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_ppo(args=get_args()):
env = create_atari_environment(args.task)
args.state_shape = env.observation_space.shape or env.observation_space.n
args.action_shape = env.action_space().shape or env.action_space().n
# train_envs = gym.make(args.task)
train_envs = SubprocVectorEnv([
lambda: create_atari_environment(args.task)
for _ in range(args.training_num)])
# test_envs = gym.make(args.task)
test_envs = SubprocVectorEnv([
lambda: create_atari_environment(args.task)
for _ in range(args.test_num)])
# seed
np.random.seed(args.seed)
torch.manual_seed(args.seed)
train_envs.seed(args.seed)
test_envs.seed(args.seed)
# model
net = Net(args.layer_num, args.state_shape, device=args.device)
actor = Actor(net, args.action_shape).to(args.device)
critic = Critic(net).to(args.device)
optim = torch.optim.Adam(list(
actor.parameters()) + list(critic.parameters()), lr=args.lr)
dist = torch.distributions.Categorical
policy = PPOPolicy(
actor, critic, optim, dist, args.gamma,
max_grad_norm=args.max_grad_norm,
eps_clip=args.eps_clip,
vf_coef=args.vf_coef,
ent_coef=args.ent_coef,
action_range=None)
# collector
train_collector = Collector(
policy, train_envs, ReplayBuffer(args.buffer_size),
preprocess_fn=preprocess_fn)
test_collector = Collector(policy, test_envs, preprocess_fn=preprocess_fn)
# log
writer = SummaryWriter(args.logdir + '/' + 'ppo')
def stop_fn(x):
if env.env.spec.reward_threshold:
return x >= env.spec.reward_threshold
else:
return False
# trainer
result = onpolicy_trainer(
policy, train_collector, test_collector, args.epoch,
args.step_per_epoch, args.collect_per_step, args.repeat_per_collect,
args.test_num, args.batch_size, stop_fn=stop_fn, writer=writer)
train_collector.close()
test_collector.close()
if __name__ == '__main__':
pprint.pprint(result)
# Let's watch its performance!
env = create_atari_environment(args.task)
collector = Collector(policy, env, preprocess_fn=preprocess_fn)
result = collector.collect(n_step=2000, render=args.render)
print(f'Final reward: {result["rew"]}, length: {result["len"]}')
collector.close()
def test_a2c(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)
# 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'])
train_collector.close()
test_collector.close()
if __name__ == '__main__':
pprint.pprint(result)
# Let's watch its performance!
env = gym.make(args.task)
collector = Collector(policy, env)
result = collector.collect(n_episode=1, render=args.render)
print(f'Final reward: {result["rew"]}, length: {result["len"]}')
collector.close()
def test_pg(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 = 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,
args.action_shape,
device=args.device,
softmax=True)
net = net.to(args.device)
optim = torch.optim.Adam(net.parameters(), lr=args.lr)
dist = torch.distributions.Categorical
policy = PGPolicy(net, optim, dist, args.gamma)
# collector
train_collector = Collector(policy, train_envs,
ReplayBuffer(args.buffer_size))
test_collector = Collector(policy, test_envs)
# log
writer = SummaryWriter(args.logdir + '/' + 'pg')
def stop_fn(x):
return x >= env.spec.reward_threshold
# trainer
result = onpolicy_trainer(policy,
train_collector,
test_collector,
args.epoch,
args.step_per_epoch,
args.collect_per_step,
args.repeat_per_collect,
args.test_num,
args.batch_size,
stop_fn=stop_fn,
writer=writer)
assert stop_fn(result['best_reward'])
train_collector.close()
test_collector.close()
if __name__ == '__main__':
pprint.pprint(result)
# Let's watch its performance!
env = gym.make(args.task)
collector = Collector(policy, env)
result = collector.collect(n_episode=1, render=args.render)
print(f'Final reward: {result["rew"]}, length: {result["len"]}')
collector.close()
def test_pg(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
net = Net(args.state_shape, args.action_shape,
hidden_sizes=args.hidden_sizes,
device=args.device, softmax=True).to(args.device)
optim = torch.optim.Adam(net.parameters(), lr=args.lr)
dist = torch.distributions.Categorical
policy = PGPolicy(net, optim, dist, args.gamma,
reward_normalization=args.rew_norm,
action_space=env.action_space)
# collector
train_collector = Collector(
policy, train_envs,
VectorReplayBuffer(args.buffer_size, len(train_envs)),
exploration_noise=True)
test_collector = Collector(policy, test_envs)
# log
log_path = os.path.join(args.logdir, args.task, 'pg')
writer = SummaryWriter(log_path)
logger = BasicLogger(writer)
def save_fn(policy):
torch.save(policy.state_dict(), os.path.join(log_path, 'policy.pth'))
def stop_fn(mean_rewards):
return mean_rewards >= env.spec.reward_threshold
# trainer
result = onpolicy_trainer(
policy, train_collector, test_collector, args.epoch,
args.step_per_epoch, args.repeat_per_collect, args.test_num, args.batch_size,
episode_per_collect=args.episode_per_collect, stop_fn=stop_fn, save_fn=save_fn,
logger=logger)
assert stop_fn(result['best_reward'])
if __name__ == '__main__':
pprint.pprint(result)
