def test_ppo(args=get_args()):
env = create_atari_environment(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
# train_envs = gym.make(args.task)
train_envs = SubprocVectorEnv([
lambda: create_atari_environment(args.task)
for _ in range(args.training_num)])
# test_envs = gym.make(args.task)
test_envs = SubprocVectorEnv([
lambda: create_atari_environment(args.task)
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)
# model
net = Net(args.layer_num, args.state_shape, device=args.device)
actor = Actor(net, args.action_shape).to(args.device)
critic = Critic(net).to(args.device)
optim = torch.optim.Adam(list(
actor.parameters()) + list(critic.parameters()), lr=args.lr)
dist = torch.distributions.Categorical
policy = PPOPolicy(
actor, critic, optim, dist, args.gamma,
max_grad_norm=args.max_grad_norm,
eps_clip=args.eps_clip,
vf_coef=args.vf_coef,
ent_coef=args.ent_coef,
action_range=None)
# collector
train_collector = Collector(
policy, train_envs, ReplayBuffer(args.buffer_size),
preprocess_fn=preprocess_fn)
test_collector = Collector(policy, test_envs, preprocess_fn=preprocess_fn)
# log
writer = SummaryWriter(args.logdir + '/' + 'ppo')
def stop_fn(x):
if env.env.spec.reward_threshold:
return x >= env.spec.reward_threshold
else:
return False
# trainer
result = onpolicy_trainer(
policy, train_collector, test_collector, args.epoch,
args.step_per_epoch, args.collect_per_step, args.repeat_per_collect,
args.test_num, args.batch_size, stop_fn=stop_fn, writer=writer)
train_collector.close()
test_collector.close()
if __name__ == '__main__':
pprint.pprint(result)
# Let's watch its performance!
env = create_atari_environment(args.task)
collector = Collector(policy, env, preprocess_fn=preprocess_fn)
result = collector.collect(n_step=2000, render=args.render)
print(f'Final reward: {result["rew"]}, length: {result["len"]}')
collector.close()
def _ppo_policy():
actor = PolicyNet(2, True)
critic = PolicyNet()
policy = PPOPolicy(
actor,
critic,
torch.optim.Adam(
tuple(actor.parameters()) + tuple(critic.parameters())),
torch.distributions.Categorical,
action_space=NoopActionInterpreter().action_space,
)
return policy
def _default_policy_fn(env):
net = Preprocessor(env.observation_space)
actor = Actor(env.action_space, net)
critic = Critic(net)
optim = torch.optim.Adam(set(actor.parameters()).union(
critic.parameters()),
lr=1e-4)
return PPOPolicy(actor,
critic,
optim,
torch.distributions.Categorical,
discount_factor=1.,
action_space=env.action_space)
def __init__(self, args, mask=None, name='full'):
env = gym.make(args.task)
self.stop_fn = lambda x: x >= env.spec.reward_threshold
self.state_shape = env.observation_space.shape or env.observation_space.n
self.action_shape = env.action_space.shape or env.action_space.n
self.buffer = ReplayBuffer(400000)
# Env
# train_envs = gym.make(args.task)
self.train_envs = SubprocVectorEnv(
[lambda: gym.make(args.task) for _ in range(args.training_num)])
# test_envs = gym.make(args.task)
self.test_envs = SubprocVectorEnv(
[lambda: gym.make(args.task) for _ in range(args.test_num)])
# Mask
state_dim = int(np.prod(self.state_shape))
self._view_mask = torch.ones(state_dim)
if mask == 'even':
for i in range(0, state_dim, 2):
self._view_mask[i] = 0
elif mask == "odd":
for i in range(1, state_dim, 2):
self._view_mask[i] = 0
elif type(mask) == int:
self._view_mask[mask] = 0
# Model
net = NetWithView(args.layer_num, self.state_shape, device=args.device,
mask=self._view_mask)
self.actor = Actor(net, self.action_shape).to(args.device)
self.critic = Critic(net).to(args.device)
optim = torch.optim.Adam(list(
self.actor.parameters()) + list(self.critic.parameters()), lr=args.lr)
dist = torch.distributions.Categorical
self.policy = PPOPolicy(
self.actor, self.critic, optim, dist, args.gamma,
max_grad_norm=args.max_grad_norm,
eps_clip=args.eps_clip,
vf_coef=args.vf_coef,
ent_coef=args.ent_coef,
action_range=None)
# Collector
self.train_collector = Collector(
self.policy, self.train_envs, ReplayBuffer(args.buffer_size))
self.test_collector = Collector(self.policy, self.test_envs)
# Log
self.writer = SummaryWriter(f'{args.logdir}/{args.task}/ppo/{args.note}/{name}')
def test_ppo(args=get_args()):
torch.set_num_threads(1) # for poor CPU
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
# train_envs = gym.make(args.task)
# you can also use tianshou.env.SubprocVectorEnv
train_envs = VectorEnv(
[lambda: gym.make(args.task) for _ in range(args.training_num)])
# test_envs = gym.make(args.task)
test_envs = VectorEnv(
[lambda: gym.make(args.task) 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)
# model
net = Net(args.layer_num, args.state_shape, device=args.device)
actor = Actor(net, args.action_shape).to(args.device)
critic = Critic(net).to(args.device)
optim = torch.optim.Adam(list(
actor.parameters()) + list(critic.parameters()), lr=args.lr)
dist = torch.distributions.Categorical
policy = PPOPolicy(
actor, critic, optim, dist, args.gamma,
max_grad_norm=args.max_grad_norm,
eps_clip=args.eps_clip,
vf_coef=args.vf_coef,
ent_coef=args.ent_coef,
action_range=None,
gae_lambda=args.gae_lambda,
