Merge branch 'main' of https://github.com/sisl/InteractionImitation

ogail-ppo-options-setobs2.py

@@ -0,0 +1,174 @@
+# %%
+import os
+
+import gym
+from src.safe_options.options import gail_ppo, Buffer
+from src.core.value import SetValue
+from src.safe_options.policy import SetMaskedDiscretePolicy
+from src.core.discriminator import DeepsetDiscriminator
+import torch
+
+from intersim.envs import IntersimpleLidarFlatRandom
+from intersim.envs.intersimple import speed_reward
+import functools
+from src.util.wrappers import CollisionPenaltyWrapper, TransformObservation, Setobs
+import numpy as np
+from src.safe_options.options import SafeOptionsEnv
+from torch.utils.tensorboard import SummaryWriter
+from ray import tune
+
+DIR = os.path.dirname(os.path.abspath(__file__))
+option_list = [[(vel, time) for vel in [0, 1, 2, 4, 6, 8, 10] for time in [5]],
+                [(vel, time) for vel in [0, 1, 2, 5, 7.5, 10] for time in [5, 20]],
+                [(vel, time) for vel in [0, 1, 2, 4, 6, 8, 10] for time in [5, 10, 20]], # was the best in training with single hidden layer, but very slow
+                [(vel, time) for vel in [0, 1, 2, 5, 7.5, 10] for time in [5, 20, 40]], 
+                [(vel, time) for vel in [0, 2, 5, 10] for time in [5, 10, 20]],
+                [(vel, time) for vel in [0, 3, 10] for time in [5, 20, 40]]
+]
+
+obs_min = np.array([
+    [-1000, -1000, 0, -np.pi, -1e-1, 0.],
+    [0, -np.pi, -20, -20, -np.pi, -1e-1],
+    [0, -np.pi, -20, -20, -np.pi, -1e-1],
+    [0, -np.pi, -20, -20, -np.pi, -1e-1],
+    [0, -np.pi, -20, -20, -np.pi, -1e-1],
+    [0, -np.pi, -20, -20, -np.pi, -1e-1],
+]).reshape(-1)
+
+obs_max = np.array([
+    [1000, 1000, 20, np.pi, 1e-1, 0.],
+    [50, np.pi, 20, 20, np.pi, 1e-1],
+    [50, np.pi, 20, 20, np.pi, 1e-1],
+    [50, np.pi, 20, 20, np.pi, 1e-1],
+    [50, np.pi, 20, 20, np.pi, 1e-1],
+    [50, np.pi, 20, 20, np.pi, 1e-1],
+]).reshape(-1)
+
+def training_function(config):
+    np.random.seed(config['seed'])
+    torch.manual_seed(config['seed'])
+
+    envs = sum([[SafeOptionsEnv(Setobs(
+        TransformObservation(CollisionPenaltyWrapper(IntersimpleLidarFlatRandom(
+            n_rays=5,
+            reward=functools.partial(
+                speed_reward,
+                collision_penalty=0
+            ),
+            stop_on_collision=config['stop_on_collision'], track=track,
+        ), collision_distance=6, collision_penalty=100), lambda obs: (obs - obs_min) / (obs_max - obs_min + 1e-10))
+    ), options=option_list[config['policy']['option']], safe_actions_collision_method=None, 
+    abort_unsafe_collision_method=None) for _ in range(20)] for track in range(4)],[])
+
+    env_fn = lambda i: envs[i]
+
+    policy = SetMaskedDiscretePolicy(env_fn(0).action_space.n, 
+        n_hidden_layers=config['policy']['n_hidden_layers'],
+        hidden_layer_size=config['policy']['hidden_layer_size'],
+        activation=config['policy']['activation'] ) # config net architecture
+    pi_opt = torch.optim.Adam(policy.parameters(), lr=config['policy']['learning_rate'])
+    pi_lr_scheduler = torch.optim.lr_scheduler.ExponentialLR(pi_opt, gamma=config['policy']['learning_rate_decay'])
+
+    value = SetValue() # config net architecture
+    v_opt = torch.optim.Adam(value.parameters(), lr=config['value']['learning_rate'])
+
+    discriminator = DeepsetDiscriminator() # config net architecture
+    disc_opt = torch.optim.Adam(discriminator.parameters(), lr=config['discriminator']['learning_rate'], weight_decay=config['discriminator']['weight_decay'])
+
+    expert_data = [
+        torch.load(os.path.join(DIR, 'intersimple-expert-data-setobs2-loc0-track0.pt')),
+        torch.load(os.path.join(DIR, 'intersimple-expert-data-setobs2-loc0-track1.pt')),