# Let's watch its performance!
env = gym.make(args.task)
policy.eval()
collector = Collector(policy, env)
result = collector.collect(n_episode=1, render=args.render)
rews, lens = result["rews"], result["lens"]
print(f"Final reward: {rews.mean()}, length: {lens.mean()}")
def train(hyper: dict):
env_id = 'CartPole-v1'
env = gym.make(env_id)
hyper['state_dim'] = 4
hyper['action_dim'] = 2
train_envs = VectorEnv([lambda: gym.make(env_id) for _ in range(hyper['training_num'])])
test_envs = SubprocVectorEnv([lambda: gym.make(env_id) for _ in range(hyper['test_num'])])
if hyper['seed']:
np.random.seed(hyper['random_seed'])
torch.manual_seed(hyper['random_seed'])
train_envs.seed(hyper['random_seed'])
test_envs.seed(hyper['random_seed'])
device = Pytorch.device()
net = Net(hyper['layer_num'], hyper['state_dim'], device=device)
actor = Actor(net, hyper['action_dim']).to(device)
critic = Critic(net).to(device)
optim = torch.optim.Adam(list(
actor.parameters()) + list(critic.parameters()), lr=hyper['learning_rate'])
dist = torch.distributions.Categorical
policy = A2CPolicy(
actor, critic, optim, dist, hyper['gamma'], vf_coef=hyper['vf_coef'],
ent_coef=hyper['ent_coef'], max_grad_norm=hyper['max_grad_norm'])
# collector
train_collector = Collector(
policy, train_envs, ReplayBuffer(hyper['capacity']))
test_collector = Collector(policy, test_envs)
writer = SummaryWriter('./a2c')
def stop_fn(x):
if env.env.spec.reward_threshold:
return x >= env.spec.reward_threshold
else:
return False
result = onpolicy_trainer(
policy, train_collector, test_collector, hyper['epoch'],
hyper['step_per_epoch'], hyper['collect_per_step'], hyper['repeat_per_collect'],
hyper['test_num'], hyper['batch_size'], stop_fn=stop_fn, writer=writer,
task=env_id)
train_collector.close()
test_collector.close()
pprint.pprint(result)
# 测试
env = gym.make(env_id)
collector = Collector(policy, env)
result = collector.collect(n_episode=1, render=hyper['render'])
print(f'Final reward: {result["rew"]}, length: {result["len"]}')
collector.close()
def test_ppo(args=get_args()):
torch.set_num_threads(1) # we just need only one thread for NN
env = gym_make()
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.action_range = [env.action_shape.low[0], env.action_space.high[0]]
# you can also use tianshou.env.SubprocVectorEnv
# train_envs = gym_make()
train_envs = VectorEnv(
[lambda: gym_make() for _ in range(args.training_num)])
# test_envs = gym_make()
test_envs = VectorEnv([lambda: gym_make() 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
policy = init_policy(args, env)
# collector
train_collector = Collector(policy, train_envs,
ReplayBuffer(args.buffer_size))
test_collector = Collector(policy, test_envs)
# log
log_path = os.path.join(args.logdir, args.task, 'ppo')
writer = SummaryWriter(log_path)
def save_fn(policy):
torch.save(policy.state_dict(), os.path.join(log_path, 'policy.pth'))
# 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,
save_fn=save_fn,
writer=writer)
train_collector.close()
test_collector.close()
if __name__ == '__main__':
pprint.pprint(result)
# Let's watch its performance!
env = gym_make()
collector = Collector(policy, env)
result = collector.collect(n_episode=1, render=args.render)
print(f'Final reward: {result["rew"]}, length: {result["len"]}')
collector.close()
def test_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 train_agent(
args: argparse.Namespace = get_args(),
agents: Optional[List[BasePolicy]] = None,
optims: Optional[List[torch.optim.Optimizer]] = None,
) -> Tuple[dict, BasePolicy]:
train_envs = DummyVectorEnv([get_env for _ in range(args.training_num)])
test_envs = DummyVectorEnv([get_env for _ in range(args.test_num)])
# seed
np.random.seed(args.seed)
torch.manual_seed(args.seed)
train_envs.seed(args.seed)
test_envs.seed(args.seed)
policy, optim, agents = get_agents(args, agents=agents, optims=optims)
# collector
train_collector = Collector(
policy,
train_envs,
VectorReplayBuffer(args.buffer_size, len(train_envs)),
exploration_noise=False # True
)
test_collector = Collector(policy, test_envs)
# train_collector.collect(n_step=args.batch_size * args.training_num)
# log
log_path = os.path.join(args.logdir, 'pistonball', 'dqn')
writer = SummaryWriter(log_path)
writer.add_text("args", str(args))
logger = TensorboardLogger(writer)
def save_best_fn(policy):
pass
def stop_fn(mean_rewards):
return False
def train_fn(epoch, env_step):
[agent.set_eps(args.eps_train) for agent in policy.policies.values()]
def test_fn(epoch, env_step):
[agent.set_eps(args.eps_test) for agent in policy.policies.values()]
def reward_metric(rews):
return rews[:, 0]
# trainer
result = 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,
resume_from_log=args.resume)
return result, policy
def test_acktr(args=get_args()):
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,
activation=nn.Tanh,
device=args.device)
actor = ActorProb(net,
args.action_shape,
max_action=args.max_action,
unbounded=True,
device=args.device).to(args.device)
critic = Critic(Net(args.state_shape,
hidden_sizes=args.hidden_sizes,
device=args.device,
activation=nn.Tanh),
device=args.device).to(args.device)
torch.nn.init.constant_(actor.sigma_param._bias, -0.5)
# orthogonal initialization
for m in list(actor.modules()) + list(critic.modules()):
if isinstance(m, torch.nn.Linear):
torch.nn.init.orthogonal_(m.weight)
torch.nn.init.zeros_(m.bias)
optim = KFACOptimizer(actor, critic, lr=0.25)
# replace DiagGuassian with Independent(Normal) which is equivalent
# pass *logits to be consistent with policy.forward
def dist(*logits):
return Independent(Normal(*logits), 1)
policy = ACKTRPolicy(actor,
critic,
optim,
dist,
discount_factor=args.gamma,
reward_normalization=args.rew_norm,
advantage_normalization=args.norm_adv,
gae_lambda=args.gae_lambda,
action_space=env.action_space)
# collector
train_collector = Collector(
policy, train_envs,
VectorReplayBuffer(args.buffer_size, len(train_envs)))
test_collector = Collector(policy, test_envs)
# log
log_path = os.path.join(args.logdir, args.task, 'acktr')
writer = SummaryWriter(log_path)
logger = BasicLogger(writer)
def save_fn(policy):
torch.save(policy.state_dict(), os.path.join(log_path, 'policy.pth'))
def stop_fn(mean_rewards):
return mean_rewards >= env.spec.reward_threshold
# trainer
result = onpolicy_trainer(policy,
train_collector,
test_collector,
args.epoch,
args.step_per_epoch,
args.repeat_per_collect,
args.test_num,
args.batch_size,
step_per_collect=args.step_per_collect,
stop_fn=stop_fn,
save_fn=save_fn,
logger=logger)
assert stop_fn(result['best_reward'])
if __name__ == '__main__':
pprint.pprint(result)