reward_normalization=args.rew_norm,
dual_clip=args.dual_clip,
value_clip=args.value_clip)
# collector
train_collector = Collector(
policy, train_envs, ReplayBuffer(args.buffer_size))
test_collector = Collector(policy, test_envs)
def init_policy(args, env):
actor = Actor(layer=None,
state_shape=args.state_shape,
action_shape=args.action_shape,
action_range=args.action_range,
device=args.device).to(args.device)
critic = Critic(layer=None,
state_shape=args.state_shape,
action_shape=args.action_shape,
device=args.device).to(args.device)
# orthogonal initialization
for m in list(actor.modules()) + list(critic.modules()):
if isinstance(m, torch.nn.Linear):
torch.nn.init.orthogonal_(m.weight)
torch.nn.init.zeros_(m.bias)
optim = torch.optim.Adam(list(actor.parameters()) +
list(critic.parameters()),
lr=args.lr)
dist = DiagGaussian
policy = PPOPolicy(
actor,
critic,
optim,
dist,
args.gamma,
max_grad_norm=args.max_grad_norm,
eps_clip=args.eps_clip,
vf_coef=args.vf_coef,
ent_coef=args.ent_coef,
reward_normalization=args.rew_norm,
# dual_clip=args.dual_clip,
# dual clip cause monotonically increasing log_std :)
value_clip=args.value_clip,
# action_range=[env.action_space.low[0], env.action_space.high[0]],)
# if clip the action, ppo would not converge :)
gae_lambda=args.gae_lambda)
return policy
def test_ppo(args=get_args()):
torch.set_num_threads(1) # we just need only one thread for NN
env = gym.make(args.task)
if args.task == 'Pendulum-v0':
env.spec.reward_threshold = -250
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]
# you can also use tianshou.env.SubprocVectorEnv
# train_envs = gym.make(args.task)
train_envs = DummyVectorEnv(
[lambda: gym.make(args.task) for _ in range(args.training_num)])
# test_envs = gym.make(args.task)
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)
train_envs.seed(args.seed)
test_envs.seed(args.seed)
# model
net = Net(args.state_shape, hidden_sizes=args.hidden_sizes,
device=args.device)
actor = ActorProb(net, args.action_shape, max_action=args.max_action,
device=args.device).to(args.device)
critic = Critic(Net(
args.state_shape, hidden_sizes=args.hidden_sizes, device=args.device
), device=args.device).to(args.device)
# orthogonal initialization
for m in list(actor.modules()) + list(critic.modules()):
if isinstance(m, torch.nn.Linear):
torch.nn.init.orthogonal_(m.weight)
torch.nn.init.zeros_(m.bias)
optim = torch.optim.Adam(set(
actor.parameters()).union(critic.parameters()), lr=args.lr)
# replace DiagGuassian with Independent(Normal) which is equivalent
# pass *logits to be consistent with policy.forward
def dist(*logits):
return Independent(Normal(*logits), 1)
policy = PPOPolicy(
actor, critic, optim, dist,
discount_factor=args.gamma,
max_grad_norm=args.max_grad_norm,
eps_clip=args.eps_clip,
vf_coef=args.vf_coef,
ent_coef=args.ent_coef,
reward_normalization=args.rew_norm,
advantage_normalization=args.norm_adv,
recompute_advantage=args.recompute_adv,
# dual_clip=args.dual_clip,
# dual clip cause monotonically increasing log_std :)
value_clip=args.value_clip,
gae_lambda=args.gae_lambda,
action_space=env.action_space)
# collector
train_collector = Collector(
policy, train_envs,
VectorReplayBuffer(args.buffer_size, len(train_envs)),
exploration_noise=True)
test_collector = Collector(policy, test_envs)
# log
log_path = os.path.join(args.logdir, args.task, 'ppo')
writer = SummaryWriter(log_path)
logger = BasicLogger(writer)
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
# trainer
result = onpolicy_trainer(
policy, train_collector, test_collector, args.epoch,
args.step_per_epoch, args.repeat_per_collect, args.test_num, args.batch_size,
episode_per_collect=args.episode_per_collect, stop_fn=stop_fn, save_fn=save_fn,
logger=logger)
assert stop_fn(result['best_reward'])
if __name__ == '__main__':
pprint.pprint(result)
# Let's watch its performance!
env = gym.make(args.task)
policy.eval()
collector = Collector(policy, env)
result = collector.collect(n_episode=1, render=args.render)
rews, lens = result["rews"], result["lens"]
print(f"Final reward: {rews.mean()}, length: {lens.mean()}")
def test_ppo(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]
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))
# train_envs = gym.make(args.task)
train_envs = SubprocVectorEnv(
[lambda: gym.make(args.task) for _ in range(args.training_num)],
norm_obs=True)
# test_envs = gym.make(args.task)
test_envs = SubprocVectorEnv(
[lambda: gym.make(args.task) for _ in range(args.test_num)],
norm_obs=True,
obs_rms=train_envs.obs_rms,
update_obs_rms=False)
# seed
np.random.seed(args.seed)
torch.manual_seed(args.seed)
train_envs.seed(args.seed)
test_envs.seed(args.seed)
# model
net_a = Net(args.state_shape,
hidden_sizes=args.hidden_sizes,
activation=nn.Tanh,
device=args.device)
actor = ActorProb(net_a,
args.action_shape,
max_action=args.max_action,
unbounded=True,
device=args.device).to(args.device)
net_c = Net(args.state_shape,
hidden_sizes=args.hidden_sizes,
activation=nn.Tanh,
device=args.device)
critic = Critic(net_c, device=args.device).to(args.device)
torch.nn.init.constant_(actor.sigma_param._bias, -0.5)
