def test_psrl(args=get_args()):
env = gym.make(args.task)
if args.task == "NChain-v0":
env.spec.reward_threshold = 3647 # described in PSRL paper
print("reward threshold:", env.spec.reward_threshold)
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 = 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 = 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
n_action = args.action_shape
n_state = args.state_shape
trans_count_prior = np.ones((n_state, n_action, n_state))
rew_mean_prior = np.full((n_state, n_action), args.rew_mean_prior)
rew_std_prior = np.full((n_state, n_action), args.rew_std_prior)
policy = PSRLPolicy(
trans_count_prior, rew_mean_prior, rew_std_prior, args.gamma, args.eps,
args.add_done_loop)
# collector
train_collector = Collector(
policy, train_envs, ReplayBuffer(args.buffer_size))
test_collector = Collector(policy, test_envs)
# log
writer = SummaryWriter(args.logdir + '/' + args.task)
def stop_fn(x):
if env.spec.reward_threshold:
return x >= env.spec.reward_threshold
else:
return False
train_collector.collect(n_step=args.buffer_size, random=True)
# trainer
result = onpolicy_trainer(
policy, train_collector, test_collector, args.epoch,
args.step_per_epoch, args.collect_per_step, 1,
args.test_num, 0, stop_fn=stop_fn, writer=writer,
test_in_train=False)
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=[1] * args.test_num,
render=args.render)
print(f'Final reward: {result["rew"]}, length: {result["len"]}')
elif env.spec.reward_threshold:
assert result["best_reward"] >= env.spec.reward_threshold
def test_ddpg(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]
# 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
net = Net(args.layer_num, args.state_shape, device=args.device)
actor = Actor(net, args.action_shape, args.max_action,
args.device).to(args.device)
actor_optim = torch.optim.Adam(actor.parameters(), lr=args.actor_lr)
net = Net(args.layer_num, args.state_shape,
args.action_shape, concat=True, device=args.device)
critic = Critic(net, args.device).to(args.device)
critic_optim = torch.optim.Adam(critic.parameters(), lr=args.critic_lr)
policy = DDPGPolicy(
actor, actor_optim, critic, critic_optim,
action_range=[env.action_space.low[0], env.action_space.high[0]],
tau=args.tau, gamma=args.gamma,
exploration_noise=GaussianNoise(sigma=args.exploration_noise),
reward_normalization=True, ignore_done=True)
# collector
train_collector = Collector(
policy, train_envs, ReplayBuffer(args.buffer_size))
test_collector = Collector(policy, test_envs)
# log
writer = SummaryWriter(args.logdir + '/' + 'ddpg')
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.collect_per_step, args.test_num,
args.batch_size, stop_fn=stop_fn, writer=writer)
assert stop_fn(result['best_reward'])
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=[1] * args.test_num,
render=args.render)
print(f'Final reward: {result["rew"]}, length: {result["len"]}')
def test_collector_nstep():
policy = MyPolicy()
env_fns = [lambda x=i: MyTestEnv(size=x) for i in np.arange(2, 11)]
dum = DummyVectorEnv(env_fns)
num = len(env_fns)
c3 = Collector(policy, dum, VectorReplayBuffer(total_size=40000, buffer_num=num))
for i in tqdm.trange(1, 400, desc="test step collector n_step"):
c3.reset()
result = c3.collect(n_step=i * len(env_fns))
assert result['n/st'] >= i
def train(net, save_path):
landlord_path = save_path
landlord_policy, upper_policy, lower_policy = get_agents()
train_envs = DummyVectorEnv([
lambda: LandlordEnv(upper_policy, lower_policy)
for _ in range(args.training_num)
])
test_envs = DummyVectorEnv([
lambda: LandlordEnv(upper_policy, lower_policy)
for _ in range(args.test_num)
])
# train_envs = SubprocVectorEnv(
# [lambda: LandlordEnv(upper_policy, lower_policy) for _ in range(args.training_num)])
# test_envs = SubprocVectorEnv(
# [lambda: LandlordEnv(upper_policy, lower_policy) 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)
# collect expert experiences
train_collector = Collector(landlord_policy, train_envs,
ReplayBuffer(args.buffer_size))
# either create or load network
optim = get_optim(net, args.il_lr)
# build test policy
il_policy = MyImitationPolicy(net, optim)
il_test_collector = Collector(il_policy, test_envs)
train_collector.reset()
# il_test_collector.collect(n_episode=1)
# writer
log_path = path.join('models', 'landlord_il')
writer = SummaryWriter(log_path)
def stop_fn(x):
return x > -0.2
def save_fn(policy):
torch.save(policy.model.state_dict(), landlord_path)
result = offpolicy_trainer(il_policy,
train_collector,
il_test_collector,
args.epoch,
args.step_per_epoch // 5,
args.collect_per_step,
args.test_num,
args.batch_size,
stop_fn=stop_fn,
save_fn=save_fn,
writer=writer)
