sac mujoco result (#246)

examples/atari/atari_dqn.py

@@ -58,8 +58,8 @@ def test_dqn(args=get_args()):
     args.state_shape = env.observation_space.shape or env.observation_space.n
     args.action_shape = env.env.action_space.shape or env.env.action_space.n
     # should be N_FRAMES x H x W
-    print("Observations shape: ", args.state_shape)
-    print("Actions shape: ", args.action_shape)
+    print("Observations shape:", args.state_shape)
+    print("Actions shape:", args.action_shape)
     # make environments
     train_envs = SubprocVectorEnv([lambda: make_atari_env(args)
                                    for _ in range(args.training_num)])
@@ -79,7 +79,9 @@ def test_dqn(args=get_args()):
                        target_update_freq=args.target_update_freq)
     # load a previous policy
     if args.resume_path:
-        policy.load_state_dict(torch.load(args.resume_path))
+        policy.load_state_dict(torch.load(
+            args.resume_path, map_location=args.device
+        ))
         print("Loaded agent from: ", args.resume_path)
     # replay buffer: `save_last_obs` and `stack_num` can be removed together
     # when you have enough RAM

examples/box2d/README.md

@@ -1,6 +1,6 @@
 # Bipedal-Hardcore-SAC
 
-- Our default choice: remove the done flag penalty, will soon converge to \~270 reward within 100 epochs (10M env steps, 3~4 hours, see the image below)
+- Our default choice: remove the done flag penalty, will soon converge to \~280 reward within 100 epochs (10M env steps, 3~4 hours, see the image below)
 - If the done penalty is not removed, it converges much slower than before, about 200 epochs (20M env steps) to reach the same performance (\~200 reward)
 
 ![](results/sac/BipedalHardcore.png)

examples/box2d/bipedal_hardcore_sac.py

@@ -6,11 +6,11 @@ import argparse
 import numpy as np
 from torch.utils.tensorboard import SummaryWriter
 
+from tianshou.policy import SACPolicy
+from tianshou.utils.net.common import Net
 from tianshou.env import SubprocVectorEnv
 from tianshou.trainer import offpolicy_trainer
 from tianshou.data import Collector, ReplayBuffer
-from tianshou.policy import SACPolicy
-from tianshou.utils.net.common import Net
 from tianshou.utils.net.continuous import ActorProb, Critic
 
 
@@ -24,8 +24,8 @@ def get_args():
     parser.add_argument('--gamma', type=float, default=0.99)
     parser.add_argument('--tau', type=float, default=0.005)
     parser.add_argument('--alpha', type=float, default=0.1)
-    parser.add_argument('--auto_alpha', type=int, default=1)
-    parser.add_argument('--alpha_lr', type=float, default=3e-4)
+    parser.add_argument('--auto-alpha', type=int, default=1)
+    parser.add_argument('--alpha-lr', type=float, default=3e-4)
     parser.add_argument('--epoch', type=int, default=100)
     parser.add_argument('--step-per-epoch', type=int, default=10000)
     parser.add_argument('--collect-per-step', type=int, default=10)
@@ -35,54 +35,50 @@ def get_args():
     parser.add_argument('--test-num', type=int, default=100)
     parser.add_argument('--logdir', type=str, default='log')
     parser.add_argument('--render', type=float, default=0.)
-    parser.add_argument('--rew-norm', type=int, default=0)
-    parser.add_argument('--ignore-done', type=int, default=0)
     parser.add_argument('--n-step', type=int, default=4)
     parser.add_argument(
         '--device', type=str,
         default='cuda' if torch.cuda.is_available() else 'cpu')
-    parser.add_argument('--resume_path', type=str, default=None)
+    parser.add_argument('--resume-path', type=str, default=None)
     return parser.parse_args()
 
 
-class EnvWrapper(object):
-    """Env wrapper for reward scale, action repeat and action noise"""
+class Wrapper(gym.Wrapper):
+    """Env wrapper for reward scale, action repeat and removing done penalty"""
 