for m in list(actor.modules()) + list(critic.modules()):
if isinstance(m, torch.nn.Linear):
# orthogonal initialization
torch.nn.init.orthogonal_(m.weight, gain=np.sqrt(2))
torch.nn.init.zeros_(m.bias)
# do last policy layer scaling, this will make initial actions have (close to)
# 0 mean and std, and will help boost performances,
# see https://arxiv.org/abs/2006.05990, Fig.24 for details
for m in actor.mu.modules():
if isinstance(m, torch.nn.Linear):
torch.nn.init.zeros_(m.bias)
m.weight.data.copy_(0.01 * m.weight.data)
optim = torch.optim.Adam(list(actor.parameters()) +
list(critic.parameters()),
lr=args.lr)
lr_scheduler = None
if args.lr_decay:
# decay learning rate to 0 linearly
max_update_num = np.ceil(
args.step_per_epoch / args.step_per_collect) * args.epoch
lr_scheduler = LambdaLR(
optim, lr_lambda=lambda epoch: 1 - epoch / max_update_num)
def dist(*logits):
return Independent(Normal(*logits), 1)
policy = PPOPolicy(actor,
critic,
optim,
dist,
discount_factor=args.gamma,
gae_lambda=args.gae_lambda,
max_grad_norm=args.max_grad_norm,
vf_coef=args.vf_coef,
ent_coef=args.ent_coef,
reward_normalization=args.rew_norm,
action_scaling=True,
action_bound_method=args.bound_action_method,
lr_scheduler=lr_scheduler,
action_space=env.action_space,
eps_clip=args.eps_clip,
value_clip=args.value_clip,
dual_clip=args.dual_clip,
advantage_normalization=args.norm_adv,
recompute_advantage=args.recompute_adv)
# 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
if args.training_num > 1:
buffer = VectorReplayBuffer(args.buffer_size, len(train_envs))
else:
buffer = ReplayBuffer(args.buffer_size)
train_collector = Collector(policy,
train_envs,
buffer,
exploration_noise=True)
test_collector = Collector(policy, test_envs)
# log
t0 = datetime.datetime.now().strftime("%m%d_%H%M%S")
log_file = f'seed_{args.seed}_{t0}-{args.task.replace("-", "_")}_ppo'
log_path = os.path.join(args.logdir, args.task, 'ppo', log_file)
writer = SummaryWriter(log_path)
writer.add_text("args", str(args))
logger = BasicLogger(writer, update_interval=100, train_interval=100)
def save_fn(policy):
torch.save(policy.state_dict(), os.path.join(log_path, 'policy.pth'))
if not args.watch:
# trainer
result = onpolicy_trainer(policy,
train_collector,
test_collector,
args.epoch,
args.step_per_epoch,
args.repeat_per_collect,
args.test_num,
args.batch_size,
step_per_collect=args.step_per_collect,
save_fn=save_fn,
logger=logger,
test_in_train=False)
pprint.pprint(result)
# 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()}'
)
def test_ppo(args=get_args()):
torch.set_num_threads(1) # we just need only one thread for NN
env = gym.make(args.task)
if args.task == 'Pendulum-v0':
env.spec.reward_threshold = -250
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]
# you can also use tianshou.env.SubprocVectorEnv
# train_envs = gym.make(args.task)
train_envs = VectorEnv(
[lambda: gym.make(args.task) for _ in range(args.training_num)])
# test_envs = gym.make(args.task)
test_envs = VectorEnv(
[lambda: gym.make(args.task) 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)
# model
actor = ActorProb(args.layer_num, args.state_shape, args.action_shape,
args.max_action, args.device).to(args.device)
critic = Critic(args.layer_num, args.state_shape,
device=args.device).to(args.device)
optim = torch.optim.Adam(list(actor.parameters()) +
list(critic.parameters()),
lr=args.lr)
dist = torch.distributions.Normal
policy = PPOPolicy(
actor,
critic,
optim,
dist,
args.gamma,
max_grad_norm=args.max_grad_norm,
eps_clip=args.eps_clip,
vf_coef=args.vf_coef,
ent_coef=args.ent_coef,
reward_normalization=args.rew_norm,
dual_clip=args.dual_clip,
value_clip=args.value_clip,
# action_range=[env.action_space.low[0], env.action_space.high[0]],)
# if clip the action, ppo would not converge :)
gae_lambda=args.gae_lambda)
# collector
train_collector = Collector(policy, train_envs,
ReplayBuffer(args.buffer_size))
test_collector = Collector(policy, test_envs)
# log
log_path = os.path.join(args.logdir, args.task, 'ppo')
writer = SummaryWriter(log_path)
def save_fn(policy):
torch.save(policy.state_dict(), os.path.join(log_path, 'policy.pth'))
def stop_fn(x):
return x >= env.spec.reward_threshold
# trainer
result = onpolicy_trainer(policy,
train_collector,
test_collector,
args.epoch,
args.step_per_epoch,
args.collect_per_step,
args.repeat_per_collect,
args.test_num,
args.batch_size,
stop_fn=stop_fn,
save_fn=save_fn,
writer=writer)
assert stop_fn(result['best_reward'])
train_collector.close()
test_collector.close()
if __name__ == '__main__':
pprint.pprint(result)
# Let's watch its performance!
env = gym.make(args.task)
collector = Collector(policy, env)
result = collector.collect(n_episode=1, render=args.render)
print(f'Final reward: {result["rew"]}, length: {result["len"]}')
collector.close()
def get_agents(
args: argparse.Namespace = get_args(),
agents: Optional[List[BasePolicy]] = None,
optims: Optional[List[torch.optim.Optimizer]] = 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
args.max_action = env.action_space.high[0]
if agents is None:
agents = []
optims = []
for _ in range(args.n_pistons):