print(f'Final reward: {result["rew"]}, length: {result["len"]}')
def test_sac_bipedal(args=get_args()):
env = EnvWrapper(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: EnvWrapper(args.task) for _ in range(args.training_num)])
# test_envs = gym.make(args.task)
test_envs = SubprocVectorEnv([lambda: EnvWrapper(args.task, reward_scale=1)
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.layer_num, args.state_shape, device=args.device)
actor = ActorProb(
net_a, args.action_shape, args.max_action, args.device, unbounded=True
).to(args.device)
actor_optim = torch.optim.Adam(actor.parameters(), lr=args.actor_lr)
net_c1 = Net(args.layer_num, args.state_shape,
args.action_shape, concat=True, device=args.device)
critic1 = Critic(net_c1, args.device).to(args.device)
critic1_optim = torch.optim.Adam(critic1.parameters(), lr=args.critic_lr)
net_c2 = Net(args.layer_num, args.state_shape,
args.action_shape, concat=True, device=args.device)
critic2 = Critic(net_c2, 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,
action_range=[env.action_space.low[0], env.action_space.high[0]],
tau=args.tau, gamma=args.gamma, alpha=args.alpha,
reward_normalization=args.rew_norm,
ignore_done=args.ignore_done,
estimation_step=args.n_step)
# 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, ReplayBuffer(args.buffer_size))
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)
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.collect_per_step, args.test_num,
args.batch_size, stop_fn=stop_fn, save_fn=save_fn, writer=writer,
test_in_train=False)
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=[1] * args.test_num,
render=args.render)
print(f'Final reward: {result["rew"]}, length: {result["len"]}')
def test_ddpg(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]
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))
# train_envs = gym.make(args.task)
if args.training_num > 1:
train_envs = SubprocVectorEnv(
[lambda: gym.make(args.task) for _ in range(args.training_num)])
else:
train_envs = gym.make(args.task)
# 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
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
log_path = os.path.join(
args.logdir, args.task, 'ddpg', 'seed_' + str(args.seed) + '_' +
datetime.datetime.now().strftime('%m%d_%H%M%S') + '-' +
args.task.replace('-', '_') + '_ddpg')
writer = SummaryWriter(log_path)
writer.add_text("args", str(args))
logger = BasicLogger(writer)
def save_fn(policy):
torch.save(policy.state_dict(), os.path.join(log_path, 'policy.pth'))
# 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_fn=save_fn,
logger=logger,
update_per_step=args.update_per_step,
test_in_train=False)
# 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_sac_with_il(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)
actor_optim = torch.optim.Adam(actor.parameters(), lr=args.actor_lr)
critic1 = Critic(args.layer_num, args.state_shape, args.action_shape,
args.device).to(args.device)
critic1_optim = torch.optim.Adam(critic1.parameters(), lr=args.critic_lr)
critic2 = Critic(args.layer_num, args.state_shape, args.action_shape,
args.device).to(args.device)
critic2_optim = torch.optim.Adam(critic2.parameters(), lr=args.critic_lr)
policy = SACPolicy(actor,
actor_optim,
critic1,
critic1_optim,
critic2,
critic2_optim,
args.tau,
args.gamma,
args.alpha,
[env.action_space.low[0], env.action_space.high[0]],
reward_normalization=args.rew_norm,
ignore_done=args.ignore_done,
estimation_step=args.n_step)
# collector
train_collector = Collector(policy, train_envs,
ReplayBuffer(args.buffer_size))
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)
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 = offpolicy_trainer(policy,
train_collector,
test_collector,
args.epoch,
args.step_per_epoch,
args.collect_per_step,
args.test_num,
args.batch_size,
stop_fn=stop_fn,
save_fn=save_fn,
writer=writer)
assert stop_fn(result['best_reward'])
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()
# here we define an imitation collector with a trivial policy
if args.task == 'Pendulum-v0':
env.spec.reward_threshold = -300 # lower the goal
net = Actor(1, args.state_shape, args.action_shape, args.max_action,
args.device).to(args.device)
optim = torch.optim.Adam(net.parameters(), lr=args.il_lr)
il_policy = ImitationPolicy(net, optim, mode='continuous')
il_test_collector = Collector(il_policy, test_envs)
train_collector.reset()
result = offpolicy_trainer(il_policy,
train_collector,
il_test_collector,
args.epoch,
args.step_per_epoch // 5,
args.collect_per_step,
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()
il_test_collector.close()
if __name__ == '__main__':
pprint.pprint(result)
