implement TD3+BC for offline RL

README.md

@@ -31,6 +31,7 @@
 - Vanilla Imitation Learning
 - [Batch-Constrained deep Q-Learning (BCQ)](https://arxiv.org/pdf/1812.02900.pdf)
 - [Conservative Q-Learning (CQL)](https://arxiv.org/pdf/2006.04779.pdf)
+- [Twin Delayed DDPG with Behavior Cloning (TD3+BC)](https://arxiv.org/pdf/2106.06860.pdf)
 - [Discrete Batch-Constrained deep Q-Learning (BCQ-Discrete)](https://arxiv.org/pdf/1910.01708.pdf)
 - [Discrete Conservative Q-Learning (CQL-Discrete)](https://arxiv.org/pdf/2006.04779.pdf)
 - [Discrete Critic Regularized Regression (CRR-Discrete)](https://arxiv.org/pdf/2006.15134.pdf)

docs/api/tianshou.policy.rst

@@ -129,6 +129,11 @@ Imitation
    :undoc-members:
    :show-inheritance:
 
+.. autoclass:: tianshou.policy.TD3BCPolicy
+   :members:
+   :undoc-members:
+   :show-inheritance:
+
 .. autoclass:: tianshou.policy.DiscreteBCQPolicy
    :members:
    :undoc-members:

docs/index.rst

@@ -31,6 +31,7 @@ Welcome to Tianshou!
 * :class:`~tianshou.policy.ImitationPolicy` Imitation Learning
 * :class:`~tianshou.policy.BCQPolicy` `Batch-Constrained deep Q-Learning <https://arxiv.org/pdf/1812.02900.pdf>`_
 * :class:`~tianshou.policy.CQLPolicy` `Conservative Q-Learning <https://arxiv.org/pdf/2006.04779.pdf>`_
+* :class:`~tianshou.policy.TD3BCPolicy` `Twin Delayed DDPG with Behavior Cloning <https://arxiv.org/pdf/2106.06860.pdf>`_
 * :class:`~tianshou.policy.DiscreteBCQPolicy` `Discrete Batch-Constrained deep Q-Learning <https://arxiv.org/pdf/1910.01708.pdf>`_
 * :class:`~tianshou.policy.DiscreteCQLPolicy` `Discrete Conservative Q-Learning <https://arxiv.org/pdf/2006.04779.pdf>`_
 * :class:`~tianshou.policy.DiscreteCRRPolicy` `Critic Regularized Regression <https://arxiv.org/pdf/2006.15134.pdf>`_

examples/offline/README.md

@@ -35,6 +35,24 @@ Tianshou provides an `offline_trainer` for offline reinforcement learning. You c
 | HalfCheetah-v2        | halfcheetah-expert-v2 | 2864.37         | `python3 d4rl_cql.py --task HalfCheetah-v2 --expert-data-task halfcheetah-expert-v2` |
 | HalfCheetah-v2        | halfcheetah-medium-v2 | 6505.41         | `python3 d4rl_cql.py --task HalfCheetah-v2 --expert-data-task halfcheetah-medium-v2` |
 
+### TD3+BC
+
+| Environment           | Dataset               | CQL             | Parameters                                               |
+| --------------------- | --------------------- | --------------- | -------------------------------------------------------- |
+| HalfCheetah-v2        | halfcheetah-expert-v2 | 11788.25         | `python3 d4rl_td3_bc.py --task HalfCheetah-v2 --expert-data-task halfcheetah-expert-v2` |
+| HalfCheetah-v2        | halfcheetah-medium-v2 | 5741.13         | `python3 d4rl_td3_bc.py --task HalfCheetah-v2 --expert-data-task halfcheetah-medium-v2` |
+
+#### Observation normalization
+
+Following the original paper, we use observation normalization by default. You can turn it off by setting `--norm-obs 0`. The difference are small but consistent.
+
+| Dataset | w/ norm-obs | w/o norm-obs |
+| :--- | :--- | :--- |
+| halfcheeta-medium-v2 | 5741.13 | 5724.41 |
+| halfcheeta-expert-v2 | 11788.25 | 11665.77 |
+| walker2d-medium-v2 | 4051.76 | 3985.59 |
+| walker2d-expert-v2 | 5068.15 | 5027.75 |
+
 ## Discrete control
 
 For discrete control, we currently use ad hoc Atari data generated from a trained QRDQN agent.