# model
net = DQN(observation_space.shape[2],
observation_space.shape[1],
observation_space.shape[0],
device=args.device).to(args.device)
actor = ActorProb(net,
args.action_shape,
max_action=args.max_action,
device=args.device).to(args.device)
net2 = DQN(observation_space.shape[2],
observation_space.shape[1],
observation_space.shape[0],
device=args.device).to(args.device)
critic = Critic(net2, device=args.device).to(args.device)
for m in set(actor.modules()).union(critic.modules()):
if isinstance(m, torch.nn.Linear):
torch.nn.init.orthogonal_(m.weight)
torch.nn.init.zeros_(m.bias)
optim = torch.optim.Adam(set(actor.parameters()).union(
critic.parameters()),
lr=args.lr)
def dist(*logits):
return Independent(Normal(*logits), 1)
agent = PPOPolicy(
actor,
critic,
optim,
dist,
discount_factor=args.gamma,
max_grad_norm=args.max_grad_norm,
eps_clip=args.eps_clip,
vf_coef=args.vf_coef,
ent_coef=args.ent_coef,
reward_normalization=args.rew_norm,
advantage_normalization=args.norm_adv,
recompute_advantage=args.recompute_adv,
# dual_clip=args.dual_clip,
# dual clip cause monotonically increasing log_std :)
value_clip=args.value_clip,
gae_lambda=args.gae_lambda,
action_space=env.action_space)
agents.append(agent)
optims.append(optim)
policy = MultiAgentPolicyManager(agents,
env,
action_scaling=True,
action_bound_method='clip')
return policy, optims, env.agents
for m in list(actor.modules()) + list(critic.modules()):
if isinstance(m, torch.nn.Linear):
torch.nn.init.orthogonal_(m.weight)
torch.nn.init.zeros_(m.bias)
optim = torch.optim.Adam(list(actor.parameters()) + list(critic.parameters()),
lr=training_config['learning_rate'])
dist = torch.distributions.Categorical
policy = PPOPolicy(actor,
critic,
optim,
dist,
training_config['gamma'],
eps_clip=training_config["eps_clip"],
vf_coef=training_config["vf_coef"],
ent_coef=training_config["ent_coef"],
action_range=None,
gae_lambda=training_config["gae_lambda"],
reward_normalization=training_config["rew_norm"],
dual_clip=None,
value_clip=training_config["value_clip"])
buf = ReplayBuffer(training_config['buffer_size'])
train_collector = Collector(policy, train_envs, buf)
train_fn = lambda e: [
torch.save(policy.state_dict(),
os.path.join(save_path, 'policy_%d.pth' % (e)))
]
def test_ppo(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
if args.reward_threshold is None:
default_reward_threshold = {"CartPole-v0": 195}
args.reward_threshold = default_reward_threshold.get(
args.task, env.spec.reward_threshold)
# train_envs = gym.make(args.task)
# you can also use tianshou.env.SubprocVectorEnv
train_envs = DummyVectorEnv(
[lambda: gym.make(args.task) for _ in range(args.training_num)])
# test_envs = gym.make(args.task)
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)
train_envs.seed(args.seed)
test_envs.seed(args.seed)
# model
net = Net(args.state_shape,
hidden_sizes=args.hidden_sizes,
device=args.device)
actor = Actor(net, args.action_shape, device=args.device).to(args.device)
critic = Critic(net, device=args.device).to(args.device)
actor_critic = ActorCritic(actor, critic)
# orthogonal initialization
for m in actor_critic.modules():
if isinstance(m, torch.nn.Linear):
torch.nn.init.orthogonal_(m.weight)
torch.nn.init.zeros_(m.bias)
optim = torch.optim.Adam(actor_critic.parameters(), lr=args.lr)
dist = torch.distributions.Categorical
policy = PPOPolicy(actor,
critic,
optim,
dist,
discount_factor=args.gamma,
max_grad_norm=args.max_grad_norm,
eps_clip=args.eps_clip,
vf_coef=args.vf_coef,
ent_coef=args.ent_coef,
gae_lambda=args.gae_lambda,
reward_normalization=args.rew_norm,
dual_clip=args.dual_clip,
value_clip=args.value_clip,
action_space=env.action_space,
deterministic_eval=True,
advantage_normalization=args.norm_adv,
recompute_advantage=args.recompute_adv)
feature_dim = args.hidden_sizes[-1]
feature_net = MLP(np.prod(args.state_shape),
output_dim=feature_dim,
hidden_sizes=args.hidden_sizes[:-1],
device=args.device)
action_dim = np.prod(args.action_shape)
icm_net = IntrinsicCuriosityModule(feature_net,
feature_dim,
action_dim,
hidden_sizes=args.hidden_sizes[-1:],
device=args.device).to(args.device)
icm_optim = torch.optim.Adam(icm_net.parameters(), lr=args.lr)
policy = ICMPolicy(policy, icm_net, icm_optim, args.lr_scale,
args.reward_scale, args.forward_loss_weight)
# collector
train_collector = Collector(
policy, train_envs,
VectorReplayBuffer(args.buffer_size, len(train_envs)))
test_collector = Collector(policy, test_envs)
# log
log_path = os.path.join(args.logdir, args.task, 'ppo_icm')
writer = SummaryWriter(log_path)
logger = TensorboardLogger(writer)
def save_best_fn(policy):
torch.save(policy.state_dict(), os.path.join(log_path, 'policy.pth'))
def stop_fn(mean_rewards):
return mean_rewards >= args.reward_threshold
# trainer
result = onpolicy_trainer(policy,
train_collector,
test_collector,
args.epoch,
args.step_per_epoch,
args.repeat_per_collect,
args.test_num,
args.batch_size,
step_per_collect=args.step_per_collect,
stop_fn=stop_fn,
save_best_fn=save_best_fn,
logger=logger)
assert stop_fn(result['best_reward'])
if __name__ == '__main__':
pprint.pprint(result)