-    def __init__(self, task, action_repeat=3, reward_scale=5, act_noise=0.0):
-        self._env = gym.make(task)
+    def __init__(self, env, action_repeat=3, reward_scale=5, rm_done=True):
+        super().__init__(env)
         self.action_repeat = action_repeat
         self.reward_scale = reward_scale
-        self.act_noise = act_noise
-
-    def __getattr__(self, name):
-        return getattr(self._env, name)
+        self.rm_done = rm_done
 
     def step(self, action):
-        # add action noise
-        action += self.act_noise * (-2 * np.random.random(4) + 1)
         r = 0.0
         for _ in range(self.action_repeat):
-            obs_, reward_, done_, info_ = self._env.step(action)
+            obs, reward, done, info = self.env.step(action)
             # remove done reward penalty
-            if done_:
+            if not done or not self.rm_done:
+                r = r + reward
+            if done:
                 break
-            r = r + reward_
         # scale reward
-        return obs_, self.reward_scale * r, done_, info_
+        return obs, self.reward_scale * r, done, info
 
 
 def test_sac_bipedal(args=get_args()):
-    env = EnvWrapper(args.task)
+    env = Wrapper(gym.make(args.task))
 
     args.state_shape = env.observation_space.shape or env.observation_space.n
     args.action_shape = env.action_space.shape or env.action_space.n
     args.max_action = env.action_space.high[0]
 
-    train_envs = SubprocVectorEnv(
-        [lambda: EnvWrapper(args.task) for _ in range(args.training_num)])
+    train_envs = SubprocVectorEnv([
+        lambda: Wrapper(gym.make(args.task))
+        for _ in range(args.training_num)])
     # test_envs = gym.make(args.task)
-    test_envs = SubprocVectorEnv([lambda: EnvWrapper(args.task, reward_scale=1)
-                                  for _ in range(args.test_num)])
+    test_envs = SubprocVectorEnv([
+        lambda: Wrapper(gym.make(args.task), reward_scale=1, rm_done=False)
+        for _ in range(args.test_num)])
 
     # seed
     np.random.seed(args.seed)
@@ -117,8 +113,6 @@ def test_sac_bipedal(args=get_args()):
         actor, actor_optim, critic1, critic1_optim, critic2, critic2_optim,
         action_range=[env.action_space.low[0], env.action_space.high[0]],
         tau=args.tau, gamma=args.gamma, alpha=args.alpha,
-        reward_normalization=args.rew_norm,
-        ignore_done=args.ignore_done,
         estimation_step=args.n_step)
     # load a previous policy
     if args.resume_path:

examples/box2d/mcc_sac.py

@@ -67,11 +67,13 @@ def test_sac(args=get_args()):
         args.max_action, args.device, unbounded=True
     ).to(args.device)
     actor_optim = torch.optim.Adam(actor.parameters(), lr=args.actor_lr)
-    net = Net(args.layer_num, args.state_shape,
-              args.action_shape, concat=True, device=args.device)
-    critic1 = Critic(net, args.device).to(args.device)
+    net_c1 = Net(args.layer_num, args.state_shape,
+                 args.action_shape, concat=True, device=args.device)
+    critic1 = Critic(net_c1, args.device).to(args.device)
     critic1_optim = torch.optim.Adam(critic1.parameters(), lr=args.critic_lr)
-    critic2 = Critic(net, args.device).to(args.device)
+    net_c2 = Net(args.layer_num, args.state_shape,
+                 args.action_shape, concat=True, device=args.device)
+    critic2 = Critic(net_c2, args.device).to(args.device)
     critic2_optim = torch.optim.Adam(critic2.parameters(), lr=args.critic_lr)
 
     if args.auto_alpha:

examples/mujoco/README.md

@@ -1,3 +1,27 @@
-Result of Ant-v2:
+# Mujoco Result
+
+
+
+## SAC (single run)
+
+The best reward computes from 100 episodes returns in the test phase.
+
+SAC on Swimmer-v3 always stops at 47\~48.
+
+| task           | 3M best reward    | parameters                                              | time cost (3M) |
+| -------------- | ----------------- | ------------------------------------------------------- | -------------- |
+| HalfCheetah-v3 | 10157.70 ± 171.70 | `python3 mujoco_sac.py --task HalfCheetah-v3`           | 2~3h           |
+| Walker2d-v3    | 5143.04 ± 15.57   | `python3 mujoco_sac.py --task Walker2d-v3`              | 2~3h           |
+| Hopper-v3      | 3604.19 ± 169.55  | `python3 mujoco_sac.py --task Hopper-v3`                | 2~3h           |
+| Humanoid-v3    | 6579.20 ± 1470.57 | `python3 mujoco_sac.py --task Humanoid-v3 --alpha 0.05` | 2~3h           |
+| Ant-v3         | 6281.65 ± 686.28  | `python3 mujoco_sac.py --task Ant-v3`                   | 2~3h           |
+
+![](results/sac/all.png)
+
+### Which parts are important?
+
+0. DO NOT share the same network with two critic networks.
+1. The sigma (of the Gaussian policy) MUST be conditioned on input.
+2. The network size should not be less than 256.
+3. The deterministic evaluation helps a lot :)
 
-![](/docs/_static/images/Ant-v2.png)

examples/mujoco/mujoco_sac.py

@@ -16,27 +16,36 @@ from tianshou.utils.net.continuous import ActorProb, Critic
 
 def get_args():
     parser = argparse.ArgumentParser()
-    parser.add_argument('--task', type=str, default='Ant-v2')
+    parser.add_argument('--task', type=str, default='Ant-v3')
     parser.add_argument('--seed', type=int, default=1626)
-    parser.add_argument('--buffer-size', type=int, default=20000)
+    parser.add_argument('--buffer-size', type=int, default=1000000)
     parser.add_argument('--actor-lr', type=float, default=3e-4)
-    parser.add_argument('--critic-lr', type=float, default=1e-3)
+    parser.add_argument('--critic-lr', type=float, default=3e-4)
     parser.add_argument('--gamma', type=float, default=0.99)
     parser.add_argument('--tau', type=float, default=0.005)
     parser.add_argument('--alpha', type=float, default=0.2)
+    parser.add_argument('--auto-alpha', default=False, action='store_true')
+    parser.add_argument('--alpha-lr', type=float, default=3e-4)
+    parser.add_argument('--n-step', type=int, default=2)
     parser.add_argument('--epoch', type=int, default=100)
-    parser.add_argument('--step-per-epoch', type=int, default=2400)
-    parser.add_argument('--collect-per-step', type=int, default=10)
-    parser.add_argument('--batch-size', type=int, default=128)
+    parser.add_argument('--step-per-epoch', type=int, default=10000)
+    parser.add_argument('--collect-per-step', type=int, default=4)
+    parser.add_argument('--update-per-step', type=int, default=1)
+    parser.add_argument('--pre-collect-step', type=int, default=10000)
+    parser.add_argument('--batch-size', type=int, default=256)
+    parser.add_argument('--hidden-layer-size', type=int, default=256)
     parser.add_argument('--layer-num', type=int, default=1)
-    parser.add_argument('--training-num', type=int, default=8)
+    parser.add_argument('--training-num', type=int, default=16)
     parser.add_argument('--test-num', type=int, default=100)
     parser.add_argument('--logdir', type=str, default='log')
     parser.add_argument('--render', type=float, default=0.)
-    parser.add_argument('--rew-norm', type=bool, default=True)
+    parser.add_argument('--log-interval', type=int, default=1000)
     parser.add_argument(
         '--device', type=str,
         default='cuda' if torch.cuda.is_available() else 'cpu')
+    parser.add_argument('--resume-path', type=str, default=None)
+    parser.add_argument('--watch', default=False, action='store_true',
+                        help='watch the play of pre-trained policy only')
     return parser.parse_args()
 
