rebrac.py

# source: https://github.com/tinkoff-ai/ReBRAC
# https://arxiv.org/abs/2305.09836

import os

os.environ["TF_CUDNN_DETERMINISTIC"] = "1"  # For reproducibility

import math
import uuid
from copy import deepcopy
from dataclasses import asdict, dataclass
from functools import partial
from typing import Any, Callable, Dict, Sequence, Tuple, Union

import chex
import d4rl  # noqa
import flax.linen as nn
import gym
import jax
import jax.numpy as jnp
import numpy as np
import optax
import pyrallis
import wandb
from flax.core import FrozenDict
from flax.training.train_state import TrainState
from tqdm.auto import trange

default_kernel_init = nn.initializers.lecun_normal()
default_bias_init = nn.initializers.zeros


@dataclass
class Config:
    # wandb params
    project: str = "CORL"
    group: str = "rebrac"
    name: str = "rebrac"
    # model params
    actor_learning_rate: float = 1e-3
    critic_learning_rate: float = 1e-3
    hidden_dim: int = 256
    actor_n_hiddens: int = 3
    critic_n_hiddens: int = 3
    gamma: float = 0.99
    tau: float = 5e-3
    actor_bc_coef: float = 1.0
    critic_bc_coef: float = 1.0
    actor_ln: bool = False
    critic_ln: bool = True
    policy_noise: float = 0.2
    noise_clip: float = 0.5
    policy_freq: int = 2
    normalize_q: bool = True
    # training params
    dataset_name: str = "halfcheetah-medium-v2"
    batch_size: int = 1024
    num_epochs: int = 1000
    num_updates_on_epoch: int = 1000
    normalize_reward: bool = False
    normalize_states: bool = False
    # evaluation params
    eval_episodes: int = 10
    eval_every: int = 5
    # general params
    train_seed: int = 0
    eval_seed: int = 42

    def __post_init__(self):
        self.name = f"{self.name}-{self.dataset_name}-{str(uuid.uuid4())[:8]}"


def pytorch_init(fan_in: float) -> Callable:
    """
    Default init for PyTorch Linear layer weights and biases:
    https://pytorch.org/docs/stable/generated/torch.nn.Linear.html
    """
    bound = math.sqrt(1 / fan_in)

    def _init(key: jax.random.PRNGKey, shape: Tuple, dtype: type) -> jax.Array:
        return jax.random.uniform(
            key, shape=shape, minval=-bound, maxval=bound, dtype=dtype
        )

    return _init


def uniform_init(bound: float) -> Callable:
    def _init(key: jax.random.PRNGKey, shape: Tuple, dtype: type) -> jax.Array:
        return jax.random.uniform(
            key, shape=shape, minval=-bound, maxval=bound, dtype=dtype
        )

    return _init


def identity(x: Any) -> Any:
    return x


class DetActor(nn.Module):
    action_dim: int
    hidden_dim: int = 256
    layernorm: bool = True
    n_hiddens: int = 3

    @nn.compact
    def __call__(self, state: jax.Array) -> jax.Array:
        s_d, h_d = state.shape[-1], self.hidden_dim
        # Initialization as in the EDAC paper
        layers = [
            nn.Dense(
                self.hidden_dim,
                kernel_init=pytorch_init(s_d),
                bias_init=nn.initializers.constant(0.1),
            ),
            nn.relu,
            nn.LayerNorm() if self.layernorm else identity,
        ]
        for _ in range(self.n_hiddens - 1):
            layers += [
                nn.Dense(
                    self.hidden_dim,
                    kernel_init=pytorch_init(h_d),
                    bias_init=nn.initializers.constant(0.1),
                ),
                nn.relu,
                nn.LayerNorm() if self.layernorm else identity,
            ]
        layers += [
            nn.Dense(
                self.action_dim,
                kernel_init=uniform_init(1e-3),
                bias_init=uniform_init(1e-3),
            ),
            nn.tanh,
        ]
        net = nn.Sequential(layers)
        actions = net(state)
        return actions


class Critic(nn.Module):
    hidden_dim: int = 256
    layernorm: bool = True
    n_hiddens: int = 3