# Let's watch its performance!
env = gym.make(args.task)
policy.eval()
collector = Collector(policy, env)
result = collector.collect(n_episode=1, render=args.render)
rews, lens = result["rews"], result["lens"]
print(f"Final reward: {rews.mean()}, length: {lens.mean()}")
### orthogonal initialization
for m in list(actor.modules()) + list(critic.modules()):
if isinstance(m, torch.nn.Linear):
torch.nn.init.orthogonal_(m.weight)
torch.nn.init.zeros_(m.bias)
### Set the probability distribution
if isinstance(env.action_space, gym.spaces.Discrete):
dist_fn = torch.distributions.Categorical
else:
dist_fn = DiagGaussian
ppo_config["action_range"] = [float(env.action_space.low),
float(env.action_space.high)]
# Setup policy and collectors:
policy = PPOPolicy(actor, critic, optim, dist_fn, **ppo_config)
train_collector = Collector(policy, train_envs, ReplayBuffer(buffer_size))
test_collector = Collector(policy, test_envs)
# Setup the trainer and run the learning process
### Define a custom on-policy trainer
def onpolicy_trainer(
policy: BasePolicy,
train_collector: Collector,
test_collector: Collector,
max_epoch: int,
frame_per_epoch: int,
collect_per_step: int,
repeat_per_collect: int,
episode_per_test: Union[int, List[int]],
def test_ppo(args=get_args()):
args.cfg_path = f"maps/{args.task}.cfg"
args.wad_path = f"maps/{args.task}.wad"
args.res = (args.skip_num, 84, 84)
env = Env(args.cfg_path, args.frames_stack, args.res)
args.state_shape = args.res
args.action_shape = env.action_space.shape or env.action_space.n
# should be N_FRAMES x H x W
print("Observations shape:", args.state_shape)
print("Actions shape:", args.action_shape)
# make environments
train_envs = ShmemVectorEnv([
lambda: Env(args.cfg_path, args.frames_stack, args.res)
for _ in range(args.training_num)
])
test_envs = ShmemVectorEnv([
lambda: Env(args.cfg_path, args.frames_stack, args.res, args.save_lmp)
for _ in range(min(os.cpu_count() - 1, args.test_num))
])
# seed
np.random.seed(args.seed)
torch.manual_seed(args.seed)
train_envs.seed(args.seed)
test_envs.seed(args.seed)
# define model
net = DQN(*args.state_shape,
args.action_shape,
device=args.device,
features_only=True,
output_dim=args.hidden_size)
actor = Actor(net,
args.action_shape,
device=args.device,
softmax_output=False)
critic = Critic(net, device=args.device)
optim = torch.optim.Adam(ActorCritic(actor, critic).parameters(),
lr=args.lr)
lr_scheduler = None
if args.lr_decay:
# decay learning rate to 0 linearly
max_update_num = np.ceil(
args.step_per_epoch / args.step_per_collect) * args.epoch
lr_scheduler = LambdaLR(
optim, lr_lambda=lambda epoch: 1 - epoch / max_update_num)
# define policy
def dist(p):
return torch.distributions.Categorical(logits=p)
policy = PPOPolicy(actor,
critic,
optim,
dist,
discount_factor=args.gamma,
gae_lambda=args.gae_lambda,
max_grad_norm=args.max_grad_norm,
vf_coef=args.vf_coef,
ent_coef=args.ent_coef,
reward_normalization=args.rew_norm,
action_scaling=False,
lr_scheduler=lr_scheduler,
action_space=env.action_space,
eps_clip=args.eps_clip,
value_clip=args.value_clip,
dual_clip=args.dual_clip,
advantage_normalization=args.norm_adv,
recompute_advantage=args.recompute_adv).to(args.device)
if args.icm_lr_scale > 0:
feature_net = DQN(*args.state_shape,
args.action_shape,
device=args.device,
features_only=True,
output_dim=args.hidden_size)
action_dim = np.prod(args.action_shape)
feature_dim = feature_net.output_dim
icm_net = IntrinsicCuriosityModule(feature_net.net,
feature_dim,
action_dim,
device=args.device)
icm_optim = torch.optim.Adam(icm_net.parameters(), lr=args.lr)
policy = ICMPolicy(policy, icm_net, icm_optim, args.icm_lr_scale,
args.icm_reward_scale,
args.icm_forward_loss_weight).to(args.device)
# 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)
# replay buffer: `save_last_obs` and `stack_num` can be removed together
# when you have enough RAM
buffer = VectorReplayBuffer(args.buffer_size,
buffer_num=len(train_envs),
ignore_obs_next=True,
save_only_last_obs=True,
stack_num=args.frames_stack)
# collector
train_collector = Collector(policy,
train_envs,
buffer,
exploration_noise=True)
test_collector = Collector(policy, test_envs, exploration_noise=True)
# log
log_name = 'ppo_icm' if args.icm_lr_scale > 0 else 'ppo'
log_path = os.path.join(args.logdir, args.task, log_name)
writer = SummaryWriter(log_path)
writer.add_text("args", str(args))
logger = TensorboardLogger(writer)
def save_best_fn(policy):
torch.save(policy.state_dict(), os.path.join(log_path, 'policy.pth'))
def stop_fn(mean_rewards):
if env.spec.reward_threshold:
return mean_rewards >= env.spec.reward_threshold
elif 'Pong' in args.task:
return mean_rewards >= 20
else:
return False
# watch agent's performance
def watch():
print("Setup test envs ...")
policy.eval()
test_envs.seed(args.seed)
if args.save_buffer_name:
print(f"Generate buffer with size {args.buffer_size}")
buffer = VectorReplayBuffer(args.buffer_size,
buffer_num=len(test_envs),
ignore_obs_next=True,
save_only_last_obs=True,
stack_num=args.frames_stack)
collector = Collector(policy,
test_envs,
buffer,
exploration_noise=True)
result = collector.collect(n_step=args.buffer_size)
print(f"Save buffer into {args.save_buffer_name}")
# Unfortunately, pickle will cause oom with 1M buffer size
buffer.save_hdf5(args.save_buffer_name)
else:
print("Testing agent ...")