# Let's watch its performance!
env = gym.make(args.task)
collector = Collector(il_policy, env)
result = collector.collect(n_episode=1, render=args.render)
print(f'Final reward: {result["rew"]}, length: {result["len"]}')
collector.close()
def test_bcq():
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] # float
print("device:", args.device)
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))
args.state_dim = args.state_shape[0]
args.action_dim = args.action_shape[0]
print("Max_action", args.max_action)
# 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)
test_envs.seed(args.seed)
# model
# perturbation network
net_a = MLP(
input_dim=args.state_dim + args.action_dim,
output_dim=args.action_dim,
hidden_sizes=args.hidden_sizes,
device=args.device,
)
actor = Perturbation(net_a,
max_action=args.max_action,
device=args.device,
phi=args.phi).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,
)
net_c2 = 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)
critic2 = Critic(net_c2, device=args.device).to(args.device)
critic2_optim = torch.optim.Adam(critic2.parameters(), lr=args.critic_lr)
# vae
# output_dim = 0, so the last Module in the encoder is ReLU
vae_encoder = MLP(
input_dim=args.state_dim + args.action_dim,
hidden_sizes=args.vae_hidden_sizes,
device=args.device,
)
if not args.latent_dim:
args.latent_dim = args.action_dim * 2
vae_decoder = MLP(
input_dim=args.state_dim + args.latent_dim,
output_dim=args.action_dim,
hidden_sizes=args.vae_hidden_sizes,
device=args.device,
)
vae = VAE(
vae_encoder,
vae_decoder,
hidden_dim=args.vae_hidden_sizes[-1],
latent_dim=args.latent_dim,
max_action=args.max_action,
device=args.device,
).to(args.device)
vae_optim = torch.optim.Adam(vae.parameters())
policy = BCQPolicy(
actor,
actor_optim,
critic1,
critic1_optim,
critic2,
critic2_optim,
vae,
vae_optim,
device=args.device,
gamma=args.gamma,
tau=args.tau,
lmbda=args.lmbda,
)
# 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
test_collector = Collector(policy, test_envs)
# log
now = datetime.datetime.now().strftime("%y%m%d-%H%M%S")
args.algo_name = "bcq"
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 watch():
if args.resume_path is None:
args.resume_path = os.path.join(log_path, "policy.pth")
policy.load_state_dict(
torch.load(args.resume_path, map_location=torch.device("cpu")))
policy.eval()
collector = Collector(policy, env)
collector.collect(n_episode=1, render=1 / 35)
if not args.watch:
dataset = d4rl.qlearning_dataset(gym.make(args.expert_data_task))
dataset_size = dataset["rewards"].size
print("dataset_size", dataset_size)
replay_buffer = ReplayBuffer(dataset_size)
for i in range(dataset_size):
replay_buffer.add(
Batch(
obs=dataset["observations"][i],
act=dataset["actions"][i],
rew=dataset["rewards"][i],
done=dataset["terminals"][i],
obs_next=dataset["next_observations"][i],
))
print("dataset loaded")
# trainer
result = offline_trainer(
policy,
replay_buffer,
test_collector,
args.epoch,
args.step_per_epoch,
args.test_num,
args.batch_size,
save_best_fn=save_best_fn,
logger=logger,
)
pprint.pprint(result)
else:
watch()
# 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_sac_bipedal(args=get_args()):
torch.set_num_threads(1) # we just need only one thread for NN
env = EnvWrapper(args.task)
def IsStop(reward):
return reward >= env.spec.reward_threshold
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: EnvWrapper(args.task) for _ in range(args.training_num)])
# test_envs = gym.make(args.task)
test_envs = SubprocVectorEnv([
lambda: EnvWrapper(args.task, reward_scale=1)
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.layer_num, args.state_shape, device=args.device)
actor = ActorProb(net_a, args.action_shape, args.max_action,
args.device).to(args.device)
actor_optim = torch.optim.Adam(actor.parameters(), lr=args.actor_lr)
net_c1 = Net(args.layer_num,
args.state_shape,
args.action_shape,
concat=True,
device=args.device)
critic1 = Critic(net_c1, args.device).to(args.device)
critic1_optim = torch.optim.Adam(critic1.parameters(), lr=args.critic_lr)
net_c2 = Net(args.layer_num,
args.state_shape,
args.action_shape,
concat=True,
device=args.device)
critic2 = Critic(net_c2, args.device).to(args.device)
critic2_optim = torch.optim.Adam(critic2.parameters(), lr=args.critic_lr)
policy = SACPolicy(actor,
actor_optim,
critic1,
critic1_optim,
critic2,
critic2_optim,
args.tau,
args.gamma,
args.alpha,
[env.action_space.low[0], env.action_space.high[0]],
reward_normalization=args.rew_norm,
ignore_done=args.ignore_done,
estimation_step=args.n_step)
# collector
train_collector = Collector(policy, train_envs,
ReplayBuffer(args.buffer_size))
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)
def save_fn(policy):
torch.save(policy.state_dict(), os.path.join(log_path, 'policy.pth'))
# trainer
result = offpolicy_trainer(policy,
train_collector,
test_collector,
args.epoch,
args.step_per_epoch,
args.collect_per_step,
args.test_num,
args.batch_size,
stop_fn=IsStop,
save_fn=save_fn,
writer=writer)
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=[1] * args.test_num,
render=args.render)
print(f'Final reward: {result["rew"]}, length: {result["len"]}')
def test_sac_with_il(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,
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)
policy = SACPolicy(
actor,
actor_optim,
critic1,
critic1_optim,
critic2,
critic2_optim,
action_range=[env.action_space.low[0], env.action_space.high[0]],
tau=args.tau,
gamma=args.gamma,
alpha=args.alpha,
reward_normalization=args.rew_norm,
estimation_step=args.n_step)
# 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 = 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 = 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,
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()}")
