500 lines
17 KiB
Python
500 lines
17 KiB
Python
import datetime
|
|
import random
|
|
import time
|
|
from dataclasses import asdict, dataclass
|
|
from functools import partial
|
|
from typing import Any
|
|
|
|
import jax
|
|
import jax.numpy as jnp
|
|
import numpy as np
|
|
import optax
|
|
from flax.training.train_state import TrainState
|
|
|
|
from experiment_logger import get_logger
|
|
|
|
from brittle_star_project.dataclasses import EpisodeStatistics, PPOArgs
|
|
from brittle_star_project.environment.BrittleStarJaxEnvWrapper import BrittleStarJaxEnvWrapper
|
|
from brittle_star_project.MLPs.mlps import (
|
|
Actor,
|
|
AgentParams,
|
|
GenericDenseLayersWithActivation,
|
|
OneDenseLayerMLP,
|
|
Storage,
|
|
)
|
|
from brittle_star_project.ppo import PPO
|
|
|
|
@jax.jit
|
|
def _clip_action(action: jnp.ndarray, low: jnp.ndarray, high: jnp.ndarray) -> jnp.ndarray:
|
|
return jnp.clip(action, low, high)
|
|
|
|
@jax.jit
|
|
def _linear_schedule(count, minibatch_count, update_epochs, num_iterations, learning_rate):
|
|
frac = 1.0 - (count // (minibatch_count * update_epochs)) / num_iterations
|
|
return learning_rate * frac
|
|
|
|
|
|
@jax.jit
|
|
def _convert_obs_dict_to_array(obs_dict: dict) -> jnp.ndarray:
|
|
return jax.vmap(lambda o: jnp.concatenate([v.flatten() for v in o.values() if v.size > 0]))(
|
|
obs_dict
|
|
)
|
|
|
|
|
|
# removed jit: used in _rollout_jit, so will be compiled with _rollout_jit
|
|
def _get_action_and_value_noise(
|
|
sensor: GenericDenseLayersWithActivation,
|
|
feature_extractor: GenericDenseLayersWithActivation,
|
|
actor: Actor,
|
|
critic: OneDenseLayerMLP,
|
|
agent_state: TrainState,
|
|
next_obs: jnp.ndarray,
|
|
key: jax.random.PRNGKey,
|
|
action_low,
|
|
action_high
|
|
):
|
|
hidden = sensor.apply(agent_state.params["sensor_params"], next_obs)
|
|
hidden_critic = feature_extractor.apply(
|
|
agent_state.params["feature_extractor_params"], next_obs
|
|
)
|
|
|
|
# Continuous actions: sample from a Gaussian parameterized by the actor
|
|
mean, log_std = actor.apply(agent_state.params["actor_params"], hidden)
|
|
key, subkey = jax.random.split(key)
|
|
noise = jax.random.normal(subkey, shape=mean.shape)
|
|
std = jnp.exp(log_std)
|
|
action = mean + noise * std
|
|
clipped_action = _clip_action(
|
|
action,
|
|
action_low,
|
|
action_high
|
|
)
|
|
logprob = -0.5 * (((clipped_action - mean) / std) ** 2 + 2 * log_std + jnp.log(2 * jnp.pi)).sum(-1)
|
|
value = critic.apply(agent_state.params["critic_params"], hidden_critic)
|
|
return clipped_action, logprob, value.squeeze(-1), key
|
|
|
|
|
|
# removed jit: used in _rollout_jit, so will be compiled with _rollout_jit
|
|
def _step_once(
|
|
carry,
|
|
_,
|
|
env_step_fn,
|
|
sensor: GenericDenseLayersWithActivation,
|
|
feature_extractor: GenericDenseLayersWithActivation,
|
|
actor: Actor,
|
|
critic: OneDenseLayerMLP,
|
|
action_low,
|
|
action_high
|
|
):
|
|
agent_state, episode_stats, obs, done, key, env_state = carry
|
|
action, logprob, value, key = _get_action_and_value_noise(
|
|
sensor, feature_extractor, actor, critic, agent_state, obs, key, action_low, action_high
|
|
)
