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()}")
def test_discrete_cql(args=get_args()):
# envs
env = make_atari_env(args)
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)
# make environments
test_envs = ShmemVectorEnv(
[lambda: make_atari_env_watch(args) for _ in range(args.test_num)]
)
# seed
np.random.seed(args.seed)
torch.manual_seed(args.seed)
test_envs.seed(args.seed)
# model
net = QRDQN(*args.state_shape, args.action_shape, args.num_quantiles, args.device)
optim = torch.optim.Adam(net.parameters(), lr=args.lr)
# define policy
policy = DiscreteCQLPolicy(
net,
optim,
args.gamma,
args.num_quantiles,
args.n_step,
args.target_update_freq,
min_q_weight=args.min_q_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)
# buffer
assert os.path.exists(args.load_buffer_name), \
"Please run atari_qrdqn.py first to get expert's data buffer."
if args.load_buffer_name.endswith('.pkl'):
buffer = pickle.load(open(args.load_buffer_name, "rb"))
elif args.load_buffer_name.endswith('.hdf5'):
buffer = VectorReplayBuffer.load_hdf5(args.load_buffer_name)
else:
print(f"Unknown buffer format: {args.load_buffer_name}")
exit(0)
# collector
test_collector = Collector(policy, test_envs, exploration_noise=True)
# log
log_path = os.path.join(
args.logdir, args.task, 'cql',
f'seed_{args.seed}_{datetime.datetime.now().strftime("%m%d-%H%M%S")}'
)
writer = SummaryWriter(log_path)
writer.add_text("args", str(args))
logger = TensorboardLogger(writer, update_interval=args.log_interval)
def save_fn(policy):
torch.save(policy.state_dict(), os.path.join(log_path, 'policy.pth'))
def stop_fn(mean_rewards):
return False
# watch agent's performance
def watch():
print("Setup test envs ...")
policy.eval()
policy.set_eps(args.eps_test)
test_envs.seed(args.seed)
print("Testing agent ...")
test_collector.reset()
result = test_collector.collect(n_episode=args.test_num, render=args.render)
pprint.pprint(result)
rew = result["rews"].mean()
print(f'Mean reward (over {result["n/ep"]} episodes): {rew}')
if args.watch:
watch()
exit(0)
result = offline_trainer(
policy,
buffer,
test_collector,
args.epoch,
args.update_per_epoch,
args.test_num,
args.batch_size,
stop_fn=stop_fn,
save_fn=save_fn,
logger=logger
)
pprint.pprint(result)
watch()
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()
def gather_data():
"""Return expert buffer data."""
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,
unbounded=True,
).to(args.device)
actor_optim = torch.optim.Adam(actor.parameters(), lr=args.actor_lr)
net_c1 = Net(
args.state_shape,
args.action_shape,
hidden_sizes=args.hidden_sizes,
concat=True,
device=args.device,
)
critic1 = Critic(net_c1, device=args.device).to(args.device)
critic1_optim = torch.optim.Adam(critic1.parameters(), lr=args.critic_lr)
net_c2 = Net(
args.state_shape,
args.action_shape,
hidden_sizes=args.hidden_sizes,
concat=True,
device=args.device,
)
critic2 = Critic(net_c2, device=args.device).to(args.device)
critic2_optim = torch.optim.Adam(critic2.parameters(), lr=args.critic_lr)
if args.auto_alpha:
target_entropy = -np.prod(env.action_space.shape)
log_alpha = torch.zeros(1, requires_grad=True, device=args.device)
alpha_optim = torch.optim.Adam([log_alpha], lr=args.alpha_lr)
args.alpha = (target_entropy, log_alpha, alpha_optim)
policy = SACPolicy(
actor,
actor_optim,
critic1,
critic1_optim,
critic2,
critic2_optim,
tau=args.tau,
gamma=args.gamma,
alpha=args.alpha,
reward_normalization=args.rew_norm,
estimation_step=args.n_step,
action_space=env.action_space,
)
# collector
buffer = VectorReplayBuffer(args.buffer_size, len(train_envs))
train_collector = Collector(policy,
train_envs,
buffer,
exploration_noise=True)
test_collector = Collector(policy, test_envs)
# train_collector.collect(n_step=args.buffer_size)
# log
log_path = os.path.join(args.logdir, args.task, 'sac')
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
offpolicy_trainer(
policy,
train_collector,
test_collector,
args.epoch,
args.step_per_epoch,
args.step_per_collect,
args.test_num,
args.batch_size,
update_per_step=args.update_per_step,
save_best_fn=save_best_fn,
stop_fn=stop_fn,
logger=logger,
)
train_collector.reset()
result = train_collector.collect(n_step=args.buffer_size)
rews, lens = result["rews"], result["lens"]