 
@@ -45,6 +54,10 @@ def test_sac(args=get_args()):
     args.state_shape = env.observation_space.shape or env.observation_space.n
     args.action_shape = env.action_space.shape or env.action_space.n
     args.max_action = env.action_space.high[0]
+    print("Observations shape:", args.state_shape)
+    print("Actions shape:", args.action_shape)
+    print("Action range:", np.min(env.action_space.low),
+          np.max(env.action_space.high))
     # train_envs = gym.make(args.task)
     train_envs = SubprocVectorEnv(
         [lambda: gym.make(args.task) for _ in range(args.training_num)])
@@ -57,53 +70,84 @@ def test_sac(args=get_args()):
     train_envs.seed(args.seed)
     test_envs.seed(args.seed)
     # model
-    net = Net(args.layer_num, args.state_shape, device=args.device)
+    net = Net(args.layer_num, args.state_shape, device=args.device,
+              hidden_layer_size=args.hidden_layer_size)
     actor = ActorProb(
-        net, args.action_shape,
-        args.max_action, args.device, unbounded=True
+        net, args.action_shape, args.max_action, args.device, unbounded=True,
+        hidden_layer_size=args.hidden_layer_size, conditioned_sigma=True,
     ).to(args.device)
     actor_optim = torch.optim.Adam(actor.parameters(), lr=args.actor_lr)
-    net = Net(args.layer_num, args.state_shape,
-              args.action_shape, concat=True, device=args.device)
-    critic1 = Critic(net, args.device).to(args.device)
+    net_c1 = Net(args.layer_num, args.state_shape, args.action_shape,
+                 concat=True, device=args.device,
+                 hidden_layer_size=args.hidden_layer_size)
+    critic1 = Critic(
+        net_c1, args.device, hidden_layer_size=args.hidden_layer_size
+    ).to(args.device)
     critic1_optim = torch.optim.Adam(critic1.parameters(), lr=args.critic_lr)
-    critic2 = Critic(net, args.device).to(args.device)
+    net_c2 = Net(args.layer_num, args.state_shape, args.action_shape,
+                 concat=True, device=args.device,
+                 hidden_layer_size=args.hidden_layer_size)
+    critic2 = Critic(
+        net_c2, args.device, hidden_layer_size=args.hidden_layer_size
+    ).to(args.device)
     critic2_optim = torch.optim.Adam(critic2.parameters(), lr=args.critic_lr)
+
+    if args.auto_alpha:
+        target_entropy = -np.prod(env.action_space.shape)
+        log_alpha = torch.zeros(1, requires_grad=True, device=args.device)
+        alpha_optim = torch.optim.Adam([log_alpha], lr=args.alpha_lr)
+        args.alpha = (target_entropy, log_alpha, alpha_optim)
+
     policy = SACPolicy(
         actor, actor_optim, critic1, critic1_optim, critic2, critic2_optim,
         action_range=[env.action_space.low[0], env.action_space.high[0]],
         tau=args.tau, gamma=args.gamma, alpha=args.alpha,
-        reward_normalization=args.rew_norm, ignore_done=True)
+        estimation_step=args.n_step)
+    # load a previous policy
+    if args.resume_path:
+        policy.load_state_dict(torch.load(
+            args.resume_path, map_location=args.device
+        ))
+        print("Loaded agent from: ", args.resume_path)
+
     # collector
     train_collector = Collector(
         policy, train_envs, ReplayBuffer(args.buffer_size))
     test_collector = Collector(policy, test_envs)
-    # train_collector.collect(n_step=args.buffer_size)
     # log
     log_path = os.path.join(args.logdir, args.task, 'sac')
     writer = SummaryWriter(log_path)
 