# here we define an imitation collector with a trivial policy
policy.eval()
if args.task == 'Pendulum-v0':
env.spec.reward_threshold = -300 # lower the goal
net = Actor(Net(args.state_shape,
hidden_sizes=args.imitation_hidden_sizes,
device=args.device),
args.action_shape,
max_action=args.max_action,
device=args.device).to(args.device)
optim = torch.optim.Adam(net.parameters(), lr=args.il_lr)
il_policy = ImitationPolicy(net, optim, mode='continuous')
il_test_collector = Collector(
il_policy,
DummyVectorEnv(
[lambda: gym.make(args.task) for _ in range(args.test_num)]))
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_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)
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(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)
if args.training_num > 1:
train_envs = SubprocVectorEnv(
[lambda: gym.make(args.task) for _ in range(args.training_num)]
)
else:
train_envs = gym.make(args.task)
# 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
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,
conditioned_sigma=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
)
net_c2 = 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)
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, 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
t0 = datetime.datetime.now().strftime("%m%d_%H%M%S")
log_file = f'seed_{args.seed}_{t0}-{args.task.replace("-", "_")}_sac'
log_path = os.path.join(args.logdir, args.task, 'sac', log_file)
writer = SummaryWriter(log_path)
writer.add_text("args", str(args))
logger = TensorboardLogger(writer)
def save_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_fn=save_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_cql():
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] # float
print("device:", args.device)
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))
args.state_dim = args.state_shape[0]
args.action_dim = args.action_shape[0]
print("Max_action", args.max_action)
# 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)
test_envs.seed(args.seed)
# model
# actor network
net_a = Net(
args.state_shape,
args.action_shape,
hidden_sizes=args.hidden_sizes,
device=args.device,
)
actor = ActorProb(net_a,
action_shape=args.action_shape,
max_action=args.max_action,
device=args.device,
unbounded=True,
conditioned_sigma=True).to(args.device)
actor_optim = torch.optim.Adam(actor.parameters(), lr=args.actor_lr)
# critic network
net_c1 = Net(
args.state_shape,
args.action_shape,
hidden_sizes=args.hidden_sizes,
concat=True,
device=args.device,
)
net_c2 = 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)
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 = CQLPolicy(
actor,
actor_optim,
critic1,
critic1_optim,
critic2,
critic2_optim,
cql_alpha_lr=args.cql_alpha_lr,
cql_weight=args.cql_weight,
tau=args.tau,
gamma=args.gamma,
alpha=args.alpha,
temperature=args.temperature,
with_lagrange=args.with_lagrange,
lagrange_threshold=args.lagrange_threshold,
min_action=np.min(env.action_space.low),
max_action=np.max(env.action_space.high),
device=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)
# collector
test_collector = Collector(policy, test_envs)
# log
now = datetime.datetime.now().strftime("%y%m%d-%H%M%S")
args.algo_name = "cql"
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 watch():
if args.resume_path is None:
args.resume_path = os.path.join(log_path, "policy.pth")
policy.load_state_dict(
torch.load(args.resume_path, map_location=torch.device("cpu")))
policy.eval()
collector = Collector(policy, env)
collector.collect(n_episode=1, render=1 / 35)
if not args.watch:
dataset = d4rl.qlearning_dataset(gym.make(args.expert_data_task))
dataset_size = dataset["rewards"].size
print("dataset_size", dataset_size)
replay_buffer = ReplayBuffer(dataset_size)
for i in range(dataset_size):
replay_buffer.add(
Batch(
obs=dataset["observations"][i],
act=dataset["actions"][i],
rew=dataset["rewards"][i],
done=dataset["terminals"][i],
obs_next=dataset["next_observations"][i],
))
print("dataset loaded")
# trainer
result = offline_trainer(
policy,
replay_buffer,
test_collector,
args.epoch,
args.step_per_epoch,
args.test_num,
args.batch_size,
save_best_fn=save_best_fn,
logger=logger,
)
pprint.pprint(result)
else:
watch()
# 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_sac_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
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
# 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 = 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
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_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,
update_per_step=args.update_per_step,
stop_fn=stop_fn,
save_best_fn=save_best_fn,
logger=logger
)
assert stop_fn(result['best_reward'])
# here we define an imitation collector with a trivial policy
policy.eval()
if args.task.startswith("Pendulum"):
args.reward_threshold -= 50 # lower the goal
net = Actor(
Net(
args.state_shape,
hidden_sizes=args.imitation_hidden_sizes,
device=args.device
),
args.action_shape,
max_action=args.max_action,
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,
action_scaling=True,
action_bound_method="clip"
)
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'])
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_psrl(args=get_args()):
env = gym.make(args.task)
if args.task == "NChain-v0":
env.spec.reward_threshold = 3400
# env.spec.reward_threshold = 3647 # described in PSRL paper
print("reward threshold:", env.spec.reward_threshold)
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 = DummyVectorEnv(