examples/offline/d4rl_td3_bc.py

@@ -0,0 +1,227 @@
+#!/usr/bin/env python3
+
+import argparse
+import datetime
+import os
+import pprint
+
+import gym
+import numpy as np
+import torch
+from torch.utils.tensorboard import SummaryWriter
+
+from examples.offline.utils import load_buffer_d4rl, normalize_all_obs_in_replay_buffer
+from tianshou.data import Collector
+from tianshou.env import SubprocVectorEnv, VectorEnvNormObs
+from tianshou.exploration import GaussianNoise
+from tianshou.policy import TD3BCPolicy
+from tianshou.trainer import offline_trainer
+from tianshou.utils import TensorboardLogger, WandbLogger
+from tianshou.utils.net.common import Net
+from tianshou.utils.net.continuous import Actor, Critic
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--task", type=str, default="HalfCheetah-v2")
+    parser.add_argument("--seed", type=int, default=0)
+    parser.add_argument(
+        "--expert-data-task", type=str, default="halfcheetah-expert-v2"
+    )
+    parser.add_argument("--buffer-size", type=int, default=1000000)
+    parser.add_argument("--hidden-sizes", type=int, nargs="*", default=[256, 256])
+    parser.add_argument("--actor-lr", type=float, default=3e-4)
+    parser.add_argument("--critic-lr", type=float, default=3e-4)
+    parser.add_argument("--epoch", type=int, default=200)
+    parser.add_argument("--step-per-epoch", type=int, default=5000)
+    parser.add_argument("--n-step", type=int, default=3)
+    parser.add_argument("--batch-size", type=int, default=256)
+
+    parser.add_argument("--alpha", type=float, default=2.5)
+    parser.add_argument("--exploration-noise", type=float, default=0.1)
+    parser.add_argument("--policy-noise", type=float, default=0.2)
+    parser.add_argument("--noise-clip", type=float, default=0.5)
+    parser.add_argument("--update-actor-freq", type=int, default=2)
+    parser.add_argument("--tau", type=float, default=0.005)
+    parser.add_argument("--gamma", type=float, default=0.99)
+    parser.add_argument("--norm-obs", type=int, default=1)
+
+    parser.add_argument("--eval-freq", type=int, default=1)
+    parser.add_argument("--test-num", type=int, default=10)
+    parser.add_argument("--logdir", type=str, default="log")
+    parser.add_argument("--render", type=float, default=1 / 35)
+    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-id", type=str, default=None)
+    parser.add_argument(
+        "--logger",
+        type=str,
+        default="tensorboard",
+        choices=["tensorboard", "wandb"],
+    )
+    parser.add_argument("--wandb-project", type=str, default="offline_d4rl.benchmark")
+    parser.add_argument(
+        "--watch",
+        default=False,
+        action="store_true",
+        help="watch the play of pre-trained policy only",
+    )
+    return parser.parse_args()
+
+
+def test_td3_bc():
+    args = get_args()
+    env = 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]  # float
+    print("device:", args.device)
+    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))
+
+    args.state_dim = args.state_shape[0]
+    args.action_dim = args.action_shape[0]
+    print("Max_action", args.max_action)
+
+    test_envs = SubprocVectorEnv(
+        [lambda: gym.make(args.task) for _ in range(args.test_num)]
+    )
+    if args.norm_obs:
+        test_envs = VectorEnvNormObs(test_envs, update_obs_rms=False)
+
+    # seed
+    np.random.seed(args.seed)
+    torch.manual_seed(args.seed)
+    test_envs.seed(args.seed)
+
+    # model
+    # actor network
+    net_a = Net(
+        args.state_shape,
+        hidden_sizes=args.hidden_sizes,
+        device=args.device,
+    )
+    actor = Actor(
+        net_a,
+        action_shape=args.action_shape,
+        max_action=args.max_action,
+        device=args.device,
+    ).to(args.device)
+    actor_optim = torch.optim.Adam(actor.parameters(), lr=args.actor_lr)
+
+    # critic network
+    net_c1 = Net(
+        args.state_shape,
+        args.action_shape,
+        hidden_sizes=args.hidden_sizes,
+        concat=True,
+        device=args.device,
+    )
+    net_c2 = Net(
+        args.state_shape,
+        args.action_shape,
+        hidden_sizes=args.hidden_sizes,