print(f"Final reward: {rews.mean()}, length: {lens.mean()}")
if args.save_buffer_name.endswith(".hdf5"):
buffer.save_hdf5(args.save_buffer_name)
else:
pickle.dump(buffer, open(args.save_buffer_name, "wb"))
return buffer
def test_a2c_with_il(args=get_args()):
# if you want to use python vector env, please refer to other test scripts
train_envs = env = envpool.make_gym(args.task,
num_envs=args.training_num,
seed=args.seed)
test_envs = envpool.make_gym(args.task,
num_envs=args.test_num,
seed=args.seed)
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)
# seed
np.random.seed(args.seed)
torch.manual_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)
optim = torch.optim.Adam(ActorCritic(actor, critic).parameters(),
lr=args.lr)
dist = torch.distributions.Categorical
policy = A2CPolicy(actor,
critic,
optim,
dist,
discount_factor=args.gamma,
gae_lambda=args.gae_lambda,
vf_coef=args.vf_coef,
ent_coef=args.ent_coef,
max_grad_norm=args.max_grad_norm,
reward_normalization=args.rew_norm,
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, 'a2c')
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,
episode_per_collect=args.episode_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()}")
policy.eval()
# here we define an imitation collector with a trivial policy
# if args.task == 'CartPole-v0':
# env.spec.reward_threshold = 190 # lower the goal
net = Net(args.state_shape,
hidden_sizes=args.hidden_sizes,
device=args.device)
net = Actor(net, args.action_shape, device=args.device).to(args.device)
optim = torch.optim.Adam(net.parameters(), lr=args.il_lr)
il_policy = ImitationPolicy(net, optim, action_space=env.action_space)
il_test_collector = Collector(
il_policy,
envpool.make_gym(args.task, num_envs=args.test_num, seed=args.seed),
)
train_collector.reset()
result = offpolicy_trainer(il_policy,
train_collector,
il_test_collector,
args.epoch,
args.il_step_per_epoch,
args.step_per_collect,
args.test_num,
args.batch_size,
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)
il_policy.eval()
collector = Collector(il_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_sac_bipedal(args=get_args()):
env = Wrapper(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]
train_envs = SubprocVectorEnv(
[lambda: Wrapper(gym.make(args.task)) for _ in range(args.training_num)]
)
# test_envs = gym.make(args.task)
test_envs = SubprocVectorEnv(
[
lambda: Wrapper(gym.make(args.task), reward_scale=1, rm_done=False)
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_a = Net(args.state_shape, hidden_sizes=args.hidden_sizes, device=args.device)
actor = ActorProb(
net_a,
args.action_shape,
max_action=args.max_action,
device=args.device,
unbounded=True
).to(args.device)
actor_optim = torch.optim.Adam(actor.parameters(), lr=args.actor_lr)
net_c1 = Net(
args.state_shape,
args.action_shape,
hidden_sizes=args.hidden_sizes,
concat=True,
device=args.device
)
critic1 = Critic(net_c1, device=args.device).to(args.device)
critic1_optim = torch.optim.Adam(critic1.parameters(), lr=args.critic_lr)
net_c2 = Net(
args.state_shape,
args.action_shape,
hidden_sizes=args.hidden_sizes,
concat=True,
device=args.device
)
critic2 = Critic(net_c2, device=args.device).to(args.device)
critic2_optim = torch.optim.Adam(critic2.parameters(), lr=args.critic_lr)
if args.auto_alpha:
target_entropy = -np.prod(env.action_space.shape)
log_alpha = torch.zeros(1, requires_grad=True, device=args.device)
alpha_optim = torch.optim.Adam([log_alpha], lr=args.alpha_lr)
args.alpha = (target_entropy, log_alpha, alpha_optim)
policy = SACPolicy(
actor,
actor_optim,
critic1,
critic1_optim,
critic2,
critic2_optim,
tau=args.tau,
gamma=args.gamma,
alpha=args.alpha,
estimation_step=args.n_step,
action_space=env.action_space
)
# load a previous policy
if args.resume_path:
policy.load_state_dict(torch.load(args.resume_path))
print("Loaded agent from: ", args.resume_path)
# collector
train_collector = Collector(
policy,
train_envs,
VectorReplayBuffer(args.buffer_size, len(train_envs)),
exploration_noise=True
)
test_collector = Collector(policy, test_envs)
# train_collector.collect(n_step=args.buffer_size)
# log
log_path = os.path.join(args.logdir, args.task, 'sac')
writer = SummaryWriter(log_path)
logger = TensorboardLogger(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 = offpolicy_trainer(
