Adding Munchausen Reinforcement Learning

benchmark/dqn_munchausen.sh

@@ -0,0 +1,6 @@
+poetry install
+OMP_NUM_THREADS=1 xvfb-run -a poetry run python -m cleanrl_utils.benchmark \
+    --env-ids CartPole-v1 Acrobot-v1 MountainCar-v0 \
+    --command "poetry run python cleanrl/dqn_munchausen.py --no_cuda --track --capture_video" \
+    --num-seeds 3 \
+    --workers 9

cleanrl/dqn_munchausen.py

@@ -0,0 +1,264 @@
+# docs and experiment results can be found at https://docs.cleanrl.dev/rl-algorithms/dqn/#dqn_munchausenpy
+# Paper link: https://arxiv.org/pdf/2007.14430
+import os
+import random
+import time
+from dataclasses import dataclass
+
+import gymnasium as gym
+import numpy as np
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import torch.optim as optim
+import tyro
+from stable_baselines3.common.buffers import ReplayBuffer
+from torch.utils.tensorboard import SummaryWriter
+
+
+@dataclass
+class Args:
+    exp_name: str = os.path.basename(__file__)[: -len(".py")]
+    """the name of this experiment"""
+    seed: int = 1
+    """seed of the experiment"""
+    torch_deterministic: bool = True
+    """if toggled, `torch.backends.cudnn.deterministic=False`"""
+    cuda: bool = True
+    """if toggled, cuda will be enabled by default"""
+    track: bool = False
+    """if toggled, this experiment will be tracked with Weights and Biases"""
+    wandb_project_name: str = "cleanRL"
+    """the wandb's project name"""
+    wandb_entity: str = None
+    """the entity (team) of wandb's project"""
+    capture_video: bool = False
+    """whether to capture videos of the agent performances (check out `videos` folder)"""
+    save_model: bool = False
+    """whether to save model into the `runs/{run_name}` folder"""
+    upload_model: bool = False
+    """whether to upload the saved model to huggingface"""
+    hf_entity: str = ""
+    """the user or org name of the model repository from the Hugging Face Hub"""
+
+    # Algorithm specific arguments
+    env_id: str = "CartPole-v1"
+    """the id of the environment"""
+    total_timesteps: int = 500000
+    """total timesteps of the experiments"""
+    learning_rate: float = 2.5e-4
+    """the learning rate of the optimizer"""
+    num_envs: int = 1
+    """the number of parallel game environments"""
+    buffer_size: int = 10000
+    """the replay memory buffer size"""
+    gamma: float = 0.99
+    """the discount factor gamma"""
+    tau: float = 1.0
+    """the target network update rate"""
+    target_network_frequency: int = 500
+    """the timesteps it takes to update the target network"""
+    batch_size: int = 128
+    """the batch size of sample from the reply memory"""
+    learning_starts: int = 10000
+    """timestep to start learning"""
+    train_frequency: int = 10
+    """the frequency of training"""
+    tau_soft: float = 0.03
+    """the temperature parameter for the soft-max policy"""
+    alpha: float = 0.9
+    """the entropy regularization parameter"""
+    l_0: float = -1.0
+    """the lower bound of the weighted log probability"""
+    epsilon_tar: float = 1e-6
+    """the epsilon term for numerical stability"""
+
+
+def make_env(env_id, seed, idx, capture_video, run_name):
+    def thunk():
+        if capture_video and idx == 0:
+            env = gym.make(env_id, render_mode="rgb_array")
+            env = gym.wrappers.RecordVideo(env, f"videos/{run_name}")
+        else:
+            env = gym.make(env_id)
+        env = gym.wrappers.RecordEpisodeStatistics(env)
+        env.action_space.seed(seed)
+
+        return env
+
+    return thunk
+
+
+# ALGO LOGIC: initialize agent here:
+class QNetwork(nn.Module):
+    def __init__(self, env):
+        super().__init__()
+        self.network = nn.Sequential(
+            nn.Linear(np.array(env.single_observation_space.shape).prod(), 120),
+            nn.ReLU(),
+            nn.Linear(120, 84),
+            nn.ReLU(),
+            nn.Linear(84, env.single_action_space.n),
+        )
+
+    def forward(self, x):
+        return self.network(x)
+
+
+if __name__ == "__main__":
+    import stable_baselines3 as sb3
+
+    if sb3.__version__ < "2.0":
+        raise ValueError(
+            """Ongoing migration: run the following command to install the new dependencies:
+
+poetry run pip install "stable_baselines3==2.0.0a1"
+"""
+        )
+    args = tyro.cli(Args)
+    assert args.num_envs == 1, "vectorized envs are not supported at the moment"
+    run_name = f"{args.env_id}__{args.exp_name}__{args.seed}__{int(time.time())}"
+    if args.track:
+        import wandb
+
+        wandb.init(
+            project=args.wandb_project_name,
+            entity=args.wandb_entity,
+            sync_tensorboard=True,
+            config=vars(args),
+            name=run_name,
+            monitor_gym=True,
+            save_code=True,
+        )
+    writer = SummaryWriter(f"runs/{run_name}")
+    writer.add_text(
+        "hyperparameters",
+        "|param|value|\n|-|-|\n%s" % ("\n".join([f"|{key}|{value}|" for key, value in vars(args).items()])),
+    )
+
+    # TRY NOT TO MODIFY: seeding
+    random.seed(args.seed)
+    np.random.seed(args.seed)
+    torch.manual_seed(args.seed)
+    torch.backends.cudnn.deterministic = args.torch_deterministic
+
+    device = torch.device("cuda" if torch.cuda.is_available() and args.cuda else "cpu")
+
+    # env setup
+    envs = gym.vector.SyncVectorEnv(
+        [make_env(args.env_id, args.seed + i, i, args.capture_video, run_name) for i in range(args.num_envs)]
+    )
+    assert isinstance(envs.single_action_space, gym.spaces.Discrete), "only discrete action space is supported"
+
+    q_network = QNetwork(envs).to(device)
