def __init__(
self,
env: gym.Env,
args: argparse.Namespace,
hyper_params: dict,
models: tuple,
optims: tuple,
noise: OUNoise,
):
"""Initialization.
Args:
env (gym.Env): openAI Gym environment with discrete action space
args (argparse.Namespace): arguments including hyperparameters and training settings
hyper_params (dict): hyper-parameters
models (tuple): models including actor and critic
optims (tuple): optimizers for actor and critic
noise (OUNoise): random noise for exploration
"""
AbstractAgent.__init__(self, env, args)
self.actor, self.actor_target, self.critic, self.critic_target = models
self.actor_optimizer, self.critic_optimizer = optims
self.hyper_params = hyper_params
self.curr_state = np.zeros((1, ))
self.noise = noise
# load the optimizer and model parameters
if args.load_from is not None and os.path.exists(args.load_from):
self.load_params(args.load_from)
# replay memory
self.memory = ReplayBuffer(hyper_params["BUFFER_SIZE"],
hyper_params["BATCH_SIZE"], self.args.seed)
def _initialize(self):
"""Initialize non-common things."""
# load demo replay memory
with open(self.args.demo_path, "rb") as f:
demo = list(pickle.load(f))
# HER
if self.hyper_params["USE_HER"]:
if self.hyper_params["DESIRED_STATES_FROM_DEMO"]:
self.her.fetch_desired_states_from_demo(demo)
self.transitions_epi: list = list()
self.desired_state = np.zeros((1,))
demo = self.her.generate_demo_transitions(demo)
if not self.args.test:
# Replay buffers
demo_batch_size = self.hyper_params["DEMO_BATCH_SIZE"]
self.demo_memory = ReplayBuffer(len(demo), demo_batch_size)
self.demo_memory.extend(demo)
self.memory = ReplayBuffer(
self.hyper_params["BUFFER_SIZE"], self.hyper_params["BATCH_SIZE"]
)
# set hyper parameters
self.lambda1 = self.hyper_params["LAMBDA1"]
self.lambda2 = self.hyper_params["LAMBDA2"] / demo_batch_size
def _initialize(self):
"""Initialize non-common things."""
self.use_n_step = self.hyper_params["N_STEP"] > 1
if not self.args.test:
# load demo replay memory
with open(self.args.demo_path, "rb") as f:
demos = pickle.load(f)
if self.use_n_step:
demos, demos_n_step = common_utils.get_n_step_info_from_demo(
demos, self.hyper_params["N_STEP"], self.hyper_params["GAMMA"]
)
# replay memory for multi-steps
self.memory_n = ReplayBuffer(
buffer_size=self.hyper_params["BUFFER_SIZE"],
n_step=self.hyper_params["N_STEP"],
gamma=self.hyper_params["GAMMA"],
demo=demos_n_step,
)
# replay memory for a single step
self.beta = self.hyper_params["PER_BETA"]
self.memory = PrioritizedReplayBuffer(
self.hyper_params["BUFFER_SIZE"],
self.hyper_params["BATCH_SIZE"],
demo=demos,
alpha=self.hyper_params["PER_ALPHA"],
epsilon_d=self.hyper_params["PER_EPS_DEMO"],
)
def _initialize(self):
"""Initialize non-common things."""
if not self.args.test:
# load demo replay memory
demos = self._load_demos()
if self.use_n_step:
demos, demos_n_step = common_utils.get_n_step_info_from_demo(
demos, self.hyper_params["N_STEP"],
self.hyper_params["GAMMA"])
self.memory_n = ReplayBuffer(
buffer_size=self.hyper_params["BUFFER_SIZE"],
n_step=self.hyper_params["N_STEP"],
gamma=self.hyper_params["GAMMA"],
demo=demos_n_step,
)
# replay memory
self.beta = self.hyper_params["PER_BETA"]
self.memory = PrioritizedReplayBuffer(
self.hyper_params["BUFFER_SIZE"],
self.hyper_params["BATCH_SIZE"],
demo=demos,
alpha=self.hyper_params["PER_ALPHA"],
epsilon_d=self.hyper_params["PER_EPS_DEMO"],
)
def _initialize(self):
"""Initialize non-common things."""
if not self.args.test:
# replay memory
self.memory = ReplayBuffer(
self.hyper_params["BUFFER_SIZE"], self.hyper_params["BATCH_SIZE"]
)
def _initialize(self):
"""Initialize non-common things."""
# load demo replay memory
with open(self.args.demo_path, "rb") as f:
demo = list(pickle.load(f))
# HER
if self.hyper_params["USE_HER"]:
self.her = HER(self.args.demo_path)
self.transitions_epi: list = list()
self.desired_state = np.zeros((1,))
self.hook_transition = True
demo = self.her.generate_demo_transitions(demo)
if not self.args.test:
# Replay buffers
self.demo_memory = ReplayBuffer(
len(demo), self.hyper_params["DEMO_BATCH_SIZE"]
)
self.demo_memory.extend(demo)
self.memory = ReplayBuffer(
self.hyper_params["BUFFER_SIZE"], self.hyper_params["BATCH_SIZE"]
)
# set hyper parameters
self.lambda1 = self.hyper_params["LAMBDA1"]
self.lambda2 = (
self.hyper_params["LAMBDA2"] / self.hyper_params["DEMO_BATCH_SIZE"]
)
def __init__(
self,
env: gym.Env,
args: argparse.Namespace,
hyper_params: dict,
models: tuple,
optims: tuple,
exploration_noise: GaussianNoise,
target_policy_noise: GaussianNoise,
):
"""Initialization.