test_collector.reset()
result = test_collector.collect(n_episode=args.test_num,
render=args.render)
rew = result["rews"].mean()
lens = result["lens"].mean() * args.skip_num
print(f'Mean reward (over {result["n/ep"]} episodes): {rew}')
print(f'Mean length (over {result["n/ep"]} episodes): {lens}')
if args.watch:
watch()
exit(0)
# test train_collector and start filling replay buffer
train_collector.collect(n_step=args.batch_size * args.training_num)
# trainer
result = onpolicy_trainer(policy,
train_collector,
test_collector,
args.epoch,
args.step_per_epoch,
args.repeat_per_collect,
args.test_num,
args.batch_size,
step_per_collect=args.step_per_collect,
stop_fn=stop_fn,
save_best_fn=save_best_fn,
logger=logger,
test_in_train=False)
pprint.pprint(result)
watch()
def build_policy(no, args):
if no == 0:
# server policy
net = Net(args.layer_num, args.state_shape, device=args.device)
actor = ServerActor(net, (10, )).to(args.device)
critic = Critic(net).to(args.device)
# orthogonal initialization
for m in list(actor.modules()) + list(critic.modules()):
if isinstance(m, torch.nn.Linear):
torch.nn.init.orthogonal_(m.weight)
torch.nn.init.zeros_(m.bias)
optim = torch.optim.Adam(list(actor.parameters()) +
list(critic.parameters()),
lr=args.lr)
dist = torch.distributions.Categorical
policy = PPOPolicy(actor,
critic,
optim,
dist,
discount_factor=args.gamma,
max_grad_norm=args.max_grad_norm,
eps_clip=args.eps_clip,
vf_coef=args.vf_coef,
ent_coef=args.ent_coef,
gae_lambda=args.gae_lambda,
reward_normalization=args.rew_norm,
dual_clip=args.dual_clip,
value_clip=args.value_clip)
elif no == 1:
# ND policy
net = Net(args.layer_num, (4, ), device=args.device)
actor = NFActor(net, (10, )).to(args.device)
critic = Critic(net).to(args.device)
# orthogonal initialization
for m in list(actor.modules()) + list(critic.modules()):
if isinstance(m, torch.nn.Linear):
torch.nn.init.orthogonal_(m.weight)
torch.nn.init.zeros_(m.bias)
optim = torch.optim.Adam(list(actor.parameters()) +
list(critic.parameters()),
lr=args.lr)
dist = torch.distributions.Categorical
policy = PPOPolicy(actor,
critic,
optim,
dist,
discount_factor=args.gamma,
max_grad_norm=args.max_grad_norm,
eps_clip=args.eps_clip,
vf_coef=args.vf_coef,
ent_coef=args.ent_coef,
gae_lambda=args.gae_lambda,
reward_normalization=args.rew_norm,
dual_clip=args.dual_clip,
value_clip=args.value_clip)
elif no == 2:
net = Net(args.layer_num, (4, ), device=args.device)
actor = RelayActor(net, (10, )).to(args.device)
critic = Critic(net).to(args.device)
# orthogonal initialization
for m in list(actor.modules()) + list(critic.modules()):
if isinstance(m, torch.nn.Linear):
torch.nn.init.orthogonal_(m.weight)
torch.nn.init.zeros_(m.bias)
optim = torch.optim.Adam(list(actor.parameters()) +
list(critic.parameters()),
lr=args.lr)
dist = torch.distributions.Categorical
policy = PPOPolicy(actor,
critic,
optim,
dist,
discount_factor=args.gamma,
max_grad_norm=args.max_grad_norm,
eps_clip=args.eps_clip,
vf_coef=args.vf_coef,
ent_coef=args.ent_coef,
gae_lambda=args.gae_lambda,
reward_normalization=args.rew_norm,
dual_clip=args.dual_clip,
value_clip=args.value_clip)
else:
net = Net(args.layer_num, (4, ), device=args.device)
actor = NFActor(net, (10, )).to(args.device)
critic = Critic(net).to(args.device)
# orthogonal initialization
for m in list(actor.modules()) + list(critic.modules()):
if isinstance(m, torch.nn.Linear):
torch.nn.init.orthogonal_(m.weight)
torch.nn.init.zeros_(m.bias)
optim = torch.optim.Adam(list(actor.parameters()) +
list(critic.parameters()),
lr=args.lr)
dist = torch.distributions.Categorical
policy = PPOPolicy(actor,
critic,
optim,
dist,
discount_factor=args.gamma,
max_grad_norm=args.max_grad_norm,
eps_clip=args.eps_clip,
vf_coef=args.vf_coef,
ent_coef=args.ent_coef,
gae_lambda=args.gae_lambda,
reward_normalization=args.rew_norm,
dual_clip=args.dual_clip,
value_clip=args.value_clip)
return policy
# Create actor and critic
actor, critic, dist_fn = build_actor_critic(
env_creator, vf_share_layers=False, free_log_std=True)
actor = actor.to("cuda")
critic = critic.to("cuda")
# Set the action range in continuous mode
env = env_creator()
if isinstance(env.action_space, gym.spaces.Box):
ppo_config["action_range"] = [
float(env.action_space.low), float(env.action_space.high)]
# Optimizer parameters
lr = 1.0e-3
optimizer = torch.optim.Adam(
list(actor.parameters()) + list(critic.parameters()), lr=lr)
# Create the training agent
train_agent = PPOPolicy(actor, critic, optimizer, dist_fn, **ppo_config)
# Run the learning process
checkpoint_path = train(
train_agent, train_envs, test_envs, writer, trainer_config)
# ===================== Enjoy the trained agent ======================
test_agent = PPOPolicy(actor, critic, optimizer, dist_fn, **ppo_config)
test_agent.load_state_dict(torch.load(checkpoint_path))
test(test_agent, env_creator, num_episodes=1)
def _test_ppo(args=get_args()):
# just a demo, I have not made it work :(
env = gym.make(args.task)
if args.task == 'Pendulum-v0':
env.spec.reward_threshold = -250
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]
# train_envs = gym.make(args.task)
train_envs = SubprocVectorEnv(
[lambda: gym.make(args.task) for _ in range(args.training_num)])
# test_envs = gym.make(args.task)
test_envs = SubprocVectorEnv(
[lambda: gym.make(args.task) 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)
# model
actor = ActorProb(args.layer_num, args.state_shape, args.action_shape,
args.max_action, args.device).to(args.device)
critic = Critic(args.layer_num, args.state_shape,
device=args.device).to(args.device)
optim = torch.optim.Adam(list(actor.parameters()) +
list(critic.parameters()),
lr=args.lr)
dist = torch.distributions.Normal
policy = PPOPolicy(
actor,
critic,
optim,
dist,
args.gamma,
max_grad_norm=args.max_grad_norm,
eps_clip=args.eps_clip,
vf_coef=args.vf_coef,
ent_coef=args.ent_coef,
action_range=[env.action_space.low[0], env.action_space.high[0]])
# collector
train_collector = Collector(policy, train_envs,
ReplayBuffer(args.buffer_size))
test_collector = Collector(policy, test_envs)
train_collector.collect(n_step=args.step_per_epoch)
# log
writer = SummaryWriter(args.logdir + '/' + 'ppo')
def stop_fn(x):
return x >= env.spec.reward_threshold
# trainer
result = onpolicy_trainer(policy,
train_collector,
test_collector,
args.epoch,
args.step_per_epoch,
args.collect_per_step,
args.repeat_per_collect,
args.test_num,
args.batch_size,
stop_fn=stop_fn,
writer=writer,
task=args.task)
assert stop_fn(result['best_reward'])
train_collector.close()
test_collector.close()
if __name__ == '__main__':
pprint.pprint(result)