+        torch.load(os.path.join(DIR, 'intersimple-expert-data-setobs2-loc0-track2.pt')),
+        torch.load(os.path.join(DIR, 'intersimple-expert-data-setobs2-loc0-track3.pt')),
+    ]
+    d0 = [d[0] for d in expert_data]
+    d1 = [d[1] for d in expert_data]
+    d2 = [d[2] for d in expert_data]
+    d3 = [d[3] for d in expert_data]
+    expert_data = (torch.cat(d0), torch.cat(d1), torch.cat(d2), torch.cat(d3))
+    expert_data = Buffer(*expert_data)
+
+    def callback(info):
+        tune.report(gen_mean_reward_per_episode=info['gen/mean_reward_per_episode'],
+                    disc_mean_reward_per_episode=info['disc/mean_reward_per_episode'], 
+                    mean_episode_length=info['gen/mean_episode_length'])
+        
+        # save model checkpoints
+        ep = info['epoch'] + 1
+        if (ep % 25 == 0):
+            torch.save(info['policy'].state_dict(), f'policy_epoch{ep}.pt')
+
+    value, policy = gail_ppo(
+        env_fn=env_fn,
+        expert_data=expert_data,
+        discriminator=discriminator,
+        disc_opt=disc_opt,
+        disc_iters=config['discriminator']['iterations_per_epoch'],
+        policy=policy,
+        value=value,
+        v_opt=v_opt,
+        v_iters=config['value']['iterations_per_epoch'],
+        epochs=200,
+        rollout_episodes=60,
+        rollout_steps=60,
+        gamma=0.99,
+        gae_lambda=0.9,
+        clip_ratio=config['policy']['clip_ratio'],
+        pi_opt=pi_opt,
+        pi_iters=config['policy']['iterations_per_epoch'],
+        logger=SummaryWriter(comment='sgail-ppo-options-setobs2'),
+        callback=callback,
+        lr_schedulers=[pi_lr_scheduler],
+    )
+
+    # save model
+    torch.save(policy.state_dict(), 'policy_final.pt')
+
+analysis = tune.run(
+    training_function,
+    config={
+        'stop_on_collision': tune.grid_search([True, False]),
+        'policy': {
+            'learning_rate': 3e-4, # tune.grid_search([3e-4]),
+            'learning_rate_decay': 1.0, #tune.grid_search([1.0]),
+            'clip_ratio': 0.2, #tune.grid_search([0.2]),
+            'iterations_per_epoch': 100, #tune.grid_search([100]),
+            'hidden_layer_size': tune.grid_search([10, 20, 40]),
+            'n_hidden_layers': tune.grid_search([2, 3, 4]),
+            'activation':tune.grid_search([torch.nn.LeakyReLU, torch.nn.Tanh]),
+            'option': tune.grid_search(list(range(len(option_list))))
+        },
+        'value': {
+            'learning_rate': 1e-3, # tune.grid_search([1e-3]),
+            'iterations_per_epoch': 1000, #tune.grid_search([1000]),
+        },
+        'discriminator': {
+            'learning_rate': 1e-3, #tune.grid_search([1e-3]),
+            'weight_decay': 1e-4, #tune.grid_search([1e-4]),
+            'iterations_per_epoch': 100, #tune.grid_search([100]),
+        },
+        'seed': 0,
+    }
+)
+
+print('Best config: ', analysis.get_best_config(metric='gen_mean_reward_per_episode', mode='max'))
+
+# %%
+# policy = SetMaskedDiscretePolicy(env_fn(0).action_space.n)
+# policy(torch.zeros(env_fn(0).observation_space['observation'].shape), torch.zeros(env_fn(0).observation_space['safe_actions'].shape))
+# policy.load_state_dict(torch.load('sgail-ppo-options-setobs2.pt'))
+
+# env = env_fn(0)
+# obs = env.reset()
+# env.render(mode='post')
+# for i in range(300):
+#     action = policy.sample(policy(
+#         torch.tensor(obs['observation'], dtype=torch.float32),
+#         torch.tensor(obs['safe_actions'], dtype=torch.float32),
+#     ))
+#     obs, reward, done, _ = env.step(action, render_mode='post')
+#     print('step', i, 'reward', reward, 'safe actions', obs['safe_actions'])
+#     if done:
+#         break
+# env.close()
+# %%