[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
n_action = args.action_shape
n_state = args.state_shape
trans_count_prior = np.ones((n_state, n_action, n_state))
rew_mean_prior = np.full((n_state, n_action), args.rew_mean_prior)
rew_std_prior = np.full((n_state, n_action), args.rew_std_prior)
policy = PSRLPolicy(trans_count_prior, rew_mean_prior, rew_std_prior,
args.gamma, args.eps, args.add_done_loop)
# collector
train_collector = Collector(policy,
train_envs,
VectorReplayBuffer(args.buffer_size,
len(train_envs)),
exploration_noise=True)
test_collector = Collector(policy, test_envs)
# Logger
if args.logger == "wandb":
logger = WandbLogger(save_interval=1,
project='psrl',
name='wandb_test',
config=args)
elif args.logger == "tensorboard":
log_path = os.path.join(args.logdir, args.task, 'psrl')
writer = SummaryWriter(log_path)
writer.add_text("args", str(args))
logger = TensorboardLogger(writer)
else:
logger = LazyLogger()
def stop_fn(mean_rewards):
if env.spec.reward_threshold:
return mean_rewards >= env.spec.reward_threshold
else:
return False
train_collector.collect(n_step=args.buffer_size, random=True)
# trainer, test it without logger
result = onpolicy_trainer(
policy,
train_collector,
test_collector,
args.epoch,
args.step_per_epoch,
1,
args.test_num,
0,
episode_per_collect=args.episode_per_collect,
stop_fn=stop_fn,
logger=logger,
test_in_train=False,
)
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()}")
elif env.spec.reward_threshold:
assert result["best_reward"] >= env.spec.reward_threshold
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_dqn(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
# 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 = 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
Q_param = {"hidden_sizes": args.dueling_q_hidden_sizes}
V_param = {"hidden_sizes": args.dueling_v_hidden_sizes}
net = Net(args.state_shape, args.action_shape,
hidden_sizes=args.hidden_sizes, device=args.device,
dueling_param=(Q_param, V_param)).to(args.device)
optim = torch.optim.Adam(net.parameters(), lr=args.lr)
# prepare hyperparameters
adaptive_scheme = args.adaptive_scheme
adaptive_scheme[4] *= args.update_per_step
adaptive_scheme[5] *= args.update_per_step
reweigh_hyper = {
"hard_weight": args.tper_weight,
"linear": args.linear_hp,
"adaptive_linear": args.adaptive_scheme,
}
policy = TPDQNPolicy(
net, optim, args.gamma, args.n_step,
target_update_freq=args.target_update_freq,
bk_step=args.bk_step,
reweigh_type=args.reweigh_type,
reweigh_hyper=reweigh_hyper)
# collector
train_collector = Collector(
policy, train_envs,
TPVectorReplayBuffer(args.buffer_size, len(train_envs)),
preprocess_fn=StepPreprocess(len(train_envs), args.bk_step).get_step,
exploration_noise=True)
test_collector = Collector(policy, test_envs, exploration_noise=True)
# policy.set_eps(1)
train_collector.collect(n_step=args.batch_size * args.training_num)
# log
cur_time = time.strftime('%y-%m-%d-%H-%M-%S', time.localtime())
log_path = os.path.join(args.logdir, args.task, 'tpdqn', "%s-seed%d"%(args.exp, args.seed), cur_time)
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
def train_fn(epoch, env_step): # exp decay
eps = max(args.eps_train * (1 - 5e-6) ** env_step, args.eps_test)
policy.set_eps(eps)
def test_fn(epoch, env_step):
policy.set_eps(args.eps_test)
# 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, stop_fn=stop_fn, train_fn=train_fn,
test_fn=test_fn, save_fn=save_fn, logger=logger)
# assert stop_fn(result['best_reward'])
if __name__ == '__main__':
pprint.pprint(result)
# Let's watch its performance!
policy.eval()
policy.set_eps(args.eps_test)
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_bdq(args=get_args()):
env = gym.make(args.task)
env = ContinuousToDiscrete(env, args.action_per_branch)
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.num_branches = args.action_shape if isinstance(args.action_shape,
int) else args.action_shape[0]
print("Observations shape:", args.state_shape)
print("Num branches:", args.num_branches)
print("Actions per branch:", args.action_per_branch)
# train_envs = ContinuousToDiscrete(gym.make(args.task), args.action_per_branch)
# you can also use tianshou.env.SubprocVectorEnv
train_envs = SubprocVectorEnv(
[
lambda: ContinuousToDiscrete(gym.make(args.task), args.action_per_branch)
for _ in range(args.training_num)
]
)
# test_envs = ContinuousToDiscrete(gym.make(args.task), args.action_per_branch)
test_envs = SubprocVectorEnv(
[
lambda: ContinuousToDiscrete(gym.make(args.task), args.action_per_branch)
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 = BranchingNet(
args.state_shape,
args.num_branches,
args.action_per_branch,
args.common_hidden_sizes,
args.value_hidden_sizes,
args.action_hidden_sizes,
device=args.device,
).to(args.device)
optim = torch.optim.Adam(net.parameters(), lr=args.lr)
policy = BranchingDQNPolicy(
net, optim, args.gamma, target_update_freq=args.target_update_freq
)
# 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=False)
# policy.set_eps(1)
train_collector.collect(n_step=args.batch_size * args.training_num)
# log
current_time = datetime.datetime.now().strftime("%Y%m%d-%H%M%S")
log_path = os.path.join(args.logdir, "bdq", args.task, current_time)
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 >= getattr(env.spec.reward_threshold)
def train_fn(epoch, env_step): # exp decay
eps = max(args.eps_train * (1 - args.eps_decay)**env_step, args.eps_test)
policy.set_eps(eps)
def test_fn(epoch, env_step):
policy.set_eps(args.eps_test)
# 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,
# stop_fn=stop_fn,
train_fn=train_fn,
test_fn=test_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!
