def testing_ddpg(args=get_args()):
env = EnvThreeUsers(args.step_per_epoch)
args.state_shape = env.observation_space.shape
args.action_shape = env.action_space.shape
args.max_action = env.action_space.high[0]
# model
net = Net(args.layer_num,
args.state_shape,
0,
device=args.device,
hidden_layer_size=args.unit_num)
actor = Actor(net,
args.action_shape,
args.max_action,
args.device,
hidden_layer_size=args.unit_num).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,
hidden_layer_size=args.unit_num)
critic = Critic(net, args.device, args.unit_num).to(args.device)
critic_optim = torch.optim.Adam(critic.parameters(), lr=args.critic_lr)
policy = DDPGPolicy(
actor,
actor_optim,
critic,
critic_optim,
args.tau,
args.gamma,
OUNoise(sigma=args.exploration_noise),
# GaussianNoise(sigma=args.exploration_noise),
[env.action_space.low[0], env.action_space.high[0]],
reward_normalization=True,
ignore_done=True)
# restore model
log_path = os.path.join(args.logdir, args.task, 'ddpg')
policy.load_state_dict(torch.load(os.path.join(log_path, 'policy.pth')))
print('\nrelode model!')
env = EnvThreeUsers(args.step_per_epoch)
collector = Collector(policy, env)
ep = 10000
result = collector.collect(n_episode=ep, render=args.render)
print('''\nty1_succ_1: {:.6f}, q_len_1: {:.6f},
\nty1_succ_2: {:.2f}, q_len_2: {:.2f},
\nty1_succ_3: {:.2f}, q_len_3: {:.2f},
\nee_1: {:.2f}, ee_2: {:.2f}, ee_3: {:.2f},
\navg_rate:{:.2f}, \navg_power:{:.2f}\n'''.format(
result["ty1s_1"][0] / ep, result["ql_1"][0] / ep,
result["ty1s_2"][0] / ep, result["ql_2"][0] / ep,
result["ty1s_3"][0] / ep, result["ql_3"][0] / ep,
result["ee_1"][0] / ep, result["ee_2"][0] / ep, result["ee_3"][0] / ep,
result["avg_r"] / ep, result["avg_p"] / ep))
print('large than Qmax: users1: {}, users2: {}, users3: {}.'.format(
str(env.large_than_Q_1), str(env.large_than_Q_2),
str(env.large_than_Q_3)))
collector.close()
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 __init__(self, args, mask=None, name=None):
env = gym.make(args.task)
if args.task == 'Pendulum-v0':
env.spec.reward_threshold = -250
self.state_shape = env.observation_space.shape or env.observation_space.n
self.action_shape = env.action_space.shape or env.action_space.n
self.max_action = env.action_space.high[0]
self.stop_fn = lambda x: x >= env.spec.reward_threshold
# env
self.train_envs = VectorEnv(
[lambda: gym.make(args.task) for _ in range(args.training_num)])
self.test_envs = SubprocVectorEnv(
[lambda: gym.make(args.task) for _ in range(args.test_num)])
state_dim = int(np.prod(self.state_shape))
self._view_mask = torch.ones(state_dim)
if mask == 'even':
for i in range(0, state_dim, 2):
self._view_mask[i] = 0
elif mask == "odd":
for i in range(1, state_dim, 2):
self._view_mask[i] = 0
# policy
self.actor = ActorWithView(args.layer_num, self.state_shape,
self.action_shape, self.max_action,
self._view_mask,
args.device).to(args.device)
self.actor_optim = torch.optim.Adam(self.actor.parameters(),
lr=args.actor_lr)
self.critic = CriticWithView(args.layer_num, self.state_shape,
self._view_mask, self.action_shape,
args.device).to(args.device)
self.critic_optim = torch.optim.Adam(self.critic.parameters(),
lr=args.critic_lr)
self.policy = DDPGPolicy(
self.actor,
self.actor_optim,
self.critic,
self.critic_optim,
args.tau,
args.gamma,
args.exploration_noise,
[env.action_space.low[0], env.action_space.high[0]],
reward_normalization=True,
ignore_done=True)
# collector
self.train_collector = Collector(self.policy, self.train_envs,
ReplayBuffer(args.buffer_size))
self.test_collector = Collector(self.policy, self.test_envs)
# log
self.writer = SummaryWriter(
f"{args.logdir}/{args.task}/ddpg/{args.note}/{name}")
def test(args=get_args()):
torch.set_num_threads(1) # we just need only one thread for NN
env = gym_make()
model_path = os.path.join(args.logdir, args.task, 'ddpg/policy.pth')
layer = [1024, 512, 512, 512]
device = 'cuda'
state_shape = env.observation_space.shape or env.observation_space.n
action_shape = env.action_space.shape or env.action_space.n
action_range = [env.action_space.low, env.action_space.high]
actor = Actor(
layer, state_shape, action_shape,
action_range, device
).to(device)
critic = Critic(
layer, state_shape, action_shape, device
).to(device)
actor = actor.to(device)
actor_optim = torch.optim.Adam(actor.parameters())
critic = critic.to(device)
critic_optim = torch.optim.Adam(critic.parameters())
policy = DDPGPolicy(
actor, actor_optim, critic, critic_optim,
action_range=action_range)
policy.load_state_dict(torch.load(model_path, map_location=device))
obs = env.reset()
# env.state[0] = -30.0
# env.goal[0] = 30.0
env.render()
print(env.goal)
while True:
action, _ = policy.actor(obs.reshape(1,-1), eps=0.01)
action = action.detach().cpu().numpy()[0]
obs, reward, done, info = env.step(action)
# print(env.state)
# print(reward)
# print(info)
env.render()
if done:
break
def load_policy(env, layer, model_path):
state_shape = env.observation_space.shape or env.observation_space.n
action_shape = env.action_space.shape or env.action_space.n