|
|
|
|
episode_stats, env_state, (next_obs, reward, next_done) = env_step_fn(
|
|
episode_stats, env_state, action
|
|
)
|
|
|
|
storage = Storage(
|
|
obs=obs,
|
|
actions=action,
|
|
logprobs=logprob,
|
|
dones=done,
|
|
values=value,
|
|
rewards=reward,
|
|
returns=jnp.zeros_like(reward),
|
|
advantages=jnp.zeros_like(reward),
|
|
)
|
|
return (agent_state, episode_stats, next_obs, next_done, key, env_state), storage
|
|
|
|
|
|
# removed jit: used in _rollout_jit, so will be compiled with _rollout_jit
|
|
def _step_env_wrapped(episode_stats, env_state, action, env_step_fn):
|
|
next_env_state = env_step_fn(env_state, action)
|
|
|
|
# Extract per-environment signals from the state object
|
|
reward = next_env_state.reward # (num_envs,)
|
|
terminated = next_env_state.terminated # (num_envs,)
|
|
truncated = next_env_state.truncated # (num_envs,)
|
|
done = terminated | truncated # (num_envs,)
|
|
|
|
new_episode_return = episode_stats.episode_returns + reward
|
|
new_episode_length = episode_stats.episode_lengths + 1
|
|
|
|
episode_stats = episode_stats.replace(
|
|
episode_returns=new_episode_return * (1 - done),
|
|
episode_lengths=new_episode_length * (1 - done),
|
|
returned_episode_returns=jnp.where(
|
|
done, new_episode_return, episode_stats.returned_episode_returns
|
|
),
|
|
returned_episode_lengths=jnp.where(
|
|
done, new_episode_length, episode_stats.returned_episode_lengths
|
|
),
|
|
)
|
|
return (
|
|
episode_stats,
|
|
next_env_state,
|
|
(_convert_obs_dict_to_array(next_env_state.observations), reward, done),
|
|
)
|
|
|
|
|
|
# jit applied in wrapper method self._rollout_jit using partial
|
|
def _rollout_jit(
|
|
agent_state,
|
|
episode_stats,
|
|
env_state,
|
|
next_obs,
|
|
next_done,
|
|
key,
|
|
max_steps,
|
|
step_env_fn,
|
|
sensor: GenericDenseLayersWithActivation,
|
|
feature_extractor: GenericDenseLayersWithActivation,
|
|
actor: Actor,
|
|
critic: OneDenseLayerMLP,
|
|
action_low,
|
|
action_high
|
|
):
|
|
(agent_state, episode_stats, next_obs, next_done, key, env_state), storage = jax.lax.scan(
|
|
partial(
|
|
_step_once,
|
|
sensor=sensor,
|
|
feature_extractor=feature_extractor,
|
|
actor=actor,
|
|
critic=critic,
|
|
env_step_fn=step_env_fn,
|
|
action_low=action_low,
|
|
action_high=action_high
|
|
),
|
|
(agent_state, episode_stats, next_obs, next_done, key, env_state),
|
|
(),
|
|
max_steps,
|
|
)
|
|
return agent_state, episode_stats, next_obs, next_done, storage, key, env_state
|
|
|
|
|
|
# removed jit: used in _compute_gae_jit, so will be compiled with _compute_gae_jit
|
|
def _compute_gae_once(carry, inp, gamma, gae_lambda):
|
|
advantages = carry
|
|
nextdone, nextvalues, curvalues, reward = inp
|
|
nextnonterminal = 1.0 - nextdone
|
|
delta = reward + gamma * nextvalues * nextnonterminal - curvalues
|
|
advantages = delta + gamma * gae_lambda * nextnonterminal * advantages
|
|
return advantages, advantages
|
|
|
|
|
|
# jit applied on partial-wrapped wrapper method self._compute_gae_jit
|
|
def _compute_gae_jit(
|
|
agent_state, storage, next_obs, next_done, gamma, gae_lambda, num_envs, sensor, critic
|
|
):
|
|
next_value = critic.apply(