# Let's watch its performance!
env = gym.make(args.task)
policy.eval()
collector = Collector(policy, env)
result = collector.collect(n_episode=1, render=args.render)
rews, lens = result["rews"], result["lens"]
print(f"Final reward: {rews.mean()}, length: {lens.mean()}")
def run_pg(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 = VectorEnv(
[lambda: gym.make(args.task) for _ in range(args.training_num)])
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,
args.action_shape,
device=args.device,
softmax=True)
net = net.to(args.device)
optim = torch.optim.Adam(net.parameters(), lr=args.lr)
dist = torch.distributions.Categorical
policy = PGPolicy(net,
optim,
dist,
args.gamma,
reward_normalization=args.rew_norm)
# collector
train_collector = Collector(policy, train_envs,
ReplayBuffer(args.buffer_size))
test_collector = Collector(policy, test_envs)
# log
if not os.path.isdir(os.path.join(args.logdir)):
os.mkdir(os.path.join(args.logdir))
if not os.path.isdir(os.path.join(args.logdir, args.task)):
os.mkdir(os.path.join(args.logdir, args.task))
if not os.path.isdir(os.path.join(args.logdir, args.task, 'pg')):
os.mkdir(os.path.join(args.logdir, args.task, 'pg'))
log_path = os.path.join(args.logdir, args.task, 'pg')
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
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)
train_collector.close()
test_collector.close()
if __name__ == '__main__':
# Let's watch its performance!
env = gym.make(args.task)
collector = Collector(policy, env)
result = collector.collect(n_episode=1, render=args.render)
print(f'Final reward: {result["rew"]}, length: {result["len"]}')
collector.close()
def test_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 = 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)
optim = torch.optim.Adam(set(actor.parameters()).union(
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,
action_space=env.action_space)
# collector
train_collector = Collector(policy,
train_envs,
VectorReplayBuffer(args.buffer_size,
len(train_envs)),
exploration_noise=True)
test_collector = Collector(policy, test_envs)
# log
log_path = os.path.join(args.logdir, args.task, 'a2c')
writer = SummaryWriter(log_path)
logger = BasicLogger(writer)
def save_fn(policy):
torch.save(policy.state_dict(), os.path.join(log_path, 'policy.pth'))
def stop_fn(mean_rewards):
return mean_rewards >= env.spec.reward_threshold
# trainer
result = onpolicy_trainer(policy,
train_collector,
test_collector,
args.epoch,
args.step_per_epoch,
args.repeat_per_collect,
args.test_num,
args.batch_size,
episode_per_collect=args.episode_per_collect,
stop_fn=stop_fn,
save_fn=save_fn,
logger=logger)
assert stop_fn(result['best_reward'])
if __name__ == '__main__':
pprint.pprint(result)
# Let's watch its performance!
env = gym.make(args.task)
policy.eval()
collector = Collector(policy, env)
result = collector.collect(n_episode=1, render=args.render)
rews, lens = result["rews"], result["lens"]
print(f"Final reward: {rews.mean()}, length: {lens.mean()}")
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, mode='discrete')
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_gail(args=get_args()):
if os.path.exists(args.load_buffer_name) and os.path.isfile(
args.load_buffer_name):
if args.load_buffer_name.endswith(".hdf5"):
buffer = VectorReplayBuffer.load_hdf5(args.load_buffer_name)
else:
buffer = pickle.load(open(args.load_buffer_name, "rb"))
else:
buffer = gather_data()
env = gym.make(args.task)
if args.reward_threshold is None:
default_reward_threshold = {"Pendulum-v0": -1100, "Pendulum-v1": -1100}
args.reward_threshold = default_reward_threshold.get(
args.task, 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]
# you can also use tianshou.env.SubprocVectorEnv
# train_envs = gym.make(args.task)
train_envs = DummyVectorEnv(
[lambda: gym.make(args.task) for _ in range(args.training_num)])
# test_envs = gym.make(args.task)
test_envs = DummyVectorEnv(
[lambda: gym.make(args.task) for _ in range(args.test_num)])
# seed
np.random.seed(args.seed)
torch.manual_seed(args.seed)
train_envs.seed(args.seed)
test_envs.seed(args.seed)
# model
net = Net(args.state_shape,
hidden_sizes=args.hidden_sizes,
device=args.device)
actor = ActorProb(net,
args.action_shape,
max_action=args.max_action,
device=args.device).to(args.device)
critic = Critic(Net(args.state_shape,
hidden_sizes=args.hidden_sizes,
device=args.device),
device=args.device).to(args.device)
actor_critic = ActorCritic(actor, critic)
# orthogonal initialization
for m in actor_critic.modules():
if isinstance(m, torch.nn.Linear):
torch.nn.init.orthogonal_(m.weight)
torch.nn.init.zeros_(m.bias)
optim = torch.optim.Adam(actor_critic.parameters(), lr=args.lr)
# discriminator
disc_net = Critic(Net(
args.state_shape,
action_shape=args.action_shape,
hidden_sizes=args.hidden_sizes,
activation=torch.nn.Tanh,
device=args.device,
concat=True,
),
device=args.device).to(args.device)
for m in disc_net.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)
disc_optim = torch.optim.Adam(disc_net.parameters(), lr=args.disc_lr)
# replace DiagGuassian with Independent(Normal) which is equivalent
# pass *logits to be consistent with policy.forward
def dist(*logits):
return Independent(Normal(*logits), 1)
policy = GAILPolicy(
actor,
critic,
optim,
dist,
buffer,
disc_net,
disc_optim,
disc_update_num=args.disc_update_num,
discount_factor=args.gamma,
max_grad_norm=args.max_grad_norm,
eps_clip=args.eps_clip,
vf_coef=args.vf_coef,
ent_coef=args.ent_coef,
reward_normalization=args.rew_norm,
advantage_normalization=args.norm_adv,
recompute_advantage=args.recompute_adv,
dual_clip=args.dual_clip,
value_clip=args.value_clip,
gae_lambda=args.gae_lambda,
action_space=env.action_space,
)
# collector
train_collector = Collector(
policy, train_envs,
VectorReplayBuffer(args.buffer_size, len(train_envs)))
test_collector = Collector(policy, test_envs)
# log
log_path = os.path.join(args.logdir, args.task, "gail")
writer = SummaryWriter(log_path)
logger = TensorboardLogger(writer, save_interval=args.save_interval)
def save_best_fn(policy):
torch.save(policy.state_dict(), os.path.join(log_path, "policy.pth"))
def stop_fn(mean_rewards):
return mean_rewards >= args.reward_threshold
def save_checkpoint_fn(epoch, env_step, gradient_step):
# see also: https://pytorch.org/tutorials/beginner/saving_loading_models.html
ckpt_path = os.path.join(log_path, "checkpoint.pth")
# Example: saving by epoch num
# ckpt_path = os.path.join(log_path, f"checkpoint_{epoch}.pth")
torch.save(
{
"model": policy.state_dict(),
"optim": optim.state_dict(),
}, ckpt_path)
return ckpt_path
if args.resume:
# load from existing checkpoint
print(f"Loading agent under {log_path}")
ckpt_path = os.path.join(log_path, "checkpoint.pth")
if os.path.exists(ckpt_path):
checkpoint = torch.load(ckpt_path, map_location=args.device)
policy.load_state_dict(checkpoint["model"])
optim.load_state_dict(checkpoint["optim"])
print("Successfully restore policy and optim.")