action_range = [env.action_space.low, env.action_space.high]
actor = Actor(layer, state_shape, action_shape, action_range,
device).to(device)
critic = Critic(layer, state_shape, action_shape, device).to(device)
# actor critic
actor = actor.to(device)
actor_optim = torch.optim.Adam(actor.parameters())
critic = critic.to(device)
critic_optim = torch.optim.Adam(critic.parameters())
policy = DDPGPolicy(actor,
actor_optim,
critic,
critic_optim,
action_range=action_range)
policy.load_state_dict(torch.load(model_path, map_location=device))
return policy
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
t0 = datetime.datetime.now().strftime("%m%d_%H%M%S")
log_file = f'seed_{args.seed}_{t0}-{args.task.replace("-", "_")}_ddpg'
log_path = os.path.join(args.logdir, args.task, 'ddpg', 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_ddpg(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 = Actor(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)
critic = Critic(args.layer_num, args.state_shape, args.action_shape,
args.device).to(args.device)
critic_optim = torch.optim.Adam(critic.parameters(), lr=args.critic_lr)
policy = DDPGPolicy(actor,
actor_optim,
critic,
critic_optim,
args.tau,
args.gamma,
GaussianNoise(sigma=args.exploration_noise),
[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)
# log
log_path = os.path.join(args.logdir, args.task, 'ddpg')
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'])
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_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_ddpg(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 = Actor(net,
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 = Net(args.state_shape,
args.action_shape,
hidden_sizes=args.hidden_sizes,
concat=True,
device=args.device)
critic = Critic(net, 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),
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)
# log
log_path = os.path.join(args.logdir, args.task, 'ddpg')
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()}")
def train(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_space.low, env.action_space.high]
train_envs = VectorEnv(
[lambda: gym_make() for _ in range(args.training_num)])
# test_envs = gym.make(args.task)
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
actor = Actor(args.layer, args.state_shape, args.action_shape,
args.action_range, args.device).to(args.device)
actor_optim = torch.optim.Adam(actor.parameters(), lr=args.actor_lr)
critic = Critic(args.layer, args.state_shape, args.action_shape,
args.device).to(args.device)
critic_optim = torch.optim.Adam(critic.parameters(), lr=args.critic_lr)
policy = DDPGPolicy(actor,
actor_optim,
critic,
critic_optim,
args.tau,
args.gamma,
args.exploration_noise,
args.action_range,
reward_normalization=args.rew_norm,
ignore_done=True)
# 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, 'ddpg')
writer = SummaryWriter(log_path)
# if a model exist, continue to train it
model_path = os.path.join(log_path, 'policy.pth')
if os.path.exists(model_path):
policy.load_state_dict(torch.load(model_path))
def save_fn(policy):
torch.save(policy.state_dict(), model_path)
def stop_fn(x):
return x >= 100
# 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,
save_fn=save_fn,
writer=writer)
train_collector.close()
test_collector.close()
if __name__ == '__main__':
# 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 training_ddpg(args=get_args()):
env = EnvTwoUsers(args.step_per_epoch)
args.state_shape = env.observation_space.shape
args.action_shape = env.action_space.shape
args.max_action = env.action_space.high[0]
train_envs = VectorEnv([
lambda: EnvTwoUsers(args.step_per_epoch)
for _ in range(args.training_num)
])
test_envs = VectorEnv([
lambda: EnvTwoUsers(args.step_per_epoch) 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,
hidden_layer_size=args.unit_num)
actor = Actor(net,
args.action_shape,
args.max_action,
args.device,
hidden_layer_size=args.unit_num).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,
hidden_layer_size=args.unit_num)
critic = Critic(net, args.device,
hidden_layer_size=args.unit_num).to(args.device)
critic_optim = torch.optim.Adam(critic.parameters(), lr=args.critic_lr)
# 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)
policy = DDPGPolicy(
actor,
actor_optim,
critic,
critic_optim,
args.tau,
args.gamma,
OUNoise(sigma=args.exploration_noise),
# GaussianNoise(sigma=args.exploration_noise),
[env.action_space.low[0], env.action_space.high[0]],
reward_normalization=True,
ignore_done=True)
# 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, 'ddpg')
if not os.path.exists(log_path):
os.makedirs(log_path)
# writer = SummaryWriter(log_path)
writer = None
# policy.load_state_dict(torch.load(os.path.join(log_path, 'policy.pth')))
# print('reload model!')
def save_fn(policy):
torch.save(policy.state_dict(), os.path.join(log_path, 'policy.pth'))
def stop_fn(x):
return x >= 1e16
# 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)
train_collector.close()
test_collector.close()