# Let's watch its performance!
env = gym.make(args.task)
collector = Collector(policy, env)
result = collector.collect(n_episode=1, render=args.render)
print(f'Final reward: {result["rew"]}, length: {result["len"]}')
collector.close()
def test_ppo(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]
if args.reward_threshold is None:
default_reward_threshold = {"Pendulum-v0": -250, "Pendulum-v1": -250}
args.reward_threshold = default_reward_threshold.get(
args.task, env.spec.reward_threshold)
# you can also use tianshou.env.SubprocVectorEnv
# train_envs = gym.make(args.task)
train_envs = DummyVectorEnv(
[lambda: gym.make(args.task) for _ in range(args.training_num)])
# test_envs = gym.make(args.task)
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)
train_envs.seed(args.seed)
test_envs.seed(args.seed)
# model
net = Net(args.state_shape,
hidden_sizes=args.hidden_sizes,
device=args.device)
actor = ActorProb(net,
args.action_shape,
max_action=args.max_action,
device=args.device).to(args.device)
critic = Critic(Net(args.state_shape,
hidden_sizes=args.hidden_sizes,
device=args.device),
device=args.device).to(args.device)
actor_critic = ActorCritic(actor, critic)
# orthogonal initialization
for m in actor_critic.modules():
if isinstance(m, torch.nn.Linear):
torch.nn.init.orthogonal_(m.weight)
torch.nn.init.zeros_(m.bias)
optim = torch.optim.Adam(actor_critic.parameters(), lr=args.lr)
# replace DiagGuassian with Independent(Normal) which is equivalent
# pass *logits to be consistent with policy.forward
def dist(*logits):
return Independent(Normal(*logits), 1)
policy = PPOPolicy(
actor,
critic,
optim,
dist,
discount_factor=args.gamma,
max_grad_norm=args.max_grad_norm,
eps_clip=args.eps_clip,
vf_coef=args.vf_coef,
ent_coef=args.ent_coef,
reward_normalization=args.rew_norm,
advantage_normalization=args.norm_adv,
recompute_advantage=args.recompute_adv,
dual_clip=args.dual_clip,
value_clip=args.value_clip,
gae_lambda=args.gae_lambda,
action_space=env.action_space,
)
# collector
train_collector = Collector(
policy, train_envs,
VectorReplayBuffer(args.buffer_size, len(train_envs)))
test_collector = Collector(policy, test_envs)
# log
log_path = os.path.join(args.logdir, args.task, "ppo")
writer = SummaryWriter(log_path)
logger = TensorboardLogger(writer, save_interval=args.save_interval)
def save_best_fn(policy):
torch.save(policy.state_dict(), os.path.join(log_path, "policy.pth"))
def stop_fn(mean_rewards):
return mean_rewards >= args.reward_threshold
def save_checkpoint_fn(epoch, env_step, gradient_step):
# see also: https://pytorch.org/tutorials/beginner/saving_loading_models.html
ckpt_path = os.path.join(log_path, "checkpoint.pth")
# Example: saving by epoch num
# ckpt_path = os.path.join(log_path, f"checkpoint_{epoch}.pth")
torch.save(
{
"model": policy.state_dict(),
"optim": optim.state_dict(),
}, ckpt_path)
return ckpt_path
if args.resume:
# load from existing checkpoint
print(f"Loading agent under {log_path}")
ckpt_path = os.path.join(log_path, "checkpoint.pth")
if os.path.exists(ckpt_path):
checkpoint = torch.load(ckpt_path, map_location=args.device)
policy.load_state_dict(checkpoint["model"])
optim.load_state_dict(checkpoint["optim"])
print("Successfully restore policy and optim.")
else:
print("Fail to restore policy and optim.")
# trainer
trainer = OnpolicyTrainer(
policy,
train_collector,
test_collector,
args.epoch,
args.step_per_epoch,
args.repeat_per_collect,
args.test_num,
args.batch_size,
episode_per_collect=args.episode_per_collect,
stop_fn=stop_fn,
save_best_fn=save_best_fn,
logger=logger,
resume_from_log=args.resume,
save_checkpoint_fn=save_checkpoint_fn,
)
for epoch, epoch_stat, info in trainer:
print(f"Epoch: {epoch}")
print(epoch_stat)
print(info)
assert stop_fn(info["best_reward"])
if __name__ == "__main__":
pprint.pprint(info)
# Let's watch its performance!
env = gym.make(args.task)
policy.eval()
collector = Collector(policy, env)
result = collector.collect(n_episode=1, render=args.render)
rews, lens = result["rews"], result["lens"]
print(f"Final reward: {rews.mean()}, length: {lens.mean()}")
def test_ppo(args=get_args()):
torch.set_num_threads(1) # for poor CPU
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
# train_envs = gym.make(args.task)
# you can also use tianshou.env.SubprocVectorEnv
train_envs = VectorEnv(
[lambda: gym.make(args.task) for _ in range(args.training_num)])
# test_envs = gym.make(args.task)
test_envs = VectorEnv(
[lambda: gym.make(args.task) 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)
# model
net = Net(args.layer_num, args.state_shape, device=args.device)
actor = Actor(net, args.action_shape).to(args.device)
critic = Critic(net).to(args.device)
# orthogonal initialization
for m in list(actor.modules()) + list(critic.modules()):
if isinstance(m, torch.nn.Linear):
torch.nn.init.orthogonal_(m.weight)
torch.nn.init.zeros_(m.bias)
optim = torch.optim.Adam(list(actor.parameters()) +
list(critic.parameters()),
lr=args.lr)
dist = torch.distributions.Categorical
policy = PPOPolicy(actor,
critic,
optim,
dist,
args.gamma,
max_grad_norm=args.max_grad_norm,
eps_clip=args.eps_clip,
vf_coef=args.vf_coef,
ent_coef=args.ent_coef,
action_range=None,
gae_lambda=args.gae_lambda,
reward_normalization=args.rew_norm,
dual_clip=args.dual_clip,
value_clip=args.value_clip)
# collector
train_collector = Collector(policy, train_envs,
ReplayBuffer(args.buffer_size))
test_collector = Collector(policy, test_envs)
# log
log_path = os.path.join(args.logdir, args.task, 'ppo')
writer = SummaryWriter(log_path)
def save_fn(policy):
torch.save(policy.state_dict(), os.path.join(log_path, 'policy.pth'))
def stop_fn(x):
return x >= env.spec.reward_threshold
# trainer
result = onpolicy_trainer(policy,
train_collector,
test_collector,
args.epoch,
args.step_per_epoch,
args.collect_per_step,
args.repeat_per_collect,
args.test_num,
args.batch_size,
stop_fn=stop_fn,
save_fn=save_fn,
writer=writer)
assert stop_fn(result['best_reward'])
train_collector.close()
test_collector.close()
if __name__ == '__main__':
pprint.pprint(result)