else:
print("Fail to restore policy and optim.")
# trainer
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,
resume_from_log=args.resume,
save_checkpoint_fn=save_checkpoint_fn,
)
assert stop_fn(result["best_reward"])
if __name__ == "__main__":
pprint.pprint(result)
# Let's watch its performance!
env = gym.make(args.task)
policy.eval()
collector = Collector(policy, env)
result = collector.collect(n_episode=1, render=args.render)
rews, lens = result["rews"], result["lens"]
print(f"Final reward: {rews.mean()}, length: {lens.mean()}")
def test_ppo(args=get_args()):
env = gym.make(args.task)
args.state_shape = env.observation_space.shape or env.observation_space.n
args.action_shape = env.action_space.shape or env.action_space.n
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_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
if load:
new_lr = 1e-3
load_dict = torch.load('saved/snake_ppo2_last.pth')
policy.load_state_dict(load_dict['policy_state'])
policy.optim.load_state_dict(load_dict['optimizer_state'])
policy.scaler.load_state_dict(load_dict['scaler'])
policy.optim.param_groups[0]['lr'] = new_lr
print(f'Learning rate is set to = {new_lr}')
result = onpolicy_trainer(
policy,
train_collector,
test_collector,
max_epoch=epoch,
step_per_epoch=1600000,
repeat_per_collect=1,
episode_per_test=test_num,
batch_size=1024,
step_per_collect=train_num * 32, # 5
# episode_per_collect=10,
save_fn=save_fn,
backup_save_freq=0,
)
# Fun part see the results!
eval_env = make_env_wrapper()
eval_collector = ts.data.Collector(policy, eval_env)
policy.eval()
eval_collector.collect(n_episode=10, render=0.005)
def test_ppo(args=get_args()):
torch.set_num_threads(1) # we just need only one thread for NN
env = gym.make(args.task)
if args.task == 'Pendulum-v0':
env.spec.reward_threshold = -250
args.state_shape = env.observation_space.shape or env.observation_space.n
args.action_shape = env.action_space.shape or env.action_space.n
args.max_action = env.action_space.high[0]
# you can also use tianshou.env.SubprocVectorEnv
# train_envs = gym.make(args.task)
train_envs = DummyVectorEnv(
[lambda: gym.make(args.task) for _ in range(args.training_num)])
# test_envs = gym.make(args.task)
test_envs = DummyVectorEnv(
[lambda: gym.make(args.task) for _ in range(args.test_num)])
# seed
np.random.seed(args.seed)
torch.manual_seed(args.seed)
train_envs.seed(args.seed)
test_envs.seed(args.seed)
# model
net = Net(args.state_shape, hidden_sizes=args.hidden_sizes,
device=args.device)
actor = ActorProb(net, args.action_shape, max_action=args.max_action,
device=args.device).to(args.device)
critic = Critic(Net(
args.state_shape, hidden_sizes=args.hidden_sizes, device=args.device
), device=args.device).to(args.device)
# orthogonal initialization
for m in list(actor.modules()) + list(critic.modules()):
if isinstance(m, torch.nn.Linear):
torch.nn.init.orthogonal_(m.weight)
torch.nn.init.zeros_(m.bias)
optim = torch.optim.Adam(set(
actor.parameters()).union(critic.parameters()), lr=args.lr)
# replace DiagGuassian with Independent(Normal) which is equivalent
# pass *logits to be consistent with policy.forward
def dist(*logits):
return Independent(Normal(*logits), 1)
policy = PPOPolicy(
actor, critic, optim, dist,
discount_factor=args.gamma,
max_grad_norm=args.max_grad_norm,
eps_clip=args.eps_clip,
vf_coef=args.vf_coef,
ent_coef=args.ent_coef,
reward_normalization=args.rew_norm,
advantage_normalization=args.norm_adv,
recompute_advantage=args.recompute_adv,
# dual_clip=args.dual_clip,
# dual clip cause monotonically increasing log_std :)
value_clip=args.value_clip,
gae_lambda=args.gae_lambda,
action_space=env.action_space)
# collector
train_collector = Collector(
policy, train_envs,
VectorReplayBuffer(args.buffer_size, len(train_envs)),
exploration_noise=True)
test_collector = Collector(policy, test_envs)
# log
log_path = os.path.join(args.logdir, args.task, 'ppo')
writer = SummaryWriter(log_path)
logger = BasicLogger(writer)
def save_fn(policy):
torch.save(policy.state_dict(), os.path.join(log_path, 'policy.pth'))
def stop_fn(mean_rewards):
return mean_rewards >= env.spec.reward_threshold
# trainer
result = onpolicy_trainer(
policy, train_collector, test_collector, args.epoch,
args.step_per_epoch, args.repeat_per_collect, args.test_num, args.batch_size,
episode_per_collect=args.episode_per_collect, stop_fn=stop_fn, save_fn=save_fn,
logger=logger)
assert stop_fn(result['best_reward'])
if __name__ == '__main__':
pprint.pprint(result)