|
|
agent_state.params["critic_params"],
|
|
sensor.apply(agent_state.params["sensor_params"], next_obs),
|
|
).squeeze(-1)
|
|
|
|
advantages = jnp.zeros((num_envs,))
|
|
dones = jnp.concatenate([storage.dones, next_done[None, :]], axis=0)
|
|
values = jnp.concatenate([storage.values, next_value[None, :]], axis=0)
|
|
_, advantages = jax.lax.scan(
|
|
partial(_compute_gae_once, gamma=gamma, gae_lambda=gae_lambda),
|
|
advantages,
|
|
(dones[1:], values[1:], values[:-1], storage.rewards),
|
|
reverse=True,
|
|
)
|
|
return storage.replace(advantages=advantages, returns=advantages + storage.values)
|
|
|
|
|
|
@dataclass
|
|
class LossInfo:
|
|
# todo: better typing
|
|
loss: Any
|
|
pg_loss: Any
|
|
v_loss: Any
|
|
entropy_loss: Any
|
|
approx_kl: Any
|
|
avg_episodic_return: Any
|
|
|
|
|
|
class PPOTrainer:
|
|
def __init__(self, args: PPOArgs, env: BrittleStarJaxEnvWrapper, run_dir: str, run_name: str):
|
|
self.args = args
|
|
self.env = env
|
|
self.run_dir = run_dir
|
|
self.run_name = run_name
|
|
self.logger = get_logger()
|
|
|
|
self.key = jax.random.PRNGKey(args.seed)
|
|
|
|
self.sensor, self.feature_extractor, self.actor, self.critic = self._init_agent()
|
|
self.sensor.apply = jax.jit(self.sensor.apply)
|
|
self.feature_extractor.apply = jax.jit(self.feature_extractor.apply)
|
|
self.actor.apply = jax.jit(self.actor.apply)
|
|
self.critic.apply = jax.jit(self.critic.apply)
|
|
|
|
action_low = jnp.asarray(self.env.single_action_space.low, dtype=jnp.float32)
|
|
action_high = jnp.asarray(self.env.single_action_space.high, dtype=jnp.float32)
|
|
|
|
self._rollout_jit = jax.jit(
|
|
partial(
|
|
_rollout_jit,
|
|
max_steps=self.args.num_steps,
|
|
step_env_fn=partial(_step_env_wrapped, env_step_fn=self.env.step),
|
|
sensor=self.sensor,
|
|
feature_extractor=self.feature_extractor,
|
|
actor=self.actor,
|
|
critic=self.critic,
|
|
action_low=action_low,
|
|
action_high=action_high
|
|
)
|
|
)
|
|
self._compute_gae_jit = jax.jit(
|
|
partial(
|
|
_compute_gae_jit,
|
|
num_envs=self.args.num_envs,
|
|
gamma=self.args.gamma,
|
|
gae_lambda=self.args.gae_lambda,
|
|
sensor=self.sensor,
|
|
critic=self.critic,
|
|
)
|
|
)
|
|
|
|
self._ppo = PPO(self.args, self.sensor, self.actor, self.critic, self.feature_extractor)
|
|
|
|
self.agent_state = self._init_agent_state()
|
|
|
|
self.episode_stats = self._init_episode_stats()
|
|
|
|
self._init_random()
|
|
|
|
def _init_random(self):
|
|
self.logger.info(f"[RANDOM]: Setting random seed to {self.args.seed}")
|
|
|
|
random.seed(self.args.seed)
|
|
np.random.seed(self.args.seed)
|
|
|
|
def _init_agent(self):
|
|
self.logger.info("[AGENT]: Initializing agent...")
|
|
|
|
sensor = GenericDenseLayersWithActivation()
|
|
feature_extractor = GenericDenseLayersWithActivation()
|
|
actor = Actor(
|
|
action_dim=self.env.single_action_space.shape[0]
|
|
) # continuous actions for MJX
|
|
critic = OneDenseLayerMLP()
|
|
# messenger = OneDenseLayerMLP()
|
|
return sensor, feature_extractor, actor, critic
|
|
|
|
def _init_agent_state(self) -> TrainState:
|
|
self.logger.info("[AGENT STATE]: Initializing agent state...")