# Let's watch its performance!
env = gym.make(args.task)
collector = Collector(policy, env)
result = collector.collect(n_episode=1, render=args.render)
print(f'Final reward: {result["rew"]}, length: {result["len"]}')
collector.close()
def test_ppo(args=get_args()):
env, train_envs, test_envs = make_atari_env(
args.task,
args.seed,
args.training_num,
args.test_num,
scale=args.scale_obs,
frame_stack=args.frames_stack,
)
args.state_shape = env.observation_space.shape or env.observation_space.n
args.action_shape = env.action_space.shape or env.action_space.n
# should be N_FRAMES x H x W
print("Observations shape:", args.state_shape)
print("Actions shape:", args.action_shape)
# seed
np.random.seed(args.seed)
torch.manual_seed(args.seed)
# define model
net = DQN(*args.state_shape,
args.action_shape,
device=args.device,
features_only=True,
output_dim=args.hidden_size)
actor = Actor(net,
args.action_shape,
device=args.device,
softmax_output=False)
critic = Critic(net, device=args.device)
optim = torch.optim.Adam(ActorCritic(actor, critic).parameters(),
lr=args.lr)
lr_scheduler = None
if args.lr_decay:
# decay learning rate to 0 linearly
max_update_num = np.ceil(
args.step_per_epoch / args.step_per_collect) * args.epoch
lr_scheduler = LambdaLR(
optim, lr_lambda=lambda epoch: 1 - epoch / max_update_num)
# define policy
def dist(p):
return torch.distributions.Categorical(logits=p)
policy = PPOPolicy(
actor,
critic,
optim,
dist,
discount_factor=args.gamma,
gae_lambda=args.gae_lambda,
max_grad_norm=args.max_grad_norm,
vf_coef=args.vf_coef,
ent_coef=args.ent_coef,
reward_normalization=args.rew_norm,
action_scaling=False,
lr_scheduler=lr_scheduler,
action_space=env.action_space,
eps_clip=args.eps_clip,
value_clip=args.value_clip,
dual_clip=args.dual_clip,
advantage_normalization=args.norm_adv,
recompute_advantage=args.recompute_adv,
).to(args.device)
if args.icm_lr_scale > 0:
feature_net = DQN(*args.state_shape,
args.action_shape,
args.device,
features_only=True)
action_dim = np.prod(args.action_shape)
feature_dim = feature_net.output_dim
icm_net = IntrinsicCuriosityModule(
feature_net.net,
feature_dim,
action_dim,
hidden_sizes=args.hidden_sizes,
device=args.device,
)
icm_optim = torch.optim.Adam(icm_net.parameters(), lr=args.lr)
policy = ICMPolicy(policy, icm_net, icm_optim, args.icm_lr_scale,
args.icm_reward_scale,
args.icm_forward_loss_weight).to(args.device)
# 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)
# replay buffer: `save_last_obs` and `stack_num` can be removed together
# when you have enough RAM
buffer = VectorReplayBuffer(
args.buffer_size,
buffer_num=len(train_envs),
ignore_obs_next=True,
save_only_last_obs=True,
stack_num=args.frames_stack,
)
# collector
train_collector = Collector(policy,
train_envs,
buffer,
exploration_noise=True)
test_collector = Collector(policy, test_envs, exploration_noise=True)
# log
now = datetime.datetime.now().strftime("%y%m%d-%H%M%S")
args.algo_name = "ppo_icm" if args.icm_lr_scale > 0 else "ppo"
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 stop_fn(mean_rewards):
if env.spec.reward_threshold:
return mean_rewards >= env.spec.reward_threshold
elif "Pong" in args.task:
return mean_rewards >= 20
else:
return False
def save_checkpoint_fn(epoch, env_step, gradient_step):
# see also: https://pytorch.org/tutorials/beginner/saving_loading_models.html
ckpt_path = os.path.join(log_path, "checkpoint.pth")
torch.save({"model": policy.state_dict()}, ckpt_path)
return ckpt_path
# watch agent's performance
def watch():
print("Setup test envs ...")
policy.eval()
test_envs.seed(args.seed)
if args.save_buffer_name:
print(f"Generate buffer with size {args.buffer_size}")
buffer = VectorReplayBuffer(
args.buffer_size,
buffer_num=len(test_envs),
ignore_obs_next=True,
save_only_last_obs=True,
stack_num=args.frames_stack,
)
collector = Collector(policy,
test_envs,
buffer,
exploration_noise=True)
result = collector.collect(n_step=args.buffer_size)
print(f"Save buffer into {args.save_buffer_name}")
# Unfortunately, pickle will cause oom with 1M buffer size
buffer.save_hdf5(args.save_buffer_name)
else:
print("Testing agent ...")
test_collector.reset()
result = test_collector.collect(n_episode=args.test_num,
render=args.render)
rew = result["rews"].mean()
print(f"Mean reward (over {result['n/ep']} episodes): {rew}")
if args.watch:
watch()
exit(0)
# test train_collector and start filling replay buffer
train_collector.collect(n_step=args.batch_size * args.training_num)
# trainer
result = onpolicy_trainer(
policy,
train_collector,
test_collector,
args.epoch,
args.step_per_epoch,
args.repeat_per_collect,
args.test_num,
args.batch_size,
step_per_collect=args.step_per_collect,
stop_fn=stop_fn,
save_best_fn=save_best_fn,
logger=logger,
test_in_train=False,
resume_from_log=args.resume_id is not None,
save_checkpoint_fn=save_checkpoint_fn,
)
pprint.pprint(result)
watch()