# Let's watch its performance!
env = gym.make(args.task)
policy.eval()
collector = Collector(policy, env)
result = collector.collect(n_episode=1, render=args.render)
rews, lens = result["rews"], result["lens"]
print(f"Final reward: {rews.mean()}, length: {lens.mean()}")
def test_ppo(args=get_args()):
torch.set_num_threads(1) # for poor CPU
env = gym.make(args.task)
args.state_shape = env.observation_space.shape or env.observation_space.n
args.action_shape = env.action_space.shape or env.action_space.n
# train_envs = gym.make(args.task)
# you can also use tianshou.env.SubprocVectorEnv
train_envs = 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)
# orthogonal initialization
for m in list(actor.modules()) + list(critic.modules()):
if isinstance(m, torch.nn.Linear):
torch.nn.init.orthogonal_(m.weight)
torch.nn.init.zeros_(m.bias)
optim = torch.optim.Adam(set(actor.parameters()).union(
critic.parameters()),
lr=args.lr)
dist = torch.distributions.Categorical
policy = PPOPolicy(actor,
critic,
optim,
dist,
args.gamma,
max_grad_norm=args.max_grad_norm,
eps_clip=args.eps_clip,
vf_coef=args.vf_coef,
ent_coef=args.ent_coef,
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)
# collector
train_collector = Collector(policy,
train_envs,
VectorReplayBuffer(args.buffer_size,
len(train_envs)),
exploration_noise=True)
test_collector = Collector(policy, test_envs)
# log
log_path = os.path.join(args.logdir, args.task, 'ppo')
writer = SummaryWriter(log_path)
logger = BasicLogger(writer)
def save_fn(policy):
torch.save(policy.state_dict(), os.path.join(log_path, 'policy.pth'))
def stop_fn(mean_rewards):
return mean_rewards >= env.spec.reward_threshold
# trainer
result = onpolicy_trainer(policy,
train_collector,
test_collector,
args.epoch,
args.step_per_epoch,
args.repeat_per_collect,
args.test_num,
args.batch_size,
episode_per_collect=args.episode_per_collect,
stop_fn=stop_fn,
save_fn=save_fn,
logger=logger)
assert stop_fn(result['best_reward'])
if __name__ == '__main__':
pprint.pprint(result)
# Let's watch its performance!
env = gym.make(args.task)
policy.eval()
collector = Collector(policy, env)
result = collector.collect(n_episode=1, render=args.render)
rews, lens = result["rews"], result["lens"]
print(f"Final reward: {rews.mean()}, length: {lens.mean()}")
def test_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_ppo(args=get_args()):
env = gym.make(args.task)
if args.task == 'Pendulum-v0':
env.spec.reward_threshold = -250
args.state_shape = env.observation_space.shape or env.observation_space.n
args.action_shape = env.action_space.shape or env.action_space.n
args.max_action = env.action_space.high[0]
# you can also use tianshou.env.SubprocVectorEnv
# train_envs = gym.make(args.task)
train_envs = DummyVectorEnv(
[lambda: gym.make(args.task) for _ in range(args.training_num)])
# test_envs = gym.make(args.task)
test_envs = DummyVectorEnv(
[lambda: gym.make(args.task) for _ in range(args.test_num)])
# seed
np.random.seed(args.seed)
torch.manual_seed(args.seed)
train_envs.seed(args.seed)
test_envs.seed(args.seed)
# model
net = Net(args.state_shape,
hidden_sizes=args.hidden_sizes,
device=args.device)
actor = ActorProb(net,
args.action_shape,
max_action=args.max_action,
device=args.device).to(args.device)
critic = Critic(Net(args.state_shape,
hidden_sizes=args.hidden_sizes,
device=args.device),
device=args.device).to(args.device)
# orthogonal initialization
for m in set(actor.modules()).union(critic.modules()):
if isinstance(m, torch.nn.Linear):
torch.nn.init.orthogonal_(m.weight)
torch.nn.init.zeros_(m.bias)
optim = torch.optim.Adam(set(actor.parameters()).union(
critic.parameters()),
lr=args.lr)
# replace DiagGuassian with Independent(Normal) which is equivalent
# pass *logits to be consistent with policy.forward
def dist(*logits):
return Independent(Normal(*logits), 1)
policy = PPOPolicy(
actor,
critic,
optim,
dist,
discount_factor=args.gamma,
max_grad_norm=args.max_grad_norm,
eps_clip=args.eps_clip,
vf_coef=args.vf_coef,
ent_coef=args.ent_coef,
reward_normalization=args.rew_norm,
advantage_normalization=args.norm_adv,
recompute_advantage=args.recompute_adv,
# dual_clip=args.dual_clip,
# dual clip cause monotonically increasing log_std :)
value_clip=args.value_clip,
gae_lambda=args.gae_lambda,
action_space=env.action_space)
# collector
train_collector = Collector(
policy, train_envs,
VectorReplayBuffer(args.buffer_size, len(train_envs)))
test_collector = Collector(policy, test_envs)
# log
log_path = os.path.join(args.logdir, args.task, 'ppo')
writer = SummaryWriter(log_path)
logger = TensorboardLogger(writer, save_interval=args.save_interval)
def save_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 save_checkpoint_fn(epoch, env_step, gradient_step):
# see also: https://pytorch.org/tutorials/beginner/saving_loading_models.html
torch.save(
{
'model': policy.state_dict(),
'optim': optim.state_dict(),
}, os.path.join(log_path, 'checkpoint.pth'))
if args.resume:
# load from existing checkpoint
print(f"Loading agent under {log_path}")
ckpt_path = os.path.join(log_path, 'checkpoint.pth')
if os.path.exists(ckpt_path):
checkpoint = torch.load(ckpt_path, map_location=args.device)
policy.load_state_dict(checkpoint['model'])
optim.load_state_dict(checkpoint['optim'])
print("Successfully restore policy and optim.")
else:
print("Fail to restore policy and optim.")