|
|
|
|
self.key, sensor_key, actor_key, critic_key, feature_extractor_key = jax.random.split(
|
|
self.key, 5
|
|
)
|
|
|
|
sample_obs = jnp.concatenate(
|
|
[
|
|
v.flatten()
|
|
for v in self.env.single_observation_space.sample(
|
|
rng=jax.random.PRNGKey(0)
|
|
).values()
|
|
if v.size > 0
|
|
]
|
|
)
|
|
sensor_params = self.sensor.init(sensor_key, sample_obs)
|
|
feature_extractor_params = self.feature_extractor.init(feature_extractor_key, sample_obs)
|
|
actor_params = self.actor.init(actor_key, self.sensor.apply(sensor_params, sample_obs))
|
|
critic_params = self.critic.init(
|
|
critic_key, self.feature_extractor.apply(feature_extractor_params, sample_obs)
|
|
)
|
|
|
|
return TrainState.create(
|
|
apply_fn=None,
|
|
params=asdict(
|
|
AgentParams(sensor_params, actor_params, critic_params, feature_extractor_params)
|
|
),
|
|
tx=optax.chain(
|
|
optax.clip_by_global_norm(self.args.max_grad_norm),
|
|
optax.inject_hyperparams(optax.adam)(
|
|
learning_rate=partial(
|
|
_linear_schedule,
|
|
minibatch_count=self.args.num_minibatches,
|
|
update_epochs=self.args.update_epochs,
|
|
num_iterations=self.args.num_iterations,
|
|
learning_rate=self.args.learning_rate,
|
|
)
|
|
if self.args.anneal_lr
|
|
else self.args.learning_rate,
|
|
eps=1e-5,
|
|
),
|
|
),
|
|
)
|
|
|
|
def _init_episode_stats(self) -> EpisodeStatistics:
|
|
self.logger.info("[EPISODE STATS]: Initializing episode stats...")
|
|
|
|
return EpisodeStatistics( # type: ignore[call-arg]
|
|
episode_returns=jnp.zeros(self.args.num_envs, dtype=jnp.float32),
|
|
episode_lengths=jnp.zeros(self.args.num_envs, dtype=jnp.int32),
|
|
returned_episode_returns=jnp.zeros(self.args.num_envs, jnp.float32),
|
|
returned_episode_lengths=jnp.zeros(self.args.num_envs, dtype=jnp.int32),
|
|
)
|
|
|
|
def _rollout(self, env_state, next_obs, next_done) -> tuple[Any, ...]:
|
|
return self._rollout_jit(
|
|
self.agent_state,
|
|
self.episode_stats,
|
|
env_state,
|
|
next_obs,
|
|
next_done,
|
|
self.key,
|
|
)
|
|
|
|
def _compute_gae(self, storage, next_obs, next_done) -> Storage:
|
|
return self._compute_gae_jit(
|
|
self.agent_state,
|
|
storage,
|
|
next_obs,
|
|
next_done,
|
|
)
|
|
|
|
def _log(
|
|
self,
|
|
global_step,
|
|
episode_stats,
|
|
start_time,
|
|
iteration_time_start,
|
|
loss_info,
|
|
):
|
|
metrics = {
|
|
"charts/avg_episodic_return": loss_info.avg_episodic_return,
|
|
"charts/avg_episodic_length": np.mean(
|
|
jax.device_get(episode_stats.returned_episode_lengths)
|
|
),
|
|
"charts/learning_rate": self.agent_state.opt_state[1]
|
|
.hyperparams["learning_rate"]
|
|
.item(),
|
|
"losses/value_loss": loss_info.v_loss[-1, -1].item(),
|
|
"losses/policy_loss": loss_info.pg_loss[-1, -1].item(),
|
|
"losses/entropy": loss_info.entropy_loss[-1, -1].item(),
|
|
"losses/approx_kl": loss_info.approx_kl[-1, -1].item(),
|
|
"losses/loss": loss_info.loss[-1, -1].item(),
|
|
"charts/SPS": int(global_step / (time.time() - start_time)),
|
|
"charts/SPS_update": int(
|
|
self.args.num_envs * self.args.num_steps / (time.time() - iteration_time_start)
|
|
),
|
|
}
|
|
self.logger.log(metrics, step=global_step)
|
|
|
|
def _step(self, env_state, next_obs, next_done, iteration: int) -> tuple:
|
|
if iteration == 1:
|
|
self.logger.log_non_interactive(f"Starting first rollout (JIT): {time.ctime()}")
|
|
|
|
(
|
|
self.agent_state,
|
|
