2cc34fb72b
Closes #914 Additional changes: - Deprecate python below 11 - Remove 3rd party and throughput tests. This simplifies install and test pipeline - Remove gym compatibility and shimmy - Format with 3.11 conventions. In particular, add `zip(..., strict=True/False)` where possible Since the additional tests and gym were complicating the CI pipeline (flaky and dist-dependent), it didn't make sense to work on fixing the current tests in this PR to then just delete them in the next one. So this PR changes the build and removes these tests at the same time.
195 lines
6.7 KiB
Python
195 lines
6.7 KiB
Python
import argparse
|
|
import os
|
|
import warnings
|
|
|
|
import gymnasium as gym
|
|
import numpy as np
|
|
import torch
|
|
from pettingzoo.butterfly import pistonball_v6
|
|
from torch.utils.tensorboard import SummaryWriter
|
|
|
|
from tianshou.data import Collector, VectorReplayBuffer
|
|
from tianshou.env import DummyVectorEnv
|
|
from tianshou.env.pettingzoo_env import PettingZooEnv
|
|
from tianshou.policy import BasePolicy, DQNPolicy, MultiAgentPolicyManager
|
|
from tianshou.trainer import OffpolicyTrainer
|
|
from tianshou.utils import TensorboardLogger
|
|
from tianshou.utils.net.common import Net
|
|
|
|
|
|
def get_parser() -> argparse.ArgumentParser:
|
|
parser = argparse.ArgumentParser()
|
|
parser.add_argument("--seed", type=int, default=1626)
|
|
parser.add_argument("--eps-test", type=float, default=0.05)
|
|
parser.add_argument("--eps-train", type=float, default=0.1)
|
|
parser.add_argument("--buffer-size", type=int, default=2000)
|
|
parser.add_argument("--lr", type=float, default=1e-3)
|
|
parser.add_argument(
|
|
"--gamma",
|
|
type=float,
|
|
default=0.9,
|
|
help="a smaller gamma favors earlier win",
|
|
)
|
|
parser.add_argument(
|
|
"--n-pistons",
|
|
type=int,
|
|
default=3,
|
|
help="Number of pistons(agents) in the env",
|
|
)
|
|
parser.add_argument("--n-step", type=int, default=100)
|
|
parser.add_argument("--target-update-freq", type=int, default=320)
|
|
parser.add_argument("--epoch", type=int, default=3)
|
|
parser.add_argument("--step-per-epoch", type=int, default=500)
|
|
parser.add_argument("--step-per-collect", type=int, default=10)
|
|
parser.add_argument("--update-per-step", type=float, default=0.1)
|
|
parser.add_argument("--batch-size", type=int, default=100)
|
|
parser.add_argument("--hidden-sizes", type=int, nargs="*", default=[64, 64])
|
|
parser.add_argument("--training-num", type=int, default=10)
|
|
parser.add_argument("--test-num", type=int, default=10)
|
|
parser.add_argument("--logdir", type=str, default="log")
|
|
parser.add_argument("--render", type=float, default=0.0)
|
|
|
|
parser.add_argument(
|
|
"--watch",
|
|
default=False,
|
|
action="store_true",
|
|
help="no training, watch the play of pre-trained models",
|
|
)
|
|
parser.add_argument(
|
|
"--device",
|
|
type=str,
|
|
default="cuda" if torch.cuda.is_available() else "cpu",
|
|
)
|
|
return parser
|
|
|
|
|
|
def get_args() -> argparse.Namespace:
|
|
parser = get_parser()
|
|
return parser.parse_known_args()[0]
|
|
|
|
|
|
def get_env(args: argparse.Namespace = get_args()):
|
|
return PettingZooEnv(pistonball_v6.env(continuous=False, n_pistons=args.n_pistons))
|
|
|
|
|
|
def get_agents(
|
|
args: argparse.Namespace = get_args(),
|
|
agents: list[BasePolicy] | None = None,
|
|
optims: list[torch.optim.Optimizer] | None = None,
|
|
) -> tuple[BasePolicy, list[torch.optim.Optimizer], list]:
|
|
env = get_env()
|
|
observation_space = (
|
|