requirements.txt

@@ -8,3 +8,5 @@ ray[tune]
 hyperopt
 psutil
 fire
+stable_baselines3
+tensorboard

sgail-ppo-options-setobs2.py

@@ -6,7 +6,8 @@ from src.safe_options.options import gail_ppo, Buffer
 from src.core.value import SetValue
 from src.safe_options.policy import SetMaskedDiscretePolicy
 from src.core.discriminator import DeepsetDiscriminator
-import torch.optim
+import torch
+
 from intersim.envs import IntersimpleLidarFlatRandom
 from intersim.envs.intersimple import speed_reward
 import functools
@@ -18,8 +19,15 @@ from ray import tune
 from datetime import datetime
 import json
 
-def training_function(config):
 DIR = os.path.dirname(os.path.abspath(__file__))
+option_list = [[(vel, time) for vel in [0, 1, 2, 4, 6, 8, 10] for time in [5]],
+                [(vel, time) for vel in [0, 1, 2, 5, 7.5, 10] for time in [5, 20]],
+                [(vel, time) for vel in [0, 1, 2, 4, 6, 8, 10] for time in [5, 10, 20]], # was the best in training with single hidden layer, but very slow
+                [(vel, time) for vel in [0, 1, 2, 5, 7.5, 10] for time in [5, 20, 40]], 
+                [(vel, time) for vel in [0, 2, 5, 10] for time in [5, 10, 20]],
+                [(vel, time) for vel in [0, 3, 10] for time in [5, 20, 40]]
+]
+
 obs_min = np.array([
     [-1000, -1000, 0, -np.pi, -1e-1, 0.],
     [0, -np.pi, -20, -20, -np.pi, -1e-1],
@@ -38,7 +46,11 @@ def training_function(config):
     [50, np.pi, 20, 20, np.pi, 1e-1],
 ]).reshape(-1)
 
-    envs = [SafeOptionsEnv(Setobs(
+def training_function(config):
+    np.random.seed(config['seed'])
+    torch.manual_seed(config['seed'])
+
+    envs = sum([[SafeOptionsEnv(Setobs(
         TransformObservation(CollisionPenaltyWrapper(IntersimpleLidarFlatRandom(
             n_rays=5,
             reward=functools.partial(
@@ -46,15 +58,18 @@ def training_function(config):
                 collision_penalty=0
             ),
             check_collisions=True,
-            stop_on_collision=config['env']['stop_on_collision'],
+            stop_on_collision=config['env']['stop_on_collision'], track=track,
         ), collision_distance=6, collision_penalty=100), lambda obs: (obs - obs_min) / (obs_max - obs_min + 1e-10))
-    ), options=[(0, 5), (1, 5), (2, 5), (4, 5), (6, 5), (8, 5), (10, 5)],
+    ), options=option_list[config['policy']['option']],
         safe_actions_collision_method=config['env']['safe_actions_collision_method'], 
-        abort_unsafe_collision_method=config['env']['abort_unsafe_collision_method']) for _ in range(60)]
+        abort_unsafe_collision_method=config['env']['abort_unsafe_collision_method']) for _ in range(20)] for track in range(4)],[])
 
     env_fn = lambda i: envs[i]
 