# trainer
result = onpolicy_trainer(policy,
train_collector,
test_collector,
args.epoch,
args.step_per_epoch,
args.repeat_per_collect,
args.test_num,
args.batch_size,
episode_per_collect=args.episode_per_collect,
stop_fn=stop_fn,
save_fn=save_fn,
logger=logger,
resume_from_log=args.resume,
save_checkpoint_fn=save_checkpoint_fn)
assert stop_fn(result['best_reward'])
if __name__ == '__main__':
pprint.pprint(result)
# Let's watch its performance!
env = gym.make(args.task)
policy.eval()
collector = Collector(policy, env)
result = collector.collect(n_episode=1, render=args.render)
rews, lens = result["rews"], result["lens"]
print(f"Final reward: {rews.mean()}, length: {lens.mean()}")
def _test_ppo(args=get_args()):
# just a demo, I have not made it work :(
env = gym.make(args.task)
if args.task == 'Pendulum-v0':
env.spec.reward_threshold = -250
args.state_shape = env.observation_space.shape or env.observation_space.n
args.action_shape = env.action_space.shape or env.action_space.n
args.max_action = env.action_space.high[0]
# train_envs = gym.make(args.task)
train_envs = SubprocVectorEnv(
[lambda: gym.make(args.task) for _ in range(args.training_num)])
# test_envs = gym.make(args.task)
test_envs = SubprocVectorEnv(
[lambda: gym.make(args.task) for _ in range(args.test_num)])
# seed
np.random.seed(args.seed)
torch.manual_seed(args.seed)
train_envs.seed(args.seed)
test_envs.seed(args.seed)
# model
actor = ActorProb(args.layer_num, args.state_shape, args.action_shape,
args.max_action, args.device).to(args.device)
critic = Critic(args.layer_num, args.state_shape,
device=args.device).to(args.device)
optim = torch.optim.Adam(list(actor.parameters()) +
list(critic.parameters()),
lr=args.lr)
dist = torch.distributions.Normal
policy = PPOPolicy(
actor,
critic,
optim,
dist,
args.gamma,
max_grad_norm=args.max_grad_norm,
eps_clip=args.eps_clip,
vf_coef=args.vf_coef,
ent_coef=args.ent_coef,
action_range=[env.action_space.low[0], env.action_space.high[0]])
# collector
train_collector = Collector(policy, train_envs,
ReplayBuffer(args.buffer_size))
test_collector = Collector(policy, test_envs)
train_collector.collect(n_step=args.step_per_epoch)
# log
writer = SummaryWriter(args.logdir + '/' + 'ppo')
def stop_fn(x):
return x >= env.spec.reward_threshold
# trainer
result = onpolicy_trainer(policy,
train_collector,
test_collector,
args.epoch,
args.step_per_epoch,
args.collect_per_step,
args.repeat_per_collect,
args.test_num,
args.batch_size,
stop_fn=stop_fn,
writer=writer,
task=args.task)
assert stop_fn(result['best_reward'])
train_collector.close()
test_collector.close()
if __name__ == '__main__':
pprint.pprint(result)
# Let's watch its performance!
env = gym.make(args.task)
collector = Collector(policy, env)
result = collector.collect(n_episode=1, render=args.render)
print(f'Final reward: {result["rew"]}, length: {result["len"]}')
collector.close()
def test_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 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 test_ppo(args=get_args()):
torch.set_num_threads(1) # we just need only one thread for NN
env = gym.make(args.task)
if args.task == 'Pendulum-v0':
env.spec.reward_threshold = -250
args.state_shape = env.observation_space.shape or env.observation_space.n
args.action_shape = env.action_space.shape or env.action_space.n
args.max_action = env.action_space.high[0]
# you can also use tianshou.env.SubprocVectorEnv
# train_envs = gym.make(args.task)
train_envs = VectorEnv(
[lambda: gym.make(args.task) for _ in range(args.training_num)])
# test_envs = gym.make(args.task)
test_envs = VectorEnv(
[lambda: gym.make(args.task) for _ in range(args.test_num)])
# seed
np.random.seed(args.seed)
torch.manual_seed(args.seed)
train_envs.seed(args.seed)
test_envs.seed(args.seed)
# model
actor = ActorProb(args.layer_num, args.state_shape, args.action_shape,
args.max_action, args.device).to(args.device)
critic = Critic(args.layer_num, args.state_shape,
device=args.device).to(args.device)
optim = torch.optim.Adam(list(actor.parameters()) +
list(critic.parameters()),
lr=args.lr)
dist = torch.distributions.Normal
policy = PPOPolicy(
actor,
critic,
optim,
dist,
args.gamma,
max_grad_norm=args.max_grad_norm,
eps_clip=args.eps_clip,
vf_coef=args.vf_coef,
ent_coef=args.ent_coef,
reward_normalization=args.rew_norm,
dual_clip=args.dual_clip,
value_clip=args.value_clip,
# action_range=[env.action_space.low[0], env.action_space.high[0]],)
# if clip the action, ppo would not converge :)
gae_lambda=args.gae_lambda)
# collector
train_collector = Collector(policy, train_envs,
ReplayBuffer(args.buffer_size))
test_collector = Collector(policy, test_envs)
# log
log_path = os.path.join(args.logdir, args.task, 'ppo')
writer = SummaryWriter(log_path)
def save_fn(policy):
torch.save(policy.state_dict(), os.path.join(log_path, 'policy.pth'))
def stop_fn(x):
return x >= env.spec.reward_threshold
# trainer
result = onpolicy_trainer(policy,
train_collector,
test_collector,
args.epoch,
args.step_per_epoch,
args.collect_per_step,
args.repeat_per_collect,
args.test_num,
args.batch_size,
stop_fn=stop_fn,
save_fn=save_fn,
writer=writer)
assert stop_fn(result['best_reward'])
train_collector.close()
test_collector.close()
if __name__ == '__main__':
pprint.pprint(result)