self.episode_stats,
|
|
next_obs,
|
|
next_done,
|
|
storage,
|
|
self.key,
|
|
next_env_state,
|
|
) = self._rollout(env_state, next_obs, next_done)
|
|
|
|
if iteration == 1:
|
|
self.logger.log_non_interactive(f"First rollout completed: {time.ctime()}")
|
|
|
|
storage = self._compute_gae(storage, next_obs, next_done)
|
|
|
|
if iteration == 1:
|
|
self.logger.log_non_interactive(f"Starting first PPO update (JIT): {time.ctime()}")
|
|
|
|
self.agent_state, loss, pg_loss, v_loss, entropy_loss, approx_kl, self.key = (
|
|
self._ppo.update_ppo(self.agent_state, storage, self.key)
|
|
)
|
|
|
|
if iteration == 1:
|
|
self.logger.log_non_interactive(f"First PPO update completed: {time.ctime()}")
|
|
|
|
avg_episodic_return = float(
|
|
jnp.mean(jax.device_get(self.episode_stats.returned_episode_returns)).item()
|
|
)
|
|
|
|
return (
|
|
next_env_state,
|
|
next_obs,
|
|
next_done,
|
|
LossInfo(
|
|
loss=loss,
|
|
pg_loss=pg_loss,
|
|
v_loss=v_loss,
|
|
entropy_loss=entropy_loss,
|
|
approx_kl=approx_kl,
|
|
avg_episodic_return=avg_episodic_return,
|
|
),
|
|
)
|
|
|
|
def _close(self):
|
|
self.env.close()
|
|
|
|
def _save_model(self, model_path: str):
|
|
self.logger.info("[SAVE]: Saving the final model...")
|
|
|
|
params = [
|
|
vars(self.args),
|
|
[
|
|
self.agent_state.params["sensor_params"],
|
|
self.agent_state.params["actor_params"],
|
|
self.agent_state.params["critic_params"],
|
|
self.agent_state.params["feature_extractor_params"],
|
|
],
|
|
]
|
|
self.logger.save_final_model(params=params)
|
|
|
|
def train(self):
|
|
"""
|
|
Train the PPO agent for a specified number of iterations
|
|
(passed through PPOArgs in constructor).
|
|
Closes the environment at the end of training.
|
|
"""
|
|
self.logger.info(f"running name: {self.run_name}")
|
|
|
|
self.logger.info("[TRAIN]: Resetting environment...")
|
|
self.logger.log_non_interactive(f"Initial reset started: {time.ctime()}")
|
|
|
|
env_state = self.env.reset(seed=self.args.seed)
|
|
next_obs = _convert_obs_dict_to_array(env_state.observations)
|
|
next_done = jnp.zeros(self.args.num_envs, dtype=jnp.bool_)
|
|
|
|
self.logger.log_non_interactive(f"Initial reset completed: {time.ctime()}")
|
|
|
|
global_step = 0
|
|
start_time = time.time()
|
|
|
|
iter_bar = self.logger.progress_bar(range(1, self.args.num_iterations + 1))
|
|
for iteration in iter_bar:
|
|
iteration_time_start = time.time()
|
|
|
|
env_state, next_obs, next_done, loss_info = self._step(
|
|
env_state, next_obs, next_done, iteration=iteration
|
|
)
|
|
|
|
global_step += self.args.num_steps * self.args.num_envs
|
|
self._log(global_step, self.episode_stats, start_time, iteration_time_start, loss_info)
|
|
|
|
sps = int(global_step / (time.time() - start_time))
|
|
remaining_steps = self.args.total_timesteps - global_step
|
|
eta_seconds = int(remaining_steps / sps) if sps > 0 else 0
|
|
eta_str = str(datetime.timedelta(seconds=eta_seconds))
|
|
|
|
self.logger.log_non_interactive(
|
|
f"Iteration {iteration}/{self.args.num_iterations} | "
|
|
f"Step {global_step}/{self.args.total_timesteps} | "
|
|
f"SPS {sps} | "
|
|
f"Return {loss_info.avg_episodic_return:.4f} | "
|
|
f"ETA {eta_str}"
|
|
)
|
|
|
|
if self.args.save_model:
|
|
model_path = f"{self.run_dir}/{self.args.exp_name}.cleanrl_model"
|
|
self._save_model(model_path=model_path)
|
|
|
|
self._close()
|