-    policy = SetMaskedDiscretePolicy(env_fn(0).action_space.n, hidden_layer_size=config['policy']['hidden_layer_size']) # config net architecture
+    policy = SetMaskedDiscretePolicy(env_fn(0).action_space.n, 
+        n_hidden_layers=config['policy']['n_hidden_layers'],
+        hidden_layer_size=config['policy']['hidden_layer_size'],
+        activation=config['policy']['activation'] ) # config net architecture
     pi_opt = torch.optim.Adam(policy.parameters(), lr=config['policy']['learning_rate'])
     pi_lr_scheduler = torch.optim.lr_scheduler.ExponentialLR(pi_opt, gamma=config['policy']['learning_rate_decay'])
 
@@ -77,16 +92,20 @@ def training_function(config):
     expert_data = (torch.cat(d0), torch.cat(d1), torch.cat(d2), torch.cat(d3))
     expert_data = Buffer(*expert_data)
 
-    folder = str(datetime.now())
+    run_folder = str(datetime.now())
-    os.mkdir(os.path.join(DIR, folder))
+    os.mkdir(os.path.join(DIR, run_folder))
-    with open(os.path.join(DIR, folder, 'config.json'), 'w') as f:
+    with open(os.path.join(DIR, run_folder, 'config.json'), 'w') as f:
         json.dump(config, f, indent=4)
 
     def callback(info):
-        tune.report(gen_mean_reward_per_episode=info['gen/mean_reward_per_episode'])
+        tune.report(gen_mean_reward_per_episode=info['gen/mean_reward_per_episode'],
-        if not info['epoch'] % 10:
+                    disc_mean_reward_per_episode=info['disc/mean_reward_per_episode'], 
-            torch.save(policy.state_dict(), os.path.join(DIR, folder, f'sgail-ppo-options-setobs2-{info["epoch"]}.pt'))
+                    mean_episode_length=info['gen/mean_episode_length'])
-            torch.save(value.state_dict(), os.path.join(DIR, folder, f'sgail-ppo-options-setobs2-value-{info["epoch"]}.pt'))
+        
+        # save model checkpoints
+        ep = info['epoch'] + 1
+        if (ep % 25 == 0):
+            torch.save(info['policy'].state_dict(), os.path.join(DIR, run_folder, f'policy_epoch{ep}.pt'))
 
     value, policy = gail_ppo(
         env_fn=env_fn,
@@ -111,6 +130,9 @@ def training_function(config):
         lr_schedulers=[pi_lr_scheduler],
     )
 