# Let's watch its performance!
env = gym.make(args.task)
collector = Collector(policy, env)
result = collector.collect(n_episode=1, render=args.render)
print(f'Final reward: {result["rew"]}, length: {result["len"]}')
collector.close()
def test_ppo(args=get_args()):
args.cfg_path = f"maps/{args.task}.cfg"
args.wad_path = f"maps/{args.task}.wad"
args.res = (args.skip_num, 84, 84)
env = Env(args.cfg_path, args.frames_stack, args.res)
args.state_shape = args.res
args.action_shape = env.action_space.shape or env.action_space.n
# should be N_FRAMES x H x W
print("Observations shape:", args.state_shape)
print("Actions shape:", args.action_shape)
# make environments
train_envs = ShmemVectorEnv([
lambda: Env(args.cfg_path, args.frames_stack, args.res)
for _ in range(args.training_num)
])
test_envs = ShmemVectorEnv([
lambda: Env(args.cfg_path, args.frames_stack, args.res, args.save_lmp)
for _ in range(min(os.cpu_count() - 1, args.test_num))
])
# seed
np.random.seed(args.seed)
torch.manual_seed(args.seed)
train_envs.seed(args.seed)
test_envs.seed(args.seed)
# define model
net = DQN(*args.state_shape,
args.action_shape,
device=args.device,
features_only=True,
output_dim=args.hidden_size)
actor = Actor(net,
args.action_shape,
device=args.device,
softmax_output=False)
critic = Critic(net, device=args.device)
optim = torch.optim.Adam(ActorCritic(actor, critic).parameters(),
lr=args.lr)
lr_scheduler = None
if args.lr_decay:
# decay learning rate to 0 linearly
max_update_num = np.ceil(
args.step_per_epoch / args.step_per_collect) * args.epoch
lr_scheduler = LambdaLR(
optim, lr_lambda=lambda epoch: 1 - epoch / max_update_num)
# define policy
def dist(p):
return torch.distributions.Categorical(logits=p)
policy = PPOPolicy(actor,
critic,
optim,
dist,
discount_factor=args.gamma,
gae_lambda=args.gae_lambda,
max_grad_norm=args.max_grad_norm,
vf_coef=args.vf_coef,
ent_coef=args.ent_coef,
reward_normalization=args.rew_norm,
action_scaling=False,
lr_scheduler=lr_scheduler,
action_space=env.action_space,
eps_clip=args.eps_clip,
value_clip=args.value_clip,
dual_clip=args.dual_clip,
advantage_normalization=args.norm_adv,
recompute_advantage=args.recompute_adv).to(args.device)
if args.icm_lr_scale > 0:
feature_net = DQN(*args.state_shape,
args.action_shape,
device=args.device,
features_only=True,
output_dim=args.hidden_size)
action_dim = np.prod(args.action_shape)
feature_dim = feature_net.output_dim
icm_net = IntrinsicCuriosityModule(feature_net.net,
feature_dim,
action_dim,
device=args.device)
icm_optim = torch.optim.Adam(icm_net.parameters(), lr=args.lr)
policy = ICMPolicy(policy, icm_net, icm_optim, args.icm_lr_scale,
args.icm_reward_scale,
args.icm_forward_loss_weight).to(args.device)
# load a previous policy
if args.resume_path:
policy.load_state_dict(
torch.load(args.resume_path, map_location=args.device))
print("Loaded agent from: ", args.resume_path)
# replay buffer: `save_last_obs` and `stack_num` can be removed together
# when you have enough RAM
buffer = VectorReplayBuffer(args.buffer_size,
buffer_num=len(train_envs),
ignore_obs_next=True,
save_only_last_obs=True,
stack_num=args.frames_stack)
# collector
train_collector = Collector(policy,
train_envs,
buffer,
exploration_noise=True)
test_collector = Collector(policy, test_envs, exploration_noise=True)
# log
log_name = 'ppo_icm' if args.icm_lr_scale > 0 else 'ppo'
log_path = os.path.join(args.logdir, args.task, log_name)
writer = SummaryWriter(log_path)
writer.add_text("args", str(args))
logger = TensorboardLogger(writer)
def save_best_fn(policy):
torch.save(policy.state_dict(), os.path.join(log_path, 'policy.pth'))
def stop_fn(mean_rewards):
if env.spec.reward_threshold:
return mean_rewards >= env.spec.reward_threshold
elif 'Pong' in args.task:
return mean_rewards >= 20
else:
return False
# watch agent's performance
def watch():
print("Setup test envs ...")
policy.eval()
test_envs.seed(args.seed)
if args.save_buffer_name:
print(f"Generate buffer with size {args.buffer_size}")
buffer = VectorReplayBuffer(args.buffer_size,
buffer_num=len(test_envs),
ignore_obs_next=True,
save_only_last_obs=True,
stack_num=args.frames_stack)
collector = Collector(policy,
test_envs,
buffer,
exploration_noise=True)
result = collector.collect(n_step=args.buffer_size)
print(f"Save buffer into {args.save_buffer_name}")
# Unfortunately, pickle will cause oom with 1M buffer size
buffer.save_hdf5(args.save_buffer_name)
else:
print("Testing agent ...")
