, *args, **kwargs):
+        return self.env.close(*args, **kwargs)
+
 class CollisionPenaltyWrapper(Wrapper):
 
     def __init__(self, env, collision_distance, collision_penalty, *args, **kwargs):