-    def save_fn(policy):
-        torch.save(policy.state_dict(), os.path.join(log_path, 'policy.pth'))
-
-    def stop_fn(mean_rewards):
-        return mean_rewards >= env.spec.reward_threshold
-
-    # trainer
-    result = offpolicy_trainer(
-        policy, train_collector, test_collector, args.epoch,
-        args.step_per_epoch, args.collect_per_step, args.test_num,
-        args.batch_size, stop_fn=stop_fn, save_fn=save_fn, writer=writer)
-    assert stop_fn(result['best_reward'])
-    if __name__ == '__main__':
-        pprint.pprint(result)
-        # Let's watch its performance!
+    def watch():
+        # watch agent's performance
+        print("Testing agent ...")
         policy.eval()
         test_envs.seed(args.seed)
         test_collector.reset()
         result = test_collector.collect(n_episode=[1] * args.test_num,
                                         render=args.render)
-        print(f'Final reward: {result["rew"]}, length: {result["len"]}')
+        pprint.pprint(result)
+
+    def save_fn(policy):
+        torch.save(policy.state_dict(), os.path.join(log_path, 'policy.pth'))
+
+    def stop_fn(mean_rewards):
+        return False
+
+    if args.watch:
+        watch()
+        exit(0)
+
+    # trainer
+    train_collector.collect(n_step=args.pre_collect_step, random=True)
+    result = offpolicy_trainer(
+        policy, train_collector, test_collector, args.epoch,
+        args.step_per_epoch, args.collect_per_step, args.test_num,
+        args.batch_size, args.update_per_step,
+        stop_fn=stop_fn, save_fn=save_fn, writer=writer,
+        log_interval=args.log_interval)
+    pprint.pprint(result)
+    watch()
 
 
 if __name__ == '__main__':

examples/mujoco/runnable/halfcheetahBullet_v0_sac.py

@@ -71,6 +71,8 @@ def test_sac(args=get_args()):
               args.action_shape, concat=True, device=args.device)
     critic1 = Critic(net, args.device).to(args.device)
     critic1_optim = torch.optim.Adam(critic1.parameters(), lr=args.critic_lr)
+    net = Net(args.layer_num, args.state_shape,
+              args.action_shape, concat=True, device=args.device)
     critic2 = Critic(net, args.device).to(args.device)
     critic2_optim = torch.optim.Adam(critic2.parameters(), lr=args.critic_lr)
     policy = SACPolicy(

examples/mujoco/runnable/point_maze_td3.py

@@ -72,6 +72,8 @@ def test_td3(args=get_args()):
               args.action_shape, concat=True, device=args.device)
     critic1 = Critic(net, args.device).to(args.device)
     critic1_optim = torch.optim.Adam(critic1.parameters(), lr=args.critic_lr)
+    net = Net(args.layer_num, args.state_shape,
+              args.action_shape, concat=True, device=args.device)
     critic2 = Critic(net, args.device).to(args.device)
     critic2_optim = torch.optim.Adam(critic2.parameters(), lr=args.critic_lr)
     policy = TD3Policy(

test/continuous/test_sac_with_il.py

@@ -70,11 +70,13 @@ def test_sac_with_il(args=get_args()):
         net, args.action_shape, args.max_action, args.device, unbounded=True
     ).to(args.device)
     actor_optim = torch.optim.Adam(actor.parameters(), lr=args.actor_lr)
-    net = Net(args.layer_num, args.state_shape,
-              args.action_shape, concat=True, device=args.device)
-    critic1 = Critic(net, args.device).to(args.device)
+    net_c1 = Net(args.layer_num, args.state_shape,
+                 args.action_shape, concat=True, device=args.device)
+    critic1 = Critic(net_c1, args.device).to(args.device)
     critic1_optim = torch.optim.Adam(critic1.parameters(), lr=args.critic_lr)
-    critic2 = Critic(net, args.device).to(args.device)
+    net_c2 = Net(args.layer_num, args.state_shape,
+                 args.action_shape, concat=True, device=args.device)
+    critic2 = Critic(net_c2, args.device).to(args.device)
     critic2_optim = torch.optim.Adam(critic2.parameters(), lr=args.critic_lr)
     policy = SACPolicy(
         actor, actor_optim, critic1, critic1_optim, critic2, critic2_optim,