+        concat=True,
+        device=args.device,
+    )
+    critic1 = Critic(net_c1, device=args.device).to(args.device)
+    critic1_optim = torch.optim.Adam(critic1.parameters(), lr=args.critic_lr)
+    critic2 = Critic(net_c2, device=args.device).to(args.device)
+    critic2_optim = torch.optim.Adam(critic2.parameters(), lr=args.critic_lr)
+
+    policy = TD3BCPolicy(
+        actor,
+        actor_optim,
+        critic1,
+        critic1_optim,
+        critic2,
+        critic2_optim,
+        tau=args.tau,
+        gamma=args.gamma,
+        exploration_noise=GaussianNoise(sigma=args.exploration_noise),
+        policy_noise=args.policy_noise,
+        update_actor_freq=args.update_actor_freq,
+        noise_clip=args.noise_clip,
+        alpha=args.alpha,
+        estimation_step=args.n_step,
+        action_space=env.action_space,
+    )
+
+    # 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
+    test_collector = Collector(policy, test_envs)
+
+    # log
+    now = datetime.datetime.now().strftime("%y%m%d-%H%M%S")
+    args.algo_name = "td3_bc"
+    log_name = os.path.join(args.task, args.algo_name, str(args.seed), now)
+    log_path = os.path.join(args.logdir, log_name)
+
+    # logger
+    if args.logger == "wandb":
+        logger = WandbLogger(
+            save_interval=1,
+            name=log_name.replace(os.path.sep, "__"),
+            run_id=args.resume_id,
+            config=args,
+            project=args.wandb_project,
+        )
+    writer = SummaryWriter(log_path)
+    writer.add_text("args", str(args))
+    if args.logger == "tensorboard":
+        logger = TensorboardLogger(writer)
+    else:  # wandb
+        logger.load(writer)
+
+    def save_best_fn(policy):
+        torch.save(policy.state_dict(), os.path.join(log_path, "policy.pth"))
+
+    def watch():
+        if args.resume_path is None:
+            args.resume_path = os.path.join(log_path, "policy.pth")
+
+        policy.load_state_dict(
+            torch.load(args.resume_path, map_location=torch.device("cpu"))
+        )
+        policy.eval()
+        collector = Collector(policy, env)
+        collector.collect(n_episode=1, render=1 / 35)
+
+    if not args.watch:
+        replay_buffer = load_buffer_d4rl(args.expert_data_task)
+        if args.norm_obs:
+            replay_buffer, obs_rms = normalize_all_obs_in_replay_buffer(replay_buffer)
+            test_envs.set_obs_rms(obs_rms)
+        # trainer
+        result = offline_trainer(
+            policy,
+            replay_buffer,
+            test_collector,
+            args.epoch,
+            args.step_per_epoch,
+            args.test_num,
+            args.batch_size,
+            save_best_fn=save_best_fn,
+            logger=logger,
+        )
+        pprint.pprint(result)
+    else:
+        watch()
+
+    # Let's watch its performance!
+    policy.eval()
+    test_envs.seed(args.seed)
+    test_collector.reset()
+    result = test_collector.collect(n_episode=args.test_num, render=args.render)
+    print(f"Final reward: {result['rews'].mean()}, length: {result['lens'].mean()}")
+
+
+if __name__ == "__main__":
+    test_td3_bc()

examples/offline/utils.py

@@ -1,8 +1,12 @@
+from typing import Tuple
+
 import d4rl
 import gym
 import h5py
+import numpy as np
 
 from tianshou.data import ReplayBuffer
+from tianshou.utils import RunningMeanStd
 
 
 def load_buffer_d4rl(expert_data_task: str) -> ReplayBuffer:
@@ -27,3 +31,18 @@ def load_buffer(buffer_path: str) -> ReplayBuffer:
             obs_next=dataset["next_observations"]
         )
     return buffer
+
+
+def normalize_all_obs_in_replay_buffer(
+    replay_buffer: ReplayBuffer
+) -> Tuple[ReplayBuffer, RunningMeanStd]:
+    # compute obs mean and var
+    obs_rms = RunningMeanStd()
+    obs_rms.update(replay_buffer.obs)
+    _eps = np.finfo(np.float32).eps.item()
+    # normalize obs
+    replay_buffer._meta["obs"] = (replay_buffer.obs -
+                                  obs_rms.mean) / np.sqrt(obs_rms.var + _eps)
+    replay_buffer._meta["obs_next"] = (replay_buffer.obs_next -
+                                       obs_rms.mean) / np.sqrt(obs_rms.var + _eps)
+    return replay_buffer, obs_rms