+    optimizer = optim.Adam(q_network.parameters(), lr=args.learning_rate)
+    target_network = QNetwork(envs).to(device)
+    target_network.load_state_dict(q_network.state_dict())
+
+    rb = ReplayBuffer(
+        args.buffer_size,
+        envs.single_observation_space,
+        envs.single_action_space,
+        device,
+        handle_timeout_termination=False,
+    )
+    start_time = time.time()
+
+    # TRY NOT TO MODIFY: start the game
+    obs, _ = envs.reset(seed=args.seed)
+    for global_step in range(args.total_timesteps):
+        # ALGO LOGIC: put action logic here
+        if global_step < args.learning_starts:
+            actions = np.array([envs.single_action_space.sample() for _ in range(args.num_envs)])
+        else:
+            with torch.no_grad():
+                q_values = q_network(torch.Tensor(obs).to(device))
+                policy = F.softmax(q_values / args.tau_soft, dim=-1)
+                actions = torch.multinomial(policy, 1).squeeze(-1).cpu().numpy()
+        # TRY NOT TO MODIFY: execute the game and log data.
+        next_obs, rewards, terminations, truncations, infos = envs.step(actions)
+
+        # TRY NOT TO MODIFY: record rewards for plotting purposes
+        if "final_info" in infos:
+            for info in infos["final_info"]:
+                if info and "episode" in info:
+                    print(f"global_step={global_step}, episodic_return={info['episode']['r']}")
+                    writer.add_scalar("charts/episodic_return", info["episode"]["r"], global_step)
+                    writer.add_scalar("charts/episodic_length", info["episode"]["l"], global_step)
+
+        # TRY NOT TO MODIFY: save data to reply buffer; handle `final_observation`
+        real_next_obs = next_obs.copy()
+        for idx, trunc in enumerate(truncations):
+            if trunc:
+                real_next_obs[idx] = infos["final_observation"][idx]
+        rb.add(obs, real_next_obs, actions, rewards, terminations, infos)
+
+        # TRY NOT TO MODIFY: CRUCIAL step easy to overlook
+        obs = next_obs
+
+        # ALGO LOGIC: training.
+        if global_step > args.learning_starts:
+            if global_step % args.train_frequency == 0:
+                data = rb.sample(args.batch_size)
+                with torch.no_grad():
+                    target_q_values = target_network(data.observations)
+                    target_policy = F.softmax(target_q_values / args.tau_soft, dim=-1)
+                    target_next_q_values = target_network(data.next_observations)
+                    target_next_policy = F.softmax(target_next_q_values / args.tau_soft, dim=-1)
+                    red_term = args.alpha * (
+                        args.tau_soft * torch.log(target_policy.gather(1, data.actions)) + args.epsilon_tar).clamp(args.l_0, 0.0)  
+                    blue_term = -args.tau_soft * torch.log(target_next_policy + args.epsilon_tar)  
+                    munchausen_target = (data.rewards + red_term + args.gamma * (1 - data.dones)* (target_next_policy * (target_next_q_values + blue_term)).sum(dim=-1).unsqueeze(-1))
+                    td_target = munchausen_target.squeeze()
+                old_val = q_network(data.observations).gather(1, data.actions).squeeze()
+                loss = F.mse_loss(td_target, old_val)
+
+                if global_step % 100 == 0:
+                    writer.add_scalar("losses/td_loss", loss, global_step)
+                    writer.add_scalar("losses/q_values", old_val.mean().item(), global_step)
+                    writer.add_scalar("losses/td_target", td_target.mean().item(), global_step)
+                    writer.add_scalar("losses/log_policy", red_term.mean().item(), global_step)
+                    writer.add_scalar("losses/entropy", blue_term.mean().item(), global_step)
+                    print("SPS:", int(global_step / (time.time() - start_time)))
+                    writer.add_scalar("charts/SPS", int(global_step / (time.time() - start_time)), global_step)
+
+                # optimize the model
+                optimizer.zero_grad()
+                loss.backward()
+                optimizer.step()
+
+            # update target network
+            if global_step % args.target_network_frequency == 0:
+                for target_network_param, q_network_param in zip(target_network.parameters(), q_network.parameters()):
+                    target_network_param.data.copy_(
+                        args.tau * q_network_param.data + (1.0 - args.tau) * target_network_param.data
+                    )
+
+    if args.save_model:
+        model_path = f"runs/{run_name}/{args.exp_name}.cleanrl_model"
+        torch.save(q_network.state_dict(), model_path)
+        print(f"model saved to {model_path}")
+        from cleanrl_utils.evals.dqn_munchausen_eval import evaluate
+
+        episodic_returns = evaluate(
+            model_path,
+            make_env,
+            args.env_id,
+            eval_episodes=10,
+            run_name=f"{run_name}-eval",
+            Model=QNetwork,
+            device=device,
+            epsilon=0.05,
+        )
+        for idx, episodic_return in enumerate(episodic_returns):
+            writer.add_scalar("eval/episodic_return", episodic_return, idx)
+
+        if args.upload_model:
+            from cleanrl_utils.huggingface import push_to_hub
+
+            repo_name = f"{args.env_id}-{args.exp_name}-seed{args.seed}"
+            repo_id = f"{args.hf_entity}/{repo_name}" if args.hf_entity else repo_name
+            push_to_hub(args, episodic_returns, repo_id, "DQN_MUNCHAUSEN", f"runs/{run_name}", f"videos/{run_name}-eval")
+
+    envs.close()
+    writer.close()