test/continuous/test_td3.py

@@ -74,11 +74,13 @@ def test_td3(args=get_args()):
         args.max_action, args.device
     ).to(args.device)
     actor_optim = torch.optim.Adam(actor.parameters(), lr=args.actor_lr)
-    net = Net(args.layer_num, args.state_shape,
-              args.action_shape, concat=True, device=args.device)
-    critic1 = Critic(net, args.device).to(args.device)
+    net_c1 = Net(args.layer_num, args.state_shape,
+                 args.action_shape, concat=True, device=args.device)
+    critic1 = Critic(net_c1, args.device).to(args.device)
     critic1_optim = torch.optim.Adam(critic1.parameters(), lr=args.critic_lr)
-    critic2 = Critic(net, args.device).to(args.device)
+    net_c2 = Net(args.layer_num, args.state_shape,
+                 args.action_shape, concat=True, device=args.device)
+    critic2 = Critic(net_c2, args.device).to(args.device)
     critic2_optim = torch.optim.Adam(critic2.parameters(), lr=args.critic_lr)
     policy = TD3Policy(
         actor, actor_optim, critic1, critic1_optim, critic2, critic2_optim,

test/discrete/test_sac.py

@@ -62,11 +62,11 @@ def test_discrete_sac(args=get_args()):
     net = Net(args.layer_num, args.state_shape, device=args.device)
     actor = Actor(net, args.action_shape, softmax_output=False).to(args.device)
     actor_optim = torch.optim.Adam(actor.parameters(), lr=args.actor_lr)
-    net = Net(args.layer_num, args.state_shape, device=args.device)
-    critic1 = Critic(net, last_size=args.action_shape).to(args.device)
+    net_c1 = Net(args.layer_num, args.state_shape, device=args.device)
+    critic1 = Critic(net_c1, last_size=args.action_shape).to(args.device)
     critic1_optim = torch.optim.Adam(critic1.parameters(), lr=args.critic_lr)
-    net = Net(args.layer_num, args.state_shape, device=args.device)
-    critic2 = Critic(net, last_size=args.action_shape).to(args.device)
+    net_c2 = Net(args.layer_num, args.state_shape, device=args.device)
+    critic2 = Critic(net_c2, last_size=args.action_shape).to(args.device)
     critic2_optim = torch.optim.Adam(critic2.parameters(), lr=args.critic_lr)
 
     # better not to use auto alpha in CartPole

tianshou/policy/modelfree/sac.py

@@ -38,6 +38,9 @@ class SACPolicy(DDPGPolicy):
         defaults to False.
     :param BaseNoise exploration_noise: add a noise to action for exploration,
         defaults to None. This is useful when solving hard-exploration problem.
+    :param bool deterministic_eval: whether to use deterministic action (mean
+        of Gaussian policy) instead of stochastic action sampled by the policy,
+        defaults to True.
 
     .. seealso::
 
@@ -63,6 +66,7 @@ class SACPolicy(DDPGPolicy):
         ignore_done: bool = False,
         estimation_step: int = 1,
         exploration_noise: Optional[BaseNoise] = None,
+        deterministic_eval: bool = True,
         **kwargs: Any,
     ) -> None:
         super().__init__(None, None, None, None, action_range, tau, gamma,
@@ -86,6 +90,7 @@ class SACPolicy(DDPGPolicy):
         else:
             self._alpha = alpha
 
+        self._deterministic_eval = deterministic_eval
         self.__eps = np.finfo(np.float32).eps.item()
 
     def train(self, mode: bool = True) -> "SACPolicy":
@@ -116,13 +121,16 @@ class SACPolicy(DDPGPolicy):
         logits, h = self.actor(obs, state=state, info=batch.info)
         assert isinstance(logits, tuple)
         dist = Independent(Normal(*logits), 1)
-        x = dist.rsample()
+        if self._deterministic_eval and not self.training:
+            x = logits[0]
+        else:
+            x = dist.rsample()
         y = torch.tanh(x)
         act = y * self._action_scale + self._action_bias
         y = self._action_scale * (1 - y.pow(2)) + self.__eps
         log_prob = dist.log_prob(x).unsqueeze(-1)
         log_prob = log_prob - torch.log(y).sum(-1, keepdim=True)
-        if self._noise is not None and not self.updating:
+        if self._noise is not None and self.training and not self.updating:
             act += to_torch_as(self._noise(act.shape), act)
         act = act.clamp(self._range[0], self._range[1])
         return Batch(

tianshou/utils/net/continuous.py

@@ -6,6 +6,10 @@ from typing import Any, Dict, Tuple, Union, Optional, Sequence
 from tianshou.data import to_torch, to_torch_as
 
 
+SIGMA_MIN = -20
+SIGMA_MAX = 2
+
+
 class Actor(nn.Module):
     """Simple actor network with MLP.
 