+    # save model
+    torch.save(policy.state_dict(), 'policy_final.pt')
+
 analysis = tune.run(
     training_function,
     config={
@@ -120,21 +142,25 @@ analysis = tune.run(
             'abort_unsafe_collision_method': 'circle',
         },
         'policy': {
-            'learning_rate': tune.grid_search([3e-4]),
+            'learning_rate': 3e-4, # tune.grid_search([3e-4]),
-            'learning_rate_decay': tune.grid_search([1.0]),
+            'learning_rate_decay': 1.0, #tune.grid_search([1.0]),
-            'clip_ratio': tune.grid_search([0.2]),
+            'clip_ratio': 0.2, #tune.grid_search([0.2]),
-            'iterations_per_epoch': tune.grid_search([100]),
+            'iterations_per_epoch': 100, #tune.grid_search([100]),
-            'hidden_layer_size': tune.grid_search([25])
+            'hidden_layer_size': tune.grid_search([10, 20, 40]),
+            'n_hidden_layers': tune.grid_search([2, 3, 4]),
+            'activation':tune.grid_search([torch.nn.LeakyReLU, torch.nn.Tanh]),
+            'option': tune.grid_search(list(range(len(option_list))))
         },
         'value': {
-            'learning_rate': tune.grid_search([1e-3]),
+            'learning_rate': 1e-3, # tune.grid_search([1e-3]),
-            'iterations_per_epoch': tune.grid_search([1000]),
+            'iterations_per_epoch': 1000, #tune.grid_search([1000]),
         },
         'discriminator': {
-            'learning_rate': tune.grid_search([1e-3]),
+            'learning_rate': 1e-3, #tune.grid_search([1e-3]),
-            'weight_decay': tune.grid_search([1e-4]),
+            'weight_decay': 1e-4, #tune.grid_search([1e-4]),
-            'iterations_per_epoch': tune.grid_search([500]),
+            'iterations_per_epoch': 100, #tune.grid_search([100]),
-        }
+        },
+        'seed': 0,
     }
 )
 

src/core/policy.py

@@ -36,30 +36,31 @@ class BasePolicy(nn.Module):
 
 class Policy(BasePolicy):
     
-    def __init__(self, *args, **kwargs):
+    def __init__(self, *args, hidden_layer_size=50, n_hidden_layers=2, activation=nn.Tanh, **kwargs):
         super().__init__(*args, **kwargs)
-        self.nn = nn.Sequential(
+        layers = sum([[nn.LazyLinear(hidden_layer_size), 
-            nn.LazyLinear(50),
+            activation()] for _ in range(n_hidden_layers)],[])
-            nn.Tanh(),
+        self.nn = nn.Sequential(*layers, nn.LazyLinear(2 *self.action_dim))
-            nn.LazyLinear(50),
+        
-            nn.Tanh(),
+        # old
-            nn.LazyLinear(2 * self.action_dim),
+        # self.nn = nn.Sequential(
-        )
+        #    nn.LazyLinear(50),
+        #    nn.Tanh(),
+        #    nn.LazyLinear(50),
+        #    nn.Tanh(),
+        #    nn.LazyLinear(2 * self.action_dim),
+        #)
 
     def forward(self, states):
         return self.nn(states)
 
 class DiscretePolicy(BasePolicy):
 
-    def __init__(self, *args, hidden_layer_size=50, **kwargs):
+    def __init__(self, *args, hidden_layer_size=50, n_hidden_layers=2, activation=nn.Tanh, **kwargs):
         super().__init__(*args, **kwargs)
-        self.nn = nn.Sequential(
+        layers = sum([[nn.LazyLinear(hidden_layer_size), 
-            nn.LazyLinear(hidden_layer_size),
+            activation()] for _ in range(n_hidden_layers)],[])
-            nn.Tanh(),
+        self.nn = nn.Sequential(*layers, nn.LazyLinear(self.action_dim))
-            nn.LazyLinear(hidden_layer_size),
-            nn.Tanh(),
-            nn.LazyLinear(self.action_dim),
-        )
 
     def forward(self, states):
         return self.nn(states)