policy.eval()
policy.set_eps(args.eps_test)
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_sac(args=get_args()):
torch.set_num_threads(1)
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]
# you can also use tianshou.env.SubprocVectorEnv
# 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
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 = Net(args.state_shape,
args.action_shape,
hidden_sizes=args.hidden_sizes,
concat=True,
device=args.device)
critic1 = Critic(net, device=args.device).to(args.device)
critic1_optim = torch.optim.Adam(critic1.parameters(), lr=args.critic_lr)
net = Net(args.state_shape,
args.action_shape,
hidden_sizes=args.hidden_sizes,
concat=True,
device=args.device)
critic2 = Critic(net, device=args.device).to(args.device)
critic2_optim = torch.optim.Adam(critic2.parameters(), lr=args.critic_lr)
policy = SACPolicy(
actor,
actor_optim,
critic1,
critic1_optim,
critic2,
critic2_optim,
action_range=[env.action_space.low[0], env.action_space.high[0]],
tau=args.tau,
gamma=args.gamma,
alpha=args.alpha,
reward_normalization=True,
ignore_done=True)
# collector
train_collector = Collector(policy, train_envs,
ReplayBuffer(args.buffer_size))
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', args.run_id)
writer = SummaryWriter(log_path)
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.collect_per_step,
args.test_num,
args.batch_size,
stop_fn=stop_fn,
writer=writer,
log_interval=args.log_interval)
assert stop_fn(result['best_reward'])
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=[1] * args.test_num,
render=args.render)
print(f'Final reward: {result["rew"]}, length: {result["len"]}')
def test_bdq(args=get_args()):
env = gym.make(args.task)
env = ContinuousToDiscrete(env, args.action_per_branch)
args.state_shape = env.observation_space.shape or env.observation_space.n
args.num_branches = env.action_space.shape[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)
print("Observations shape:", args.state_shape)
print("Num branches:", args.num_branches)
print("Actions per branch:", args.action_per_branch)
train_envs = DummyVectorEnv([
lambda: ContinuousToDiscrete(gym.make(args.task), args.
action_per_branch)
for _ in range(args.training_num)
])
test_envs = DummyVectorEnv([
lambda: ContinuousToDiscrete(gym.make(
args.task), args.action_per_branch) 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 = BranchingNet(
args.state_shape,
args.num_branches,
args.action_per_branch,
args.common_hidden_sizes,
args.value_hidden_sizes,
args.action_hidden_sizes,
device=args.device,
).to(args.device)
optim = torch.optim.Adam(net.parameters(), lr=args.lr)
policy = BranchingDQNPolicy(net,
optim,
args.gamma,
target_update_freq=args.target_update_freq)
# collector
train_collector = Collector(policy,
train_envs,
VectorReplayBuffer(args.buffer_size,
args.training_num),
exploration_noise=True)
test_collector = Collector(policy, test_envs, exploration_noise=False)
# policy.set_eps(1)
train_collector.collect(n_step=args.batch_size * args.training_num)
def train_fn(epoch, env_step): # exp decay
eps = max(args.eps_train * (1 - args.eps_decay)**env_step,
args.eps_test)
policy.set_eps(eps)
def test_fn(epoch, env_step):
policy.set_eps(args.eps_test)
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,
update_per_step=args.update_per_step,
train_fn=train_fn,
test_fn=test_fn,
stop_fn=stop_fn,
)
# assert stop_fn(result["best_reward"])
if __name__ == "__main__":
pprint.pprint(result)
# Let's watch its performance!
policy.eval()
policy.set_eps(args.eps_test)
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_trpo(args=get_args()):
env, train_envs, test_envs = make_mujoco_env(args.task,
args.seed,
args.training_num,
args.test_num,
obs_norm=True)
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))
# seed
np.random.seed(args.seed)
torch.manual_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, -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(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 = TRPOPolicy(
actor,
critic,
optim,
dist,
discount_factor=args.gamma,
gae_lambda=args.gae_lambda,
reward_normalization=args.rew_norm,
action_scaling=True,
action_bound_method=args.bound_action_method,
lr_scheduler=lr_scheduler,
action_space=env.action_space,
advantage_normalization=args.norm_adv,
optim_critic_iters=args.optim_critic_iters,
max_kl=args.max_kl,
backtrack_coeff=args.backtrack_coeff,
max_backtracks=args.max_backtracks,
)
# load a previous policy
if args.resume_path:
ckpt = torch.load(args.resume_path, map_location=args.device)
policy.load_state_dict(ckpt["model"])
train_envs.set_obs_rms(ckpt["obs_rms"])
test_envs.set_obs_rms(ckpt["obs_rms"])
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
now = datetime.datetime.now().strftime("%y%m%d-%H%M%S")
args.algo_name = "trpo"
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):
state = {
"model": policy.state_dict(),
"obs_rms": train_envs.get_obs_rms()
}
torch.save(state, 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_best_fn=save_best_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 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_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
args.max_action = env.action_space.high[0]
# 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,
action_range=[env.action_space.low[0], env.action_space.high[0]],
tau=args.tau,
gamma=args.gamma,
alpha=args.alpha,
reward_normalization=args.rew_norm,
exploration_noise=OUNoise(0.0, args.noise_std))
# 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 = 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 = 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,
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!
