Add offline trainer and discrete BCQ algorithm

.gitignore

@@ -145,3 +145,4 @@ MUJOCO_LOG.TXT
 *.zip
 *.pstats
 *.swp
+*.pkl

README.md

@@ -31,6 +31,7 @@
 - [Soft Actor-Critic (SAC)](https://arxiv.org/pdf/1812.05905.pdf)
 - [Discrete Soft Actor-Critic (SAC-Discrete)](https://arxiv.org/pdf/1910.07207.pdf)
 - Vanilla Imitation Learning
+- [Discrete Batch-Constrained deep Q-Learning (BCQ-Discrete)](https://arxiv.org/pdf/1910.01708.pdf)
 - [Prioritized Experience Replay (PER)](https://arxiv.org/pdf/1511.05952.pdf)
 - [Generalized Advantage Estimator (GAE)](https://arxiv.org/pdf/1506.02438.pdf)
 - [Posterior Sampling Reinforcement Learning (PSRL)](https://www.ece.uvic.ca/~bctill/papers/learning/Strens_2000.pdf)

docs/index.rst

@@ -20,8 +20,9 @@ Welcome to Tianshou!
 * :class:`~tianshou.policy.TD3Policy` `Twin Delayed DDPG <https://arxiv.org/pdf/1802.09477.pdf>`_
 * :class:`~tianshou.policy.SACPolicy` `Soft Actor-Critic <https://arxiv.org/pdf/1812.05905.pdf>`_
 * :class:`~tianshou.policy.DiscreteSACPolicy` `Discrete Soft Actor-Critic <https://arxiv.org/pdf/1910.07207.pdf>`_
-* :class:`~tianshou.policy.PSRLPolicy` `Posterior Sampling Reinforcement Learning <https://www.ece.uvic.ca/~bctill/papers/learning/Strens_2000.pdf>`_
 * :class:`~tianshou.policy.ImitationPolicy` Imitation Learning
+* :class:`~tianshou.policy.DiscreteBCQPolicy` `Discrete Batch-Constrained deep Q-Learning <https://arxiv.org/pdf/1910.01708.pdf>`_
+* :class:`~tianshou.policy.PSRLPolicy` `Posterior Sampling Reinforcement Learning <https://www.ece.uvic.ca/~bctill/papers/learning/Strens_2000.pdf>`_
 * :class:`~tianshou.data.PrioritizedReplayBuffer` `Prioritized Experience Replay <https://arxiv.org/pdf/1511.05952.pdf>`_
 * :meth:`~tianshou.policy.BasePolicy.compute_episodic_return` `Generalized Advantage Estimator <https://arxiv.org/pdf/1506.02438.pdf>`_
 

docs/tutorials/concepts.rst

@@ -201,7 +201,7 @@ Trainer
 
 Once you have a collector and a policy, you can start writing the training method for your RL agent. Trainer, to be honest, is a simple wrapper. It helps you save energy for writing the training loop. You can also construct your own trainer: :ref:`customized_trainer`.
 
-Tianshou has two types of trainer: :func:`~tianshou.trainer.onpolicy_trainer` and :func:`~tianshou.trainer.offpolicy_trainer`, corresponding to on-policy algorithms (such as Policy Gradient) and off-policy algorithms (such as DQN). Please check out :doc:`/api/tianshou.trainer` for the usage.
+Tianshou has three types of trainer: :func:`~tianshou.trainer.onpolicy_trainer` for on-policy algorithms such as Policy Gradient, :func:`~tianshou.trainer.offpolicy_trainer` for off-policy algorithms such as DQN, and :func:`~tianshou.trainer.offline_trainer` for offline algorithms such as BCQ. Please check out :doc:`/api/tianshou.trainer` for the usage.
 
 
 .. _pseudocode:

docs/tutorials/dqn.rst

@@ -120,7 +120,7 @@ In each step, the collector will let the policy perform (at least) a specified n
 Train Policy with a Trainer
 ---------------------------
 
-Tianshou provides :class:`~tianshou.trainer.onpolicy_trainer` and :class:`~tianshou.trainer.offpolicy_trainer`. The trainer will automatically stop training when the policy reach the stop condition ``stop_fn`` on test collector. Since DQN is an off-policy algorithm, we use the :class:`~tianshou.trainer.offpolicy_trainer` as follows:
+Tianshou provides :func:`~tianshou.trainer.onpolicy_trainer`, :func:`~tianshou.trainer.offpolicy_trainer`, and :func:`~tianshou.trainer.offline_trainer`. The trainer will automatically stop training when the policy reach the stop condition ``stop_fn`` on test collector. Since DQN is an off-policy algorithm, we use the :func:`~tianshou.trainer.offpolicy_trainer` as follows:
 ::
 
     result = ts.trainer.offpolicy_trainer(
@@ -133,7 +133,7 @@ Tianshou provides :class:`~tianshou.trainer.onpolicy_trainer` and :class:`~tians
         writer=None)
     print(f'Finished training! Use {result["duration"]}')
 