policy,
train_collector,
test_collector,
args.epoch,
args.step_per_epoch,
args.step_per_collect,
args.test_num,
args.batch_size,
update_per_step=args.update_per_step,
test_in_train=False,
stop_fn=stop_fn,
save_fn=save_fn,
logger=logger
)
if __name__ == '__main__':
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)
rews, lens = result["rews"], result["lens"]
print(f"Final reward: {rews.mean()}, length: {lens.mean()}")
def test_ddpg(args=get_args()):
env, train_envs, test_envs = make_mujoco_env(args.task,
args.seed,
args.training_num,
args.test_num,
obs_norm=False)
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]
args.exploration_noise = args.exploration_noise * args.max_action
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))
# seed
np.random.seed(args.seed)
torch.manual_seed(args.seed)
# model
net_a = Net(args.state_shape,
hidden_sizes=args.hidden_sizes,
device=args.device)
actor = Actor(net_a,
args.action_shape,
max_action=args.max_action,
device=args.device).to(args.device)
actor_optim = torch.optim.Adam(actor.parameters(), lr=args.actor_lr)
net_c = Net(
args.state_shape,
args.action_shape,
hidden_sizes=args.hidden_sizes,
concat=True,
device=args.device,
)
critic = Critic(net_c, device=args.device).to(args.device)
critic_optim = torch.optim.Adam(critic.parameters(), lr=args.critic_lr)
policy = DDPGPolicy(
actor,
actor_optim,
critic,
critic_optim,
tau=args.tau,
gamma=args.gamma,
exploration_noise=GaussianNoise(sigma=args.exploration_noise),
estimation_step=args.n_step,
action_space=env.action_space,
)
# 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)
train_collector.collect(n_step=args.start_timesteps, random=True)
# log
now = datetime.datetime.now().strftime("%y%m%d-%H%M%S")
args.algo_name = "ddpg"
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"))
if not args.watch:
# trainer
result = offpolicy_trainer(
policy,
train_collector,
test_collector,
args.epoch,
args.step_per_epoch,
args.step_per_collect,
args.test_num,
args.batch_size,
save_best_fn=save_best_fn,
logger=logger,
update_per_step=args.update_per_step,
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_discrete_sac(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": 180} # lower the goal
args.reward_threshold = default_reward_threshold.get(
args.task, env.spec.reward_threshold)
train_envs = DummyVectorEnv(
[lambda: gym.make(args.task) for _ in range(args.training_num)])
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,
softmax_output=False,
device=args.device).to(args.device)
actor_optim = torch.optim.Adam(actor.parameters(), lr=args.actor_lr)
net_c1 = Net(args.state_shape,
hidden_sizes=args.hidden_sizes,
device=args.device)
critic1 = Critic(net_c1, last_size=args.action_shape,
device=args.device).to(args.device)
critic1_optim = torch.optim.Adam(critic1.parameters(), lr=args.critic_lr)
net_c2 = Net(args.state_shape,
hidden_sizes=args.hidden_sizes,
device=args.device)
critic2 = Critic(net_c2, last_size=args.action_shape,
device=args.device).to(args.device)
critic2_optim = torch.optim.Adam(critic2.parameters(), lr=args.critic_lr)
# better not to use auto alpha in CartPole
if args.auto_alpha:
target_entropy = 0.98 * np.log(np.prod(args.action_shape))
log_alpha = torch.zeros(1, requires_grad=True, device=args.device)
alpha_optim = torch.optim.Adam([log_alpha], lr=args.alpha_lr)
args.alpha = (target_entropy, log_alpha, alpha_optim)
policy = DiscreteSACPolicy(actor,
actor_optim,
critic1,
critic1_optim,
critic2,
critic2_optim,
args.tau,
args.gamma,
args.alpha,
estimation_step=args.n_step,
reward_normalization=args.rew_norm)
# collector
train_collector = Collector(
policy, train_envs,
VectorReplayBuffer(args.buffer_size, len(train_envs)))
test_collector = Collector(policy, test_envs)
# train_collector.collect(n_step=args.buffer_size)
# log
log_path = os.path.join(args.logdir, args.task, 'discrete_sac')
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 = offpolicy_trainer(policy,
train_collector,
test_collector,
args.epoch,
args.step_per_epoch,
args.step_per_collect,
args.test_num,
args.batch_size,
stop_fn=stop_fn,
save_best_fn=save_best_fn,
logger=logger,
update_per_step=args.update_per_step,
test_in_train=False)
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()}")