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()}")
writer = SummaryWriter(log_path)
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
######################################### Train #############################
# trainer
result = offline_trainer(
policy, offline_buffer, test_collector, args.epoch,
args.step_per_epoch, args.test_num, args.batch_size,
stop_fn=stop_fn, save_fn=save_fn, writer=writer
)
assert stop_fn(result['best_reward'])
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=[1] * args.test_num, render=args.render)
print(f'Final reward: {result["rew"]}, length: {result["len"]}')
if __name__ == '__main__':
main()
def test_dqn(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
# 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 = 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
Q_param = {"hidden_sizes": args.dueling_q_hidden_sizes}
V_param = {"hidden_sizes": args.dueling_v_hidden_sizes}
net = Net(args.state_shape,
args.action_shape,
hidden_sizes=args.hidden_sizes,
device=args.device,
dueling_param=(Q_param, V_param)).to(args.device)
optim = torch.optim.Adam(net.parameters(), lr=args.lr)
policy = DQNPolicy(net,
optim,
args.gamma,
args.n_step,
target_update_freq=args.target_update_freq)
# 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)
# policy.set_eps(1)
train_collector.collect(n_step=args.batch_size * args.training_num)
# log
log_path = os.path.join(args.logdir, args.task, 'dqn')
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
def train_fn(epoch, env_step): # exp decay
eps = max(args.eps_train * (1 - 5e-6)**env_step, args.eps_test)
policy.set_eps(eps)
def test_fn(epoch, env_step):
policy.set_eps(args.eps_test)
# 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,
stop_fn=stop_fn,
train_fn=train_fn,
test_fn=test_fn,
save_fn=save_fn,
logger=logger)
assert stop_fn(result['best_reward'])
if __name__ == '__main__':
pprint.pprint(result)
# Let's watch its performance!
policy.eval()
policy.set_eps(args.eps_test)
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_td3_bc():
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] # float
print("device:", args.device)
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))
args.state_dim = args.state_shape[0]
args.action_dim = args.action_shape[0]
print("Max_action", args.max_action)
test_envs = SubprocVectorEnv(
[lambda: gym.make(args.task) for _ in range(args.test_num)])
if args.norm_obs:
test_envs = VectorEnvNormObs(test_envs, update_obs_rms=False)
# seed
np.random.seed(args.seed)
torch.manual_seed(args.seed)
test_envs.seed(args.seed)
# model
# actor network
net_a = Net(
args.state_shape,
hidden_sizes=args.hidden_sizes,
device=args.device,
)
actor = Actor(
net_a,
action_shape=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)
# critic network
net_c1 = Net(
args.state_shape,
args.action_shape,
hidden_sizes=args.hidden_sizes,
concat=True,
device=args.device,
)
net_c2 = 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)
critic2 = Critic(net_c2, device=args.device).to(args.device)
critic2_optim = torch.optim.Adam(critic2.parameters(), lr=args.critic_lr)
policy = TD3BCPolicy(
actor,
actor_optim,
critic1,
critic1_optim,
critic2,
critic2_optim,
tau=args.tau,
gamma=args.gamma,
exploration_noise=GaussianNoise(sigma=args.exploration_noise),
policy_noise=args.policy_noise,
update_actor_freq=args.update_actor_freq,
noise_clip=args.noise_clip,
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, map_location=args.device))
print("Loaded agent from: ", args.resume_path)
# collector
test_collector = Collector(policy, test_envs)
# log
now = datetime.datetime.now().strftime("%y%m%d-%H%M%S")
args.algo_name = "td3_bc"
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 watch():
if args.resume_path is None:
args.resume_path = os.path.join(log_path, "policy.pth")
policy.load_state_dict(
torch.load(args.resume_path, map_location=torch.device("cpu")))
policy.eval()
collector = Collector(policy, env)
collector.collect(n_episode=1, render=1 / 35)
if not args.watch:
replay_buffer = load_buffer_d4rl(args.expert_data_task)
if args.norm_obs:
replay_buffer, obs_rms = normalize_all_obs_in_replay_buffer(
replay_buffer)
test_envs.set_obs_rms(obs_rms)
# trainer
result = offline_trainer(
policy,
replay_buffer,
test_collector,
args.epoch,
args.step_per_epoch,
args.test_num,
args.batch_size,
save_best_fn=save_best_fn,
logger=logger,
)
pprint.pprint(result)
else:
watch()
# 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_a2c_with_il(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
# 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
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 = A2CPolicy(actor,
critic,
optim,
dist,
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)
# 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, 'a2c')
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'])
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()
# 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(1, args.state_shape, device=args.device)
net = Actor(net, args.action_shape).to(args.device)
optim = torch.optim.Adam(net.parameters(), lr=args.il_lr)
il_policy = ImitationPolicy(net, optim, mode='discrete')
il_test_collector = Collector(il_policy, test_envs)
train_collector.reset()
result = offpolicy_trainer(il_policy,
train_collector,
il_test_collector,
args.epoch,
args.step_per_epoch,
args.collect_per_step,
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()
il_test_collector.close()
if __name__ == '__main__':
pprint.pprint(result)