-The meaning of each parameter is as follows (full description can be found at :meth:`~tianshou.trainer.offpolicy_trainer`):
+The meaning of each parameter is as follows (full description can be found at :func:`~tianshou.trainer.offpolicy_trainer`):
 
 * ``max_epoch``: The maximum of epochs for training. The training process might be finished before reaching the ``max_epoch``;
 * ``step_per_epoch``: The number of step for updating policy network in one epoch;

setup.py

@@ -51,7 +51,7 @@ def get_version() -> str:
         "tensorboard",
         "torch>=1.4.0",
         "numba>=0.51.0",
-        "h5py>=3.1.0"
+        "h5py>=2.10.0",  # to match tensorflow's minimal requirements
     ],
     extras_require={
         "dev": [

test/discrete/test_dqn.py

@@ -1,6 +1,7 @@
 import os
 import gym
 import torch
+import pickle
 import pprint
 import argparse
 import numpy as np
@@ -36,6 +37,9 @@ def get_args():
     parser.add_argument('--prioritized-replay', type=int, default=0)
     parser.add_argument('--alpha', type=float, default=0.6)
     parser.add_argument('--beta', type=float, default=0.4)
+    parser.add_argument(
+        '--save-buffer-name', type=str,
+        default="./expert_DQN_CartPole-v0.pkl")
     parser.add_argument(
         '--device', type=str,
         default='cuda' if torch.cuda.is_available() else 'cpu')
@@ -110,6 +114,7 @@ def test_fn(epoch, env_step):
         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!
@@ -120,6 +125,12 @@ def test_fn(epoch, env_step):
         result = collector.collect(n_episode=1, render=args.render)
         print(f'Final reward: {result["rew"]}, length: {result["len"]}')
 
+    # save buffer in pickle format, for imitation learning unittest
+    buf = ReplayBuffer(args.buffer_size)
+    collector = Collector(policy, test_envs, buf)
+    collector.collect(n_step=args.buffer_size)
+    pickle.dump(buf, open(args.save_buffer_name, "wb"))
+
 
 def test_pdqn(args=get_args()):
     args.prioritized_replay = 1

test/discrete/test_il_bcq.py

@@ -0,0 +1,114 @@
+import os
+import gym
+import torch
+import pickle
+import pprint
+import argparse
+import numpy as np
+from torch.utils.tensorboard import SummaryWriter
+
+from tianshou.data import Collector
+from tianshou.env import DummyVectorEnv
+from tianshou.utils.net.common import Net
+from tianshou.trainer import offline_trainer
+from tianshou.policy import DiscreteBCQPolicy
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--task", type=str, default="CartPole-v0")
+    parser.add_argument("--seed", type=int, default=1626)
+    parser.add_argument("--eps-test", type=float, default=0.001)
+    parser.add_argument("--lr", type=float, default=3e-4)
+    parser.add_argument("--gamma", type=float, default=0.9)
+    parser.add_argument("--n-step", type=int, default=3)
+    parser.add_argument("--target-update-freq", type=int, default=320)
+    parser.add_argument("--unlikely-action-threshold", type=float, default=0.3)
+    parser.add_argument("--imitation-logits-penalty", type=float, default=0.01)
+    parser.add_argument("--epoch", type=int, default=5)
+    parser.add_argument("--step-per-epoch", type=int, default=1000)
+    parser.add_argument("--batch-size", type=int, default=64)
+    parser.add_argument("--layer-num", type=int, default=2)
+    parser.add_argument("--hidden-layer-size", type=int, default=128)
+    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(
+        "--load-buffer-name", type=str,
+        default="./expert_DQN_CartPole-v0.pkl",
+    )
+    parser.add_argument(
+        "--device", type=str,
+        default="cuda" if torch.cuda.is_available() else "cpu",
+    )
+    args = parser.parse_known_args()[0]
+    return args
+
+
+def test_discrete_bcq(args=get_args()):
+    # envs
+    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
+    test_envs = DummyVectorEnv(
+        [lambda: gym.make(args.task) for _ in range(args.test_num)])
+    # seed
+    np.random.seed(args.seed)
+    torch.manual_seed(args.seed)
+    test_envs.seed(args.seed)
+    # model
+    policy_net = Net(
+        args.layer_num, args.state_shape, args.action_shape, args.device,
+        hidden_layer_size=args.hidden_layer_size,
+    ).to(args.device)
+    imitation_net = Net(
+        args.layer_num, args.state_shape, args.action_shape, args.device,
+        hidden_layer_size=args.hidden_layer_size,
+    ).to(args.device)
+    optim = torch.optim.Adam(
+        list(policy_net.parameters()) + list(imitation_net.parameters()),
+        lr=args.lr
+    )
+
+    policy = DiscreteBCQPolicy(
+        policy_net, imitation_net, optim, args.gamma, args.n_step,
+        args.target_update_freq, args.eps_test,
+        args.unlikely_action_threshold, args.imitation_logits_penalty,
+    )
+    # buffer
+    assert os.path.exists(args.load_buffer_name), \
+        "Please run test_dqn.py first to get expert's data buffer."
+    buffer = pickle.load(open(args.load_buffer_name, "rb"))
+
+    # collector
+    test_collector = Collector(policy, test_envs)
+
+    log_path = os.path.join(args.logdir, args.task, 'discrete_bcq')
+    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
+
+    result = offline_trainer(
+        policy, buffer, test_collector,
+        args.epoch, args.step_per_epoch, 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!
+        env = gym.make(args.task)
+        policy.eval()
+        policy.set_eps(args.eps_test)
+        collector = Collector(policy, env)
+        result = collector.collect(n_episode=1, render=args.render)
+        print(f'Final reward: {result["rew"]}, length: {result["len"]}')
+
+
+if __name__ == "__main__":
+    test_discrete_bcq(get_args())