cleanrl_utils/evals/dqn_munchausen_eval.py

@@ -0,0 +1,60 @@
+import random
+from typing import Callable
+
+import gymnasium as gym
+import numpy as np
+import torch
+
+
+def evaluate(
+    model_path: str,
+    make_env: Callable,
+    env_id: str,
+    eval_episodes: int,
+    run_name: str,
+    Model: torch.nn.Module,
+    device: torch.device = torch.device("cpu"),
+    epsilon: float = 0.05,
+    capture_video: bool = True,
+):
+    envs = gym.vector.SyncVectorEnv([make_env(env_id, 0, 0, capture_video, run_name)])
+    model = Model(envs).to(device)
+    model.load_state_dict(torch.load(model_path, map_location=device))
+    model.eval()
+
+    obs, _ = envs.reset()
+    episodic_returns = []
+    while len(episodic_returns) < eval_episodes:
+        if random.random() < epsilon:
+            actions = np.array([envs.single_action_space.sample() for _ in range(envs.num_envs)])
+        else:
+            q_values = model(torch.Tensor(obs).to(device))
+            actions = torch.argmax(q_values, dim=1).cpu().numpy()
+        next_obs, _, _, _, infos = envs.step(actions)
+        if "final_info" in infos:
+            for info in infos["final_info"]:
+                if "episode" not in info:
+                    continue
+                print(f"eval_episode={len(episodic_returns)}, episodic_return={info['episode']['r']}")
+                episodic_returns += [info["episode"]["r"]]
+        obs = next_obs
+
+    return episodic_returns
+
+
+if __name__ == "__main__":
+    from huggingface_hub import hf_hub_download
+
+    from cleanrl.dqn_munchausen import QNetwork, make_env
+
+    model_path = hf_hub_download(repo_id="cleanrl/CartPole-v1-dqn_munchausen-seed1", filename="q_network.pth")
+    evaluate(
+        model_path,
+        make_env,
+        "CartPole-v1",
+        eval_episodes=10,
+        run_name=f"eval",
+        Model=QNetwork,
+        device="cpu",
+        capture_video=False,
+    )