# Let's watch its performance!
env = gym.make(args.task)
collector = Collector(il_policy, env)
result = collector.collect(n_episode=1, render=args.render)
print(f'Final reward: {result["rew"]}, length: {result["len"]}')
collector.close()
def test_dqn(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
# 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
Q_param = {"hidden_sizes": args.dueling_q_hidden_sizes}
V_param = {"hidden_sizes": args.dueling_v_hidden_sizes}
net = Net(args.state_shape,
args.action_shape,
hidden_sizes=args.hidden_sizes,
device=args.device,
dueling_param=(Q_param, V_param)).to(args.device)
optim = torch.optim.Adam(net.parameters(), lr=args.lr)
policy = DQNPolicy(net,
optim,
args.gamma,
args.n_step,
target_update_freq=args.target_update_freq)
# collector
train_collector = Collector(policy, train_envs,
ReplayBuffer(args.buffer_size))
test_collector = Collector(policy, test_envs)
# policy.set_eps(1)
train_collector.collect(n_step=args.batch_size)
# log
log_path = os.path.join(args.logdir, args.task, 'dqn')
writer = SummaryWriter(log_path)
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
def train_fn(epoch, env_step):
if env_step <= 100000:
policy.set_eps(args.eps_train)
elif env_step <= 500000:
eps = args.eps_train - (env_step - 100000) / \
400000 * (0.5 * args.eps_train)
policy.set_eps(eps)
else:
policy.set_eps(0.5 * args.eps_train)
def test_fn(epoch, env_step):
policy.set_eps(args.eps_test)
# trainer
result = offpolicy_trainer(policy,
train_collector,
test_collector,
args.epoch,
args.step_per_epoch,
args.collect_per_step,
args.test_num,
args.batch_size,
train_fn=train_fn,
test_fn=test_fn,
stop_fn=stop_fn,
save_fn=save_fn,
writer=writer)
assert stop_fn(result['best_reward'])
if __name__ == '__main__':
pprint.pprint(result)
# Let's watch its performance!
policy.eval()
policy.set_eps(args.eps_test)
test_envs.seed(args.seed)
test_collector.reset()
result = test_collector.collect(n_episode=[1] * args.test_num,
render=args.render)
print(f'Final reward: {result["rew"]}, length: {result["len"]}')
def test_reinforce(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)
torch.nn.init.constant_(actor.sigma_param, -0.5)
for m in actor.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(actor.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 = PGPolicy(actor,
optim,
dist,
discount_factor=args.gamma,
reward_normalization=args.rew_norm,
action_scaling=True,
action_bound_method=args.action_bound_method,
lr_scheduler=lr_scheduler,
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)
# log
t0 = datetime.datetime.now().strftime("%m%d_%H%M%S")
log_file = f'seed_{args.seed}_{t0}-{args.task.replace("-", "_")}_reinforce'
log_path = os.path.join(args.logdir, args.task, 'reinforce', log_file)
writer = SummaryWriter(log_path)
writer.add_text("args", str(args))
logger = BasicLogger(writer, update_interval=10, 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_psrl(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)
if args.reward_threshold is None:
default_reward_threshold = {"NChain-v0": 3400}
args.reward_threshold = default_reward_threshold.get(
args.task, env.spec.reward_threshold)
print("reward threshold:", args.reward_threshold)
args.state_shape = env.observation_space.shape or env.observation_space.n
args.action_shape = env.action_space.shape or env.action_space.n
# seed
np.random.seed(args.seed)
torch.manual_seed(args.seed)
# model
n_action = args.action_shape
n_state = args.state_shape
trans_count_prior = np.ones((n_state, n_action, n_state))
rew_mean_prior = np.full((n_state, n_action), args.rew_mean_prior)
rew_std_prior = np.full((n_state, n_action), args.rew_std_prior)
policy = PSRLPolicy(trans_count_prior, rew_mean_prior, rew_std_prior,
args.gamma, args.eps, args.add_done_loop)
# collector
train_collector = Collector(policy,
train_envs,
VectorReplayBuffer(args.buffer_size,
len(train_envs)),
exploration_noise=True)
test_collector = Collector(policy, test_envs)
# Logger
if args.logger == "wandb":
logger = WandbLogger(save_interval=1,
project='psrl',
name='wandb_test',
config=args)
if args.logger != "none":
log_path = os.path.join(args.logdir, args.task, 'psrl')
writer = SummaryWriter(log_path)
writer.add_text("args", str(args))
if args.logger == "tensorboard":
logger = TensorboardLogger(writer)
else:
logger.load(writer)
else:
logger = LazyLogger()
def stop_fn(mean_rewards):
return mean_rewards >= args.reward_threshold
train_collector.collect(n_step=args.buffer_size, random=True)
# trainer, test it without logger
result = onpolicy_trainer(
policy,
train_collector,
test_collector,
args.epoch,
args.step_per_epoch,
1,
args.test_num,
0,
episode_per_collect=args.episode_per_collect,
stop_fn=stop_fn,
logger=logger,
test_in_train=False,
)
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()}")
elif env.spec.reward_threshold:
assert result["best_reward"] >= env.spec.reward_threshold