    @nn.compact
    def __call__(self, state: jax.Array, action: jax.Array) -> jax.Array:
        s_d, a_d, h_d = state.shape[-1], action.shape[-1], self.hidden_dim
        # Initialization as in the EDAC paper
        layers = [
            nn.Dense(
                self.hidden_dim,
                kernel_init=pytorch_init(s_d + a_d),
                bias_init=nn.initializers.constant(0.1),
            ),
            nn.relu,
            nn.LayerNorm() if self.layernorm else identity,
        ]
        for _ in range(self.n_hiddens - 1):
            layers += [
                nn.Dense(
                    self.hidden_dim,
                    kernel_init=pytorch_init(h_d),
                    bias_init=nn.initializers.constant(0.1),
                ),
                nn.relu,
                nn.LayerNorm() if self.layernorm else identity,
            ]
        layers += [
            nn.Dense(1, kernel_init=uniform_init(3e-3), bias_init=uniform_init(3e-3))
        ]
        network = nn.Sequential(layers)
        state_action = jnp.hstack([state, action])
        out = network(state_action).squeeze(-1)
        return out


class EnsembleCritic(nn.Module):
    hidden_dim: int = 256
    num_critics: int = 10
    layernorm: bool = True
    n_hiddens: int = 3

    @nn.compact
    def __call__(self, state: jax.Array, action: jax.Array) -> jax.Array:
        ensemble = nn.vmap(
            target=Critic,
            in_axes=None,
            out_axes=0,
            variable_axes={"params": 0},
            split_rngs={"params": True},
            axis_size=self.num_critics,
        )
        q_values = ensemble(self.hidden_dim, self.layernorm, self.n_hiddens)(
            state, action
        )
        return q_values


def qlearning_dataset(
    env: gym.Env,
    dataset: Dict = None,
    terminate_on_end: bool = False,
    **kwargs,
) -> Dict:
    if dataset is None:
        dataset = env.get_dataset(**kwargs)

    N = dataset["rewards"].shape[0]
    obs_ = []
    next_obs_ = []
    action_ = []
    next_action_ = []
    reward_ = []
    done_ = []

    # The newer version of the dataset adds an explicit
    # timeouts field. Keep old method for backwards compatability.
    use_timeouts = "timeouts" in dataset

    episode_step = 0
    for i in range(N - 1):
        obs = dataset["observations"][i].astype(np.float32)
        new_obs = dataset["observations"][i + 1].astype(np.float32)
        action = dataset["actions"][i].astype(np.float32)
        new_action = dataset["actions"][i + 1].astype(np.float32)
        reward = dataset["rewards"][i].astype(np.float32)
        done_bool = bool(dataset["terminals"][i])

        if use_timeouts:
            final_timestep = dataset["timeouts"][i]
        else:
            final_timestep = episode_step == env._max_episode_steps - 1
        if (not terminate_on_end) and final_timestep:
            # Skip this transition
            episode_step = 0
            continue
        if done_bool or final_timestep:
            episode_step = 0

        obs_.append(obs)
        next_obs_.append(new_obs)
        action_.append(action)
        next_action_.append(new_action)
        reward_.append(reward)
        done_.append(done_bool)
        episode_step += 1

    return {
        "observations": np.array(obs_),
        "actions": np.array(action_),
        "next_observations": np.array(next_obs_),
        "next_actions": np.array(next_action_),
        "rewards": np.array(reward_),
        "terminals": np.array(done_),
    }


def compute_mean_std(states: jax.Array, eps: float) -> Tuple[jax.Array, jax.Array]:
    mean = states.mean(0)
    std = states.std(0) + eps
    return mean, std


def normalize_states(states: jax.Array, mean: jax.Array, std: jax.Array) -> jax.Array:
    return (states - mean) / std