tianshou/__init__.py

@@ -1,7 +1,7 @@
 from tianshou import data, env, utils, policy, trainer, exploration
 
 
-__version__ = "0.3.0"
+__version__ = "0.3.1"
 
 __all__ = [
     "env",

tianshou/policy/__init__.py

@@ -10,6 +10,7 @@
 from tianshou.policy.modelfree.td3 import TD3Policy
 from tianshou.policy.modelfree.sac import SACPolicy
 from tianshou.policy.modelfree.discrete_sac import DiscreteSACPolicy
+from tianshou.policy.imitation.discrete_bcq import DiscreteBCQPolicy
 from tianshou.policy.modelbase.psrl import PSRLPolicy
 from tianshou.policy.multiagent.mapolicy import MultiAgentPolicyManager
 
@@ -27,6 +28,7 @@
     "TD3Policy",
     "SACPolicy",
     "DiscreteSACPolicy",
+    "DiscreteBCQPolicy",
     "PSRLPolicy",
     "MultiAgentPolicyManager",
 ]

tianshou/policy/imitation/base.py

@@ -57,7 +57,7 @@ def learn(self, batch: Batch, **kwargs: Any) -> Dict[str, float]:
             a_ = to_torch(batch.act, dtype=torch.float32, device=a.device)
             loss = F.mse_loss(a, a_)  # type: ignore
         elif self.mode == "discrete":  # classification
-            a = self(batch).logits
+            a = F.log_softmax(self(batch).logits, dim=-1)
             a_ = to_torch(batch.act, dtype=torch.long, device=a.device)
             loss = F.nll_loss(a, a_)  # type: ignore
         loss.backward()