env.observation_space["observation"]
|
|
if isinstance(env.observation_space, gym.spaces.Dict)
|
|
else env.observation_space
|
|
)
|
|
args.state_shape = observation_space.shape or observation_space.n
|
|
args.action_shape = env.action_space.shape or env.action_space.n
|
|
if agents is None:
|
|
agents = []
|
|
optims = []
|
|
for _ in range(args.n_pistons):
|
|
# model
|
|
net = Net(
|
|
args.state_shape,
|
|
args.action_shape,
|
|
hidden_sizes=args.hidden_sizes,
|
|
device=args.device,
|
|
).to(args.device)
|
|
optim = torch.optim.Adam(net.parameters(), lr=args.lr)
|
|
agent = DQNPolicy(
|
|
net,
|
|
optim,
|
|
args.gamma,
|
|
args.n_step,
|
|
target_update_freq=args.target_update_freq,
|
|
)
|
|
agents.append(agent)
|
|
optims.append(optim)
|
|
|
|
policy = MultiAgentPolicyManager(agents, env)
|
|
return policy, optims, env.agents
|
|
|
|
|
|
def train_agent(
|
|
args: argparse.Namespace = get_args(),
|
|
agents: list[BasePolicy] | None = None,
|
|
optims: list[torch.optim.Optimizer] | None = None,
|
|
) -> tuple[dict, BasePolicy]:
|
|
train_envs = DummyVectorEnv([get_env for _ in range(args.training_num)])
|
|
test_envs = DummyVectorEnv([get_env for _ in range(args.test_num)])
|
|
# seed
|
|
np.random.seed(args.seed)
|
|
torch.manual_seed(args.seed)
|
|
train_envs.seed(args.seed)
|
|
test_envs.seed(args.seed)
|
|
|
|
policy, optim, agents = get_agents(args, agents=agents, optims=optims)
|
|
|
|
# collector
|
|
train_collector = Collector(
|
|
policy,
|
|
train_envs,
|
|
VectorReplayBuffer(args.buffer_size, len(train_envs)),
|
|
exploration_noise=True,
|
|
)
|
|
test_collector = Collector(policy, test_envs, exploration_noise=True)
|
|
train_collector.collect(n_step=args.batch_size * args.training_num)
|
|
# log
|
|
log_path = os.path.join(args.logdir, "pistonball", "dqn")
|
|
writer = SummaryWriter(log_path)
|
|
writer.add_text("args", str(args))
|
|
logger = TensorboardLogger(writer)
|
|
|
|
def save_best_fn(policy):
|
|
pass
|
|
|
|
def stop_fn(mean_rewards):
|
|
return False
|
|
|
|
def train_fn(epoch, env_step):
|
|
[agent.set_eps(args.eps_train) for agent in policy.policies.values()]
|
|
|
|
def test_fn(epoch, env_step):
|
|
[agent.set_eps(args.eps_test) for agent in policy.policies.values()]
|
|
|
|
def reward_metric(rews):
|
|
return rews[:, 0]
|
|
|
|
# trainer
|
|
result = OffpolicyTrainer(
|
|
policy=policy,
|
|
train_collector=train_collector,
|
|
test_collector=test_collector,
|
|
max_epoch=args.epoch,
|
|
step_per_epoch=args.step_per_epoch,
|
|
step_per_collect=args.step_per_collect,
|
|
episode_per_test=args.test_num,
|
|
batch_size=args.batch_size,
|
|
train_fn=train_fn,
|
|
test_fn=test_fn,
|
|
stop_fn=stop_fn,
|
|
save_best_fn=save_best_fn,
|
|
update_per_step=args.update_per_step,
|
|
logger=logger,
|
|
test_in_train=False,
|
|
reward_metric=reward_metric,
|
|
).run()
|
|
|
|
return result, policy
|
|
|
|
|
|
def watch(args: argparse.Namespace = get_args(), policy: BasePolicy | None = None) -> None:
|
|
env = DummyVectorEnv([get_env])
|
|
if not policy:
|
|
warnings.warn(
|
|
"watching random agents, as loading pre-trained policies is currently not supported",
|
|
)
|
|
policy, _, _ = get_agents(args)
|
|
policy.eval()
|
|
[agent.set_eps(args.eps_test) for agent in policy.policies.values()]
|
|
collector = Collector(policy, env, exploration_noise=True)
|
|
result = collector.collect(n_episode=1, render=args.render)
|
|
rews, lens = result["rews"], result["lens"]
|
|
print(f"Final reward: {rews[:, 0].mean()}, length: {lens.mean()}")
|