@chex.dataclass
class ReplayBuffer:
    data: Dict[str, jax.Array] = None
    mean: float = 0
    std: float = 1

    def create_from_d4rl(
        self,
        dataset_name: str,
        normalize_reward: bool = False,
        is_normalize: bool = False,
    ):
        d4rl_data = qlearning_dataset(gym.make(dataset_name))
        buffer = {
            "states": jnp.asarray(d4rl_data["observations"], dtype=jnp.float32),
            "actions": jnp.asarray(d4rl_data["actions"], dtype=jnp.float32),
            "rewards": jnp.asarray(d4rl_data["rewards"], dtype=jnp.float32),
            "next_states": jnp.asarray(
                d4rl_data["next_observations"], dtype=jnp.float32
            ),
            "next_actions": jnp.asarray(d4rl_data["next_actions"], dtype=jnp.float32),
            "dones": jnp.asarray(d4rl_data["terminals"], dtype=jnp.float32),
        }
        if is_normalize:
            self.mean, self.std = compute_mean_std(buffer["states"], eps=1e-3)
            buffer["states"] = normalize_states(buffer["states"], self.mean, self.std)
            buffer["next_states"] = normalize_states(
                buffer["next_states"], self.mean, self.std
            )
        if normalize_reward:
            buffer["rewards"] = ReplayBuffer.normalize_reward(
                dataset_name, buffer["rewards"]
            )
        self.data = buffer

    @property
    def size(self) -> int:
        # WARN: It will use len of the dataclass, i.e. number of fields.
        return self.data["states"].shape[0]

    def sample_batch(
        self, key: jax.random.PRNGKey, batch_size: int
    ) -> Dict[str, jax.Array]:
        indices = jax.random.randint(
            key, shape=(batch_size,), minval=0, maxval=self.size
        )
        batch = jax.tree_map(lambda arr: arr[indices], self.data)
        return batch

    def get_moments(self, modality: str) -> Tuple[jax.Array, jax.Array]:
        mean = self.data[modality].mean(0)
        std = self.data[modality].std(0)
        return mean, std

    @staticmethod
    def normalize_reward(dataset_name: str, rewards: jax.Array) -> jax.Array:
        if "antmaze" in dataset_name:
            return rewards * 100.0  # like in LAPO
        else:
            raise NotImplementedError(
                "Reward normalization is implemented only for AntMaze yet!"
            )


@chex.dataclass(frozen=True)
class Metrics:
    accumulators: Dict[str, Tuple[jax.Array, jax.Array]]

    @staticmethod
    def create(metrics: Sequence[str]) -> "Metrics":
        init_metrics = {key: (jnp.array([0.0]), jnp.array([0.0])) for key in metrics}
        return Metrics(accumulators=init_metrics)

    def update(self, updates: Dict[str, jax.Array]) -> "Metrics":
        new_accumulators = deepcopy(self.accumulators)
        for key, value in updates.items():
            acc, steps = new_accumulators[key]
            new_accumulators[key] = (acc + value, steps + 1)

        return self.replace(accumulators=new_accumulators)

    def compute(self) -> Dict[str, np.ndarray]:
        # cumulative_value / total_steps
        return {k: np.array(v[0] / v[1]) for k, v in self.accumulators.items()}


def normalize(
    arr: jax.Array, mean: jax.Array, std: jax.Array, eps: float = 1e-8
) -> jax.Array:
    return (arr - mean) / (std + eps)


def make_env(env_name: str, seed: int) -> gym.Env:
    env = gym.make(env_name)
    env.seed(seed)
    env.action_space.seed(seed)
    env.observation_space.seed(seed)
    return env


def wrap_env(
    env: gym.Env,
    state_mean: Union[np.ndarray, float] = 0.0,
    state_std: Union[np.ndarray, float] = 1.0,
    reward_scale: float = 1.0,
) -> gym.Env:
    # PEP 8: E731 do not assign a lambda expression, use a def
    def normalize_state(state: np.ndarray) -> np.ndarray:
        return (
            state - state_mean
        ) / state_std  # epsilon should be already added in std.