src/eval_main.py

@@ -37,21 +37,22 @@ def load_policy(method:str,
     Returns:
         policy (Optional[BaseAlgorithm]): the policy to evaluate
     """
+    ml = torch.device('cpu') if not torch.cuda.is_available() else None
     if method == 'idm':
         policy = IDMRulePolicy(env, **policy_kwargs)
     elif method == 'bc':
         policy = SetPolicy(env.action_space.shape[-1])
-        policy.load_state_dict(torch.load(policy_file))
+        policy.load_state_dict(torch.load(policy_file, map_location=ml))
         policy.eval()
     elif method == 'gail':
         policy = SetPolicy(env.action_space.shape[-1])
         policy(torch.zeros(env.observation_space.shape))
         policy = ReparamPolicy(policy)
-        policy.load_state_dict(torch.load(policy_file))
+        policy.load_state_dict(torch.load(policy_file, map_location=ml))
         policy.eval()
     elif method == 'gail-ppo':
         policy = SetPolicy(env.action_space.shape[-1])
-        policy.load_state_dict(torch.load(policy_file))
+        policy.load_state_dict(torch.load(policy_file, map_location=ml))
         policy.eval()
     elif method == 'rail':
         raise NotImplementedError
@@ -59,11 +60,11 @@ def load_policy(method:str,
         policy = SetDiscretePolicy(env.action_space.n)
         policy(torch.zeros(env.observation_space.shape))
         policy = ReparamPolicy(policy)
-        policy.load_state_dict(torch.load(policy_file))
+        policy.load_state_dict(torch.load(policy_file, map_location=ml))
         policy.eval()
     elif method == 'ogail-ppo':
         policy = SetDiscretePolicy(env.action_space.n)
-        policy.load_state_dict(torch.load(policy_file))
+        policy.load_state_dict(torch.load(policy_file, map_location=ml))
         policy.eval()
     elif method == 'sgail':
         policy = SetMaskedDiscretePolicy(env.action_space.n)
@@ -72,11 +73,11 @@ def load_policy(method:str,
             torch.zeros(env.observation_space['safe_actions'].shape)
         )
         policy = ReparamSafePolicy(policy)
-        policy.load_state_dict(torch.load(policy_file))
+        policy.load_state_dict(torch.load(policy_file, map_location=ml))
         policy.eval()
     elif method == 'sgail-ppo':
         policy = SetMaskedDiscretePolicy(env.action_space.n)
-        policy.load_state_dict(torch.load(policy_file))
+        policy.load_state_dict(torch.load(policy_file, map_location=ml))
         policy.eval()
     else:
         raise NotImplementedError

src/safe_options/options.py

@@ -52,7 +52,8 @@ def gail(env_fn, expert_data, discriminator, disc_opt, disc_iters, policy, value
 
         generator_data.ll.actions += 0.1 * torch.randn_like(generator_data.ll.actions)
 
-        logger.add_scalar('gen/mean_episode_length', (~generator_data.ll.dones).sum() / generator_data.ll.states.shape[0], epoch)
+        gen_mean_episode_length = (~generator_data.ll.dones).sum() / generator_data.ll.states.shape[0]
+        logger.add_scalar('gen/mean_episode_length', gen_mean_episode_length , epoch)
         gen_mean_reward_per_episode = generator_data.hl.rewards[~generator_data.hl.dones].sum() / generator_data.hl.states.shape[0]
         logger.add_scalar('gen/mean_reward_per_episode', gen_mean_reward_per_episode, epoch)
         logger.add_scalar('gen/unsafe_probability_mass', policy.unsafe_probability_mass(policy(generator_data.hl.states[~generator_data.hl.dones], generator_data.hl.safe_actions[~generator_data.hl.dones])).mean(), epoch)
@@ -64,7 +65,8 @@ def gail(env_fn, expert_data, discriminator, disc_opt, disc_iters, policy, value
         else:
             generator_data.ll.rewards = -F.logsigmoid(discriminator(generator_data.ll.states, generator_data.ll.actions))
         logger.add_scalar('disc/final_loss', loss, epoch)
-        logger.add_scalar('disc/mean_reward_per_episode', generator_data.ll.rewards[~generator_data.ll.dones].sum() / generator_data.ll.states.shape[0], epoch)
+        disc_mean_reward_per_episode = generator_data.ll.rewards[~generator_data.ll.dones].sum() / generator_data.ll.states.shape[0]
+        logger.add_scalar('disc/mean_reward_per_episode', disc_mean_reward_per_episode , epoch)
 