tianshou/policy/imitation/discrete_bcq.py

@@ -0,0 +1,133 @@
+import math
+import torch
+import numpy as np
+import torch.nn.functional as F
+from typing import Any, Dict, Union, Optional
+
+from tianshou.policy import DQNPolicy
+from tianshou.data import Batch, ReplayBuffer, to_torch
+
+
+class DiscreteBCQPolicy(DQNPolicy):
+    """Implementation of discrete BCQ algorithm. arXiv:1910.01708.
+
+    :param torch.nn.Module model: a model following the rules in
+        :class:`~tianshou.policy.BasePolicy`. (s -> q_value)
+    :param torch.nn.Module imitator: a model following the rules in
+        :class:`~tianshou.policy.BasePolicy`. (s -> imtation_logits)
+    :param torch.optim.Optimizer optim: a torch.optim for optimizing the model.
+    :param float discount_factor: in [0, 1].
+    :param int estimation_step: greater than 1, the number of steps to look
+        ahead.
+    :param int target_update_freq: the target network update frequency.
+    :param float eval_eps: the epsilon-greedy noise added in evaluation.
+    :param float unlikely_action_threshold: the threshold (tau) for unlikely
+        actions, as shown in Equ. (17) in the paper, defaults to 0.3.
+    :param float imitation_logits_penalty: reguralization weight for imitation
+        logits, defaults to 1e-2.
+    :param bool reward_normalization: normalize the reward to Normal(0, 1),
+        defaults to False.
+
+    .. seealso::
+
+        Please refer to :class:`~tianshou.policy.BasePolicy` for more detailed
+        explanation.
+    """
+
+    def __init__(
+        self,
+        model: torch.nn.Module,
+        imitator: torch.nn.Module,
+        optim: torch.optim.Optimizer,
+        discount_factor: float = 0.99,
+        estimation_step: int = 1,
+        target_update_freq: int = 8000,
+        eval_eps: float = 1e-3,
+        unlikely_action_threshold: float = 0.3,
+        imitation_logits_penalty: float = 1e-2,
+        reward_normalization: bool = False,
+        **kwargs: Any,
+    ) -> None:
+        super().__init__(model, optim, discount_factor, estimation_step,
+                         target_update_freq, reward_normalization, **kwargs)
+        assert target_update_freq > 0, "BCQ needs target network setting."
+        self.imitator = imitator
+        assert (
+            0.0 <= unlikely_action_threshold < 1.0
+        ), "unlikely_action_threshold should be in [0, 1)"
+        self._log_tau = math.log(unlikely_action_threshold)
+        assert 0.0 <= eval_eps < 1.0
+        self._eps = eval_eps
+        self._weight_reg = imitation_logits_penalty
+
+    def train(self, mode: bool = True) -> "DiscreteBCQPolicy":
+        self.training = mode
+        self.model.train(mode)
+        self.imitator.train(mode)
+        return self
+
+    def _target_q(
+        self, buffer: ReplayBuffer, indice: np.ndarray
+    ) -> torch.Tensor:
+        batch = buffer[indice]  # batch.obs_next: s_{t+n}
+        # target_Q = Q_old(s_, argmax(Q_new(s_, *)))
+        with torch.no_grad():
+            act = self(batch, input="obs_next", eps=0.0).act
+            target_q, _ = self.model_old(batch.obs_next)
+            target_q = target_q[np.arange(len(act)), act]
+        return target_q
+
+    def forward(  # type: ignore
+        self,
+        batch: Batch,
+        state: Optional[Union[dict, Batch, np.ndarray]] = None,
+        input: str = "obs",
+        eps: Optional[float] = None,
+        **kwargs: Any,
+    ) -> Batch:
+        if eps is None:
+            eps = self._eps
+        obs = batch[input]
+        q_value, state = self.model(obs, state=state, info=batch.info)
+        imitation_logits, _ = self.imitator(obs, state=state, info=batch.info)
+
+        # mask actions for argmax
+        ratio = imitation_logits - imitation_logits.max(
+            dim=-1, keepdim=True).values
+        mask = (ratio < self._log_tau).float()
+        action = (q_value - np.inf * mask).argmax(dim=-1)
+
+        # add eps to act
+        if not np.isclose(eps, 0.0) and np.random.rand() < eps:
+            bsz, action_num = q_value.shape
+            action = np.random.randint(action_num, size=bsz)
+
+        return Batch(act=action, state=state, q_value=q_value,
+                     imitation_logits=imitation_logits)
+
+    def learn(self, batch: Batch, **kwargs: Any) -> Dict[str, float]:
+        if self._iter % self._freq == 0:
+            self.sync_weight()
+        self._iter += 1
+
+        target_q = batch.returns.flatten()
+        result = self(batch, eps=0.0)
+        imitation_logits = result.imitation_logits
+        current_q = result.q_value[np.arange(len(target_q)), batch.act]
+        act = to_torch(batch.act, dtype=torch.long, device=target_q.device)
+        q_loss = F.smooth_l1_loss(current_q, target_q)
+        i_loss = F.nll_loss(
+            F.log_softmax(imitation_logits, dim=-1), act)  # type: ignore
+        reg_loss = imitation_logits.pow(2).mean()
+        loss = q_loss + i_loss + self._weight_reg * reg_loss
+
+        self.optim.zero_grad()
+        loss.backward()
+        self.optim.step()
+
+        return {
+            "loss": loss.item(),
+            "q_loss": q_loss.item(),
+            "i_loss": i_loss.item(),
+            "reg_loss": reg_loss.item(),
+        }