    def scale_reward(reward: float) -> float:
        # Please be careful, here reward is multiplied by scale!
        return reward_scale * reward

    env = gym.wrappers.TransformObservation(env, normalize_state)
    if reward_scale != 1.0:
        env = gym.wrappers.TransformReward(env, scale_reward)
    return env


def evaluate(
    env: gym.Env,
    params: jax.Array,
    action_fn: Callable,
    num_episodes: int,
    seed: int,
) -> np.ndarray:
    env.seed(seed)
    env.action_space.seed(seed)
    env.observation_space.seed(seed)

    returns = []
    for _ in trange(num_episodes, desc="Eval", leave=False):
        obs, done = env.reset(), False
        total_reward = 0.0
        while not done:
            action = np.asarray(jax.device_get(action_fn(params, obs)))
            obs, reward, done, _ = env.step(action)
            total_reward += reward
        returns.append(total_reward)

    return np.array(returns)


class CriticTrainState(TrainState):
    target_params: FrozenDict


class ActorTrainState(TrainState):
    target_params: FrozenDict


def update_actor(
    key: jax.random.PRNGKey,
    actor: TrainState,
    critic: TrainState,
    batch: Dict[str, jax.Array],
    beta: float,
    tau: float,
    normalize_q: bool,
    metrics: Metrics,
) -> Tuple[jax.random.PRNGKey, TrainState, TrainState, Metrics]:
    key, random_action_key = jax.random.split(key, 2)

    def actor_loss_fn(params: jax.Array) -> Tuple[jax.Array, Metrics]:
        actions = actor.apply_fn(params, batch["states"])

        bc_penalty = ((actions - batch["actions"]) ** 2).sum(-1)
        q_values = critic.apply_fn(critic.params, batch["states"], actions).min(0)
        lmbda = 1
        if normalize_q:
            lmbda = jax.lax.stop_gradient(1 / jax.numpy.abs(q_values).mean())

        loss = (beta * bc_penalty - lmbda * q_values).mean()

        # logging stuff
        random_actions = jax.random.uniform(
            random_action_key, shape=batch["actions"].shape, minval=-1.0, maxval=1.0
        )
        new_metrics = metrics.update(
            {
                "actor_loss": loss,
                "bc_mse_policy": bc_penalty.mean(),
                "bc_mse_random": ((random_actions - batch["actions"]) ** 2)
                .sum(-1)
                .mean(),
                "action_mse": ((actions - batch["actions"]) ** 2).mean(),
            }
        )
        return loss, new_metrics

    grads, new_metrics = jax.grad(actor_loss_fn, has_aux=True)(actor.params)
    new_actor = actor.apply_gradients(grads=grads)

    new_actor = new_actor.replace(
        target_params=optax.incremental_update(actor.params, actor.target_params, tau)
    )
    new_critic = critic.replace(
        target_params=optax.incremental_update(critic.params, critic.target_params, tau)
    )

    return key, new_actor, new_critic, new_metrics


def update_critic(
    key: jax.random.PRNGKey,
    actor: TrainState,
    critic: CriticTrainState,
    batch: Dict[str, jax.Array],
    gamma: float,
    beta: float,
    tau: float,
    policy_noise: float,
    noise_clip: float,
    metrics: Metrics,
) -> Tuple[jax.random.PRNGKey, TrainState, Metrics]:
    key, actions_key = jax.random.split(key)

    next_actions = actor.apply_fn(actor.target_params, batch["next_states"])
    noise = jax.numpy.clip(
        (jax.random.normal(actions_key, next_actions.shape) * policy_noise),
        -noise_clip,
        noise_clip,
    )
    next_actions = jax.numpy.clip(next_actions + noise, -1, 1)
    bc_penalty = ((next_actions - batch["next_actions"]) ** 2).sum(-1)
    next_q = critic.apply_fn(
        critic.target_params, batch["next_states"], next_actions
    ).min(0)
    next_q = next_q - beta * bc_penalty

    target_q = batch["rewards"] + (1 - batch["dones"]) * gamma * next_q

    def critic_loss_fn(critic_params: jax.Array) -> Tuple[jax.Array, jax.Array]:
        # [N, batch_size] - [1, batch_size]
        q = critic.apply_fn(critic_params, batch["states"], batch["actions"])
        q_min = q.min(0).mean()
        loss = ((q - target_q[None, ...]) ** 2).mean(1).sum(0)
        return loss, q_min