@@ -89,12 +93,17 @@ class ActorProb(nn.Module):
         device: Union[str, int, torch.device] = "cpu",
         unbounded: bool = False,
         hidden_layer_size: int = 128,
+        conditioned_sigma: bool = False,
     ) -> None:
         super().__init__()
         self.preprocess = preprocess_net
         self.device = device
         self.mu = nn.Linear(hidden_layer_size, np.prod(action_shape))
-        self.sigma = nn.Parameter(torch.zeros(np.prod(action_shape), 1))
+        self._c_sigma = conditioned_sigma
+        if conditioned_sigma:
+            self.sigma = nn.Linear(hidden_layer_size, np.prod(action_shape))
+        else:
+            self.sigma = nn.Parameter(torch.zeros(np.prod(action_shape), 1))
         self._max = max_action
         self._unbounded = unbounded
 
@@ -109,9 +118,14 @@ class ActorProb(nn.Module):
         mu = self.mu(logits)
         if not self._unbounded:
             mu = self._max * torch.tanh(mu)
-        shape = [1] * len(mu.shape)
-        shape[1] = -1
-        sigma = (self.sigma.view(shape) + torch.zeros_like(mu)).exp()
+        if self._c_sigma:
+            sigma = torch.clamp(
+                self.sigma(logits), min=SIGMA_MIN, max=SIGMA_MAX
+            ).exp()
+        else:
+            shape = [1] * len(mu.shape)
+            shape[1] = -1
+            sigma = (self.sigma.view(shape) + torch.zeros_like(mu)).exp()
         return (mu, sigma), state
 
 
@@ -131,6 +145,7 @@ class RecurrentActorProb(nn.Module):
         device: Union[str, int, torch.device] = "cpu",
         unbounded: bool = False,
         hidden_layer_size: int = 128,
+        conditioned_sigma: bool = False,
     ) -> None:
         super().__init__()
         self.device = device
@@ -141,7 +156,11 @@ class RecurrentActorProb(nn.Module):
             batch_first=True,
         )
         self.mu = nn.Linear(hidden_layer_size, np.prod(action_shape))
-        self.sigma = nn.Parameter(torch.zeros(np.prod(action_shape), 1))
+        self._c_sigma = conditioned_sigma
+        if conditioned_sigma:
+            self.sigma = nn.Linear(hidden_layer_size, np.prod(action_shape))
+        else:
+            self.sigma = nn.Parameter(torch.zeros(np.prod(action_shape), 1))
         self._max = max_action
         self._unbounded = unbounded
 
@@ -170,9 +189,14 @@ class RecurrentActorProb(nn.Module):
         mu = self.mu(logits)
         if not self._unbounded:
             mu = self._max * torch.tanh(mu)
-        shape = [1] * len(mu.shape)
-        shape[1] = -1
-        sigma = (self.sigma.view(shape) + torch.zeros_like(mu)).exp()
+        if self._c_sigma:
+            sigma = torch.clamp(
+                self.sigma(logits), min=SIGMA_MIN, max=SIGMA_MAX
+            ).exp()
+        else:
+            shape = [1] * len(mu.shape)
+            shape[1] = -1
+            sigma = (self.sigma.view(shape) + torch.zeros_like(mu)).exp()
         # please ensure the first dim is batch size: [bsz, len, ...]
         return (mu, sigma), {"h": h.transpose(0, 1).detach(),
                              "c": c.transpose(0, 1).detach()}