         #assert generator_data.ll.rewards.shape == generator_data.ll.dones.shape
         generator_data.hl.rewards = torch.where(~generator_data.ll.dones, generator_data.ll.rewards, torch.tensor(0.)).sum(-1)
@@ -77,7 +79,9 @@ def gail(env_fn, expert_data, discriminator, disc_opt, disc_iters, policy, value
                 'epoch': epoch,
                 'value': value,
                 'policy': policy,
-                'gen/mean_reward_per_episode': gen_mean_reward_per_episode,
+                'gen/mean_episode_length': gen_mean_episode_length.item(),
+                'gen/mean_reward_per_episode': gen_mean_reward_per_episode.item(),
+                'disc/mean_reward_per_episode': disc_mean_reward_per_episode.item(),
             })
     
     return value, policy
@@ -94,8 +98,8 @@ def gail_ppo(env_fn, expert_data, discriminator, disc_opt, disc_iters, policy, v
         generator_data = OptionsRollout(HLBuffer(*hl_data), Buffer(*ll_data))
 
         generator_data.ll.actions += 0.1 * torch.randn_like(generator_data.ll.actions)
-
+        gen_mean_episode_length = (~generator_data.ll.dones).sum() / generator_data.ll.states.shape[0]
-        logger.add_scalar('gen/mean_episode_length', (~generator_data.ll.dones).sum() / generator_data.ll.states.shape[0], epoch)
+        logger.add_scalar('gen/mean_episode_length', gen_mean_episode_length, epoch)
         gen_mean_reward_per_episode = generator_data.hl.rewards[~generator_data.hl.dones].sum() / generator_data.hl.states.shape[0]
         logger.add_scalar('gen/mean_reward_per_episode', gen_mean_reward_per_episode, epoch)
         logger.add_scalar('gen/unsafe_probability_mass', policy.unsafe_probability_mass(policy(generator_data.hl.states[~generator_data.hl.dones], generator_data.hl.safe_actions[~generator_data.hl.dones])).mean(), epoch)
@@ -107,7 +111,8 @@ def gail_ppo(env_fn, expert_data, discriminator, disc_opt, disc_iters, policy, v
         else:
             generator_data.ll.rewards = -F.logsigmoid(discriminator(generator_data.ll.states, generator_data.ll.actions))
         logger.add_scalar('disc/final_loss', loss, epoch)
-        logger.add_scalar('disc/mean_reward_per_episode', generator_data.ll.rewards[~generator_data.ll.dones].sum() / generator_data.ll.states.shape[0], epoch)
+        disc_mean_reward_per_episode = generator_data.ll.rewards[~generator_data.ll.dones].sum() / generator_data.ll.states.shape[0]
+        logger.add_scalar('disc/mean_reward_per_episode', disc_mean_reward_per_episode, epoch)
         
         #assert generator_data.ll.rewards.shape == generator_data.ll.dones.shape
         generator_data.hl.rewards = torch.where(~generator_data.ll.dones, generator_data.ll.rewards, torch.tensor(0.)).sum(-1)
@@ -120,7 +125,9 @@ def gail_ppo(env_fn, expert_data, discriminator, disc_opt, disc_iters, policy, v
                 'epoch': epoch,
                 'value': value,
                 'policy': policy,
-                'gen/mean_reward_per_episode': gen_mean_reward_per_episode,
+                'gen/mean_episode_length': gen_mean_episode_length.item(),
+                'gen/mean_reward_per_episode': gen_mean_reward_per_episode.item(),
+                'disc/mean_reward_per_episode': disc_mean_reward_per_episode.item(),
             })
     
     for lr_scheduler in lr_schedulers:
@@ -248,8 +255,8 @@ class SafeOptionsEnv(OptionsEnv):
             if d:
                 break
 
-            if self.abort_unsafe_collision_method is not None and \
+            if self.abort_unsafe_collision_method is not None:
-                not feasible(self.env, plan[k:], method=self.abort_unsafe_collision_method):
+                if not feasible(self.env, plan[k:], method=self.abort_unsafe_collision_method):
                     break
         
         n_steps = k + 1