    (loss, q_min), grads = jax.value_and_grad(critic_loss_fn, has_aux=True)(
        critic.params
    )
    new_critic = critic.apply_gradients(grads=grads)
    new_metrics = metrics.update(
        {
            "critic_loss": loss,
            "q_min": q_min,
        }
    )
    return key, new_critic, new_metrics


def update_td3(
    key: jax.random.PRNGKey,
    actor: TrainState,
    critic: CriticTrainState,
    batch: Dict[str, Any],
    metrics: Metrics,
    gamma: float,
    actor_bc_coef: float,
    critic_bc_coef: float,
    tau: float,
    policy_noise: float,
    noise_clip: float,
    normalize_q: bool,
) -> Tuple[jax.random.PRNGKey, TrainState, TrainState, Metrics]:
    key, new_critic, new_metrics = update_critic(
        key,
        actor,
        critic,
        batch,
        gamma,
        critic_bc_coef,
        tau,
        policy_noise,
        noise_clip,
        metrics,
    )
    key, new_actor, new_critic, new_metrics = update_actor(
        key, actor, new_critic, batch, actor_bc_coef, tau, normalize_q, new_metrics
    )
    return key, new_actor, new_critic, new_metrics


def update_td3_no_targets(
    key: jax.random.PRNGKey,
    actor: TrainState,
    critic: CriticTrainState,
    batch: Dict[str, Any],
    gamma: float,
    metrics: Metrics,
    actor_bc_coef: float,
    critic_bc_coef: float,
    tau: float,
    policy_noise: float,
    noise_clip: float,
) -> Tuple[jax.random.PRNGKey, TrainState, TrainState, Metrics]:
    key, new_critic, new_metrics = update_critic(
        key,
        actor,
        critic,
        batch,
        gamma,
        critic_bc_coef,
        tau,
        policy_noise,
        noise_clip,
        metrics,
    )
    return key, actor, new_critic, new_metrics


def action_fn(actor: TrainState) -> Callable:
    @jax.jit
    def _action_fn(obs: jax.Array) -> jax.Array:
        action = actor.apply_fn(actor.params, obs)
        return action

    return _action_fn


@pyrallis.wrap()
def train(config: Config):
    dict_config = asdict(config)
    dict_config["mlc_job_name"] = os.environ.get("PLATFORM_JOB_NAME")

    wandb.init(
        config=dict_config,
        project=config.project,
        group=config.group,
        name=config.name,
        id=str(uuid.uuid4()),
    )
    wandb.mark_preempting()
    buffer = ReplayBuffer()
    buffer.create_from_d4rl(
        config.dataset_name, config.normalize_reward, config.normalize_states
    )

    key = jax.random.PRNGKey(seed=config.train_seed)
    key, actor_key, critic_key = jax.random.split(key, 3)

    eval_env = make_env(config.dataset_name, seed=config.eval_seed)
    eval_env = wrap_env(eval_env, buffer.mean, buffer.std)
    init_state = buffer.data["states"][0][None, ...]
    init_action = buffer.data["actions"][0][None, ...]