test_collector.reset()
result = test_collector.collect(n_episode=args.test_num,
render=args.render)
rew = result["rews"].mean()
lens = result["lens"].mean() * args.skip_num
print(f'Mean reward (over {result["n/ep"]} episodes): {rew}')
print(f'Mean length (over {result["n/ep"]} episodes): {lens}')
if args.watch:
watch()
exit(0)
# test train_collector and start filling replay buffer
train_collector.collect(n_step=args.batch_size * args.training_num)
# trainer
result = onpolicy_trainer(policy,
train_collector,
test_collector,
args.epoch,
args.step_per_epoch,
args.repeat_per_collect,
args.test_num,
args.batch_size,
step_per_collect=args.step_per_collect,
stop_fn=stop_fn,
save_best_fn=save_best_fn,
logger=logger,
test_in_train=False)
pprint.pprint(result)
watch()
def test_ppo(args=get_args()):
torch.set_num_threads(1) # for poor CPU
env = gym.make(args.task)
args.state_shape = env.observation_space.shape or env.observation_space.n
args.action_shape = env.action_space.shape or env.action_space.n
# train_envs = gym.make(args.task)
# you can also use tianshou.env.SubprocVectorEnv
train_envs = VectorEnv(
[lambda: gym.make(args.task) for _ in range(args.training_num)])
# test_envs = gym.make(args.task)
test_envs = VectorEnv(
[lambda: gym.make(args.task) for _ in range(args.test_num)])
# seed
np.random.seed(args.seed)
torch.manual_seed(args.seed)
train_envs.seed(args.seed)
test_envs.seed(args.seed)
# model
net = Net(args.layer_num, args.state_shape, device=args.device)
actor = Actor(net, args.action_shape).to(args.device)
critic = Critic(net).to(args.device)
optim = torch.optim.Adam(list(actor.parameters()) +
list(critic.parameters()),
lr=args.lr)
dist = torch.distributions.Categorical
policy = PPOPolicy(actor,
critic,
optim,
dist,
args.gamma,
max_grad_norm=args.max_grad_norm,
eps_clip=args.eps_clip,
vf_coef=args.vf_coef,
ent_coef=args.ent_coef,
action_range=None,
gae_lambda=args.gae_lambda,
reward_normalization=args.rew_norm,
dual_clip=args.dual_clip,
value_clip=args.value_clip)
# collector
train_collector = Collector(policy, train_envs,
ReplayBuffer(args.buffer_size))
test_collector = Collector(policy, test_envs)
# log
log_path = os.path.join(args.logdir, args.task, 'ppo')
writer = SummaryWriter(log_path)
def save_fn(policy):
torch.save(policy.state_dict(), os.path.join(log_path, 'policy.pth'))
def stop_fn(x):
return x >= env.spec.reward_threshold
# trainer
result = onpolicy_trainer(policy,
train_collector,
test_collector,
args.epoch,
args.step_per_epoch,
args.collect_per_step,
args.repeat_per_collect,
args.test_num,
args.batch_size,
stop_fn=stop_fn,
save_fn=save_fn,
writer=writer)
assert stop_fn(result['best_reward'])
train_collector.close()
test_collector.close()
pprint.pprint(result)
writer = SummaryWriter(f'{args.logdir}/{args.task}/pg/{args.note}')
=======
log_path = os.path.join(args.logdir, args.task, 'pg')
writer = SummaryWriter(log_path)
def save_fn(policy):
torch.save(policy.state_dict(), os.path.join(log_path, 'policy.pth'))
>>>>>>> 4fd826761c9884457928da9dac52d7ee1c51443a
def stop_fn(x):
return x >= env.spec.reward_threshold
# trainer
result = onpolicy_trainer(
policy, train_collector, test_collector, args.epoch,
args.step_per_epoch, args.collect_per_step, args.repeat_per_collect,
args.test_num, args.batch_size, stop_fn=stop_fn, save_fn=save_fn,
writer=writer)
assert stop_fn(result['best_reward'])
train_collector.close()
test_collector.close()
if __name__ == '__main__':
pprint.pprint(result)
# Let's watch its performance!
env = gym.make(args.task)
collector = Collector(policy, env)
result = collector.collect(n_episode=1, render=args.render)
print(f'Final reward: {result["rew"]}, length: {result["len"]}')
collector.close()
def test_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_a2c(args=get_args()):
env = make_atari_env(args)
args.state_shape = env.observation_space.shape or env.observation_space.n
args.action_shape = env.env.action_space.shape or env.env.action_space.n
# should be N_FRAMES x H x W
print("Observations shape: ", args.state_shape)
print("Actions shape: ", args.action_shape)
# make environments
train_envs = SubprocVectorEnv(
[lambda: make_atari_env(args)
for _ in range(args.training_num)])
# test_envs = gym.make(args.task)
test_envs = SubprocVectorEnv(
[lambda: make_atari_env_watch(args)
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 = DQN(*args.state_shape,
args.hidden_layer_size, args.device).to(args.device)
actor = Actor(net, args.action_shape,
hidden_layer_size=args.hidden_layer_size,
softmax_output=False).to(args.device)
critic = Critic(net,
hidden_layer_size=args.hidden_layer_size).to(args.device)
optim = torch.optim.Adam(list(
actor.parameters()) + list(critic.parameters()), lr=args.lr)
def dist(x):
return torch.distributions.Categorical(logits=x)
# define policy
policy = A2CPolicy(
actor, critic, optim, dist, args.gamma, vf_coef=args.vf_coef,
ent_coef=args.ent_coef, max_grad_norm=args.max_grad_norm)
# 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, ignore_obs_next=True))
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):
if env.env.spec.reward_threshold:
return x >= env.spec.reward_threshold
elif 'Pong' in args.task:
return x >= 20
# watch agent's performance
def watch():
print("Testing agent ...")
policy.eval()
policy.set_eps(args.eps_test)
envs = SubprocVectorEnv([lambda: make_atari_env_watch(args)
for _ in range(args.test_num)])
envs.seed(args.seed)
collector = Collector(policy, envs)
result = collector.collect(n_episode=args.test_num, render=args.render)
pprint.pprint(result)
if args.watch:
watch()
exit(0)
# test train_collector and start filling replay buffer
train_collector.collect(n_step=args.batch_size * 4)
# trainer
result = onpolicy_trainer(
policy, train_collector, test_collector, args.epoch,
args.step_per_epoch, args.collect_per_step, args.repeat_per_collect,
args.test_num, args.batch_size, stop_fn=stop_fn, writer=writer,
save_fn=save_fn, test_in_train=False)
pprint.pprint(result)
watch()