    actor_module = DetActor(
        action_dim=init_action.shape[-1],
        hidden_dim=config.hidden_dim,
        layernorm=config.actor_ln,
        n_hiddens=config.actor_n_hiddens,
    )
    actor = ActorTrainState.create(
        apply_fn=actor_module.apply,
        params=actor_module.init(actor_key, init_state),
        target_params=actor_module.init(actor_key, init_state),
        tx=optax.adam(learning_rate=config.actor_learning_rate),
    )

    critic_module = EnsembleCritic(
        hidden_dim=config.hidden_dim,
        num_critics=2,
        layernorm=config.critic_ln,
        n_hiddens=config.critic_n_hiddens,
    )
    critic = CriticTrainState.create(
        apply_fn=critic_module.apply,
        params=critic_module.init(critic_key, init_state, init_action),
        target_params=critic_module.init(critic_key, init_state, init_action),
        tx=optax.adam(learning_rate=config.critic_learning_rate),
    )

    update_td3_partial = partial(
        update_td3,
        gamma=config.gamma,
        actor_bc_coef=config.actor_bc_coef,
        critic_bc_coef=config.critic_bc_coef,
        tau=config.tau,
        policy_noise=config.policy_noise,
        noise_clip=config.noise_clip,
        normalize_q=config.normalize_q,
    )

    update_td3_no_targets_partial = partial(
        update_td3_no_targets,
        gamma=config.gamma,
        actor_bc_coef=config.actor_bc_coef,
        critic_bc_coef=config.critic_bc_coef,
        tau=config.tau,
        policy_noise=config.policy_noise,
        noise_clip=config.noise_clip,
    )

    def td3_loop_update_step(i: int, carry: TrainState):
        key, batch_key = jax.random.split(carry["key"])
        batch = carry["buffer"].sample_batch(batch_key, batch_size=config.batch_size)

        full_update = partial(
            update_td3_partial,
            key=key,
            actor=carry["actor"],
            critic=carry["critic"],
            batch=batch,
            metrics=carry["metrics"],
        )

        update = partial(
            update_td3_no_targets_partial,
            key=key,
            actor=carry["actor"],
            critic=carry["critic"],
            batch=batch,
            metrics=carry["metrics"],
        )

        key, new_actor, new_critic, new_metrics = jax.lax.cond(
            update_carry["delayed_updates"][i], full_update, update
        )

        carry.update(key=key, actor=new_actor, critic=new_critic, metrics=new_metrics)
        return carry

    # metrics
    bc_metrics_to_log = [
        "critic_loss",
        "q_min",
        "actor_loss",
        "batch_entropy",
        "bc_mse_policy",
        "bc_mse_random",
        "action_mse",
    ]
    # shared carry for update loops
    update_carry = {
        "key": key,
        "actor": actor,
        "critic": critic,
        "buffer": buffer,
        "delayed_updates": jax.numpy.equal(
            jax.numpy.arange(config.num_updates_on_epoch) % config.policy_freq, 0
        ).astype(int),
    }

    @jax.jit
    def actor_action_fn(params: jax.Array, obs: jax.Array):
        return actor.apply_fn(params, obs)

    for epoch in trange(config.num_epochs, desc="ReBRAC Epochs"):
        # metrics for accumulation during epoch and logging to wandb
        # we need to reset them every epoch
        update_carry["metrics"] = Metrics.create(bc_metrics_to_log)

        update_carry = jax.lax.fori_loop(
            lower=0,
            upper=config.num_updates_on_epoch,
            body_fun=td3_loop_update_step,
            init_val=update_carry,
        )
        # log mean over epoch for each metric
        mean_metrics = update_carry["metrics"].compute()
        wandb.log(
            {"epoch": epoch, **{f"ReBRAC/{k}": v for k, v in mean_metrics.items()}}
        )

        if epoch % config.eval_every == 0 or epoch == config.num_epochs - 1:
            eval_returns = evaluate(
                eval_env,
                update_carry["actor"].params,
                actor_action_fn,
                config.eval_episodes,
                seed=config.eval_seed,
            )
            normalized_score = eval_env.get_normalized_score(eval_returns) * 100.0
            wandb.log(
                {
                    "epoch": epoch,
                    "eval/return_mean": np.mean(eval_returns),
                    "eval/return_std": np.std(eval_returns),
                    "eval/normalized_score_mean": np.mean(normalized_score),
                    "eval/normalized_score_std": np.std(normalized_score),
                }
            )


if __name__ == "__main__":
    train()