Args:
env (gym.Env): openAI Gym environment
args (argparse.Namespace): arguments including hyperparameters and training settings
hyper_params (dict): hyper-parameters
models (tuple): models including actor and critic
optims (tuple): optimizers for actor and critic
exploration_noise (GaussianNoise): random noise for exploration
target_policy_noise (GaussianNoise): random noise for target values
"""
Agent.__init__(self, env, args)
self.actor, self.actor_target = models[0:2]
self.critic1, self.critic2 = models[2:4]
self.critic_target1, self.critic_target2 = models[4:6]
self.actor_optim = optims[0]
self.critic_optim = optims[1]
self.hyper_params = hyper_params
self.curr_state = np.zeros((1,))
self.exploration_noise = exploration_noise
self.target_policy_noise = target_policy_noise
self.total_step = 0
self.episode_step = 0
self.update_step = 0
self.i_episode = 0
# load the optimizer and model parameters
if args.load_from is not None and os.path.exists(args.load_from):
self.load_params(args.load_from)
if not self.args.test:
# replay memory
self.memory = ReplayBuffer(
hyper_params["BUFFER_SIZE"], hyper_params["BATCH_SIZE"]
)
def _initialize(self):
"""Initialize non-common things."""
if not self.args.test:
# replay memory for a single step
self.beta = self.hyper_params["PER_BETA"]
self.memory = PrioritizedReplayBuffer(
self.hyper_params["BUFFER_SIZE"],
self.hyper_params["BATCH_SIZE"],
alpha=self.hyper_params["PER_ALPHA"],
)
# replay memory for multi-steps
if self.use_n_step:
self.memory_n = ReplayBuffer(
self.hyper_params["BUFFER_SIZE"],
n_step=self.hyper_params["N_STEP"],
gamma=self.hyper_params["GAMMA"],
)
def _initialize(self):
"""Initialize non-common things."""
if not self.args.test:
# replay memory
self.memory = ReplayBuffer(self.hyper_params["BUFFER_SIZE"],
self.hyper_params["BATCH_SIZE"])
# HER
if self.hyper_params["USE_HER"]:
# load demo replay memory
with open(self.args.demo_path, "rb") as f:
demo = pickle.load(f)
if self.hyper_params["DESIRED_STATES_FROM_DEMO"]:
self.her.fetch_desired_states_from_demo(demo)
self.transitions_epi = list()
self.desired_state = np.zeros((1, ))
demo = self.her.generate_demo_transitions(demo)
if not self.args.test:
# Replay buffers
self.memory = ReplayBuffer(self.hyper_params["BUFFER_SIZE"],
self.hyper_params["BATCH_SIZE"])
class Agent(SACAgent):
"""SAC agent interacting with environment.
Attrtibutes:
memory (ReplayBuffer): replay memory
demo_memory (ReplayBuffer): replay memory for demo
her (HER): hinsight experience replay
transitions_epi (list): transitions per episode (for HER)
goal_state (np.ndarray): goal state to generate concatenated states
total_step (int): total step numbers
episode_step (int): step number of the current episode
"""
# pylint: disable=attribute-defined-outside-init
def _initialize(self):
"""Initialize non-common things."""
# load demo replay memory
with open(self.args.demo_path, "rb") as f:
demo = list(pickle.load(f))
# HER
if self.hyper_params["USE_HER"]:
self.her = HER(self.args.demo_path)
self.transitions_epi: list = list()
self.desired_state = np.zeros((1,))
self.hook_transition = True
demo = self.her.generate_demo_transitions(demo)
if not self.args.test:
# Replay buffers
self.demo_memory = ReplayBuffer(
len(demo), self.hyper_params["DEMO_BATCH_SIZE"]
)
self.demo_memory.extend(demo)
self.memory = ReplayBuffer(
self.hyper_params["BUFFER_SIZE"], self.hyper_params["BATCH_SIZE"]
)
# set hyper parameters
self.lambda1 = self.hyper_params["LAMBDA1"]
self.lambda2 = (
self.hyper_params["LAMBDA2"] / self.hyper_params["DEMO_BATCH_SIZE"]
)
def select_action(self, state: np.ndarray) -> np.ndarray:
"""Select an action from the input space."""
state_ = state
# HER
if self.hyper_params["USE_HER"]:
self.desired_state = self.her.sample_desired_state()
state = np.concatenate((state, self.desired_state), axis=-1)
selected_action = SACAgent.select_action(self, state)
self.curr_state = state_
return selected_action
def step(self, action: np.ndarray) -> Tuple[np.ndarray, np.float64, bool]:
"""Take an action and return the response of the env."""
next_state, reward, done = SACAgent.step(self, action)
if not self.args.test and self.hyper_params["USE_HER"]:
self.transitions_epi.append(self.hooked_transition)
if done:
# insert generated transitions if the episode is done
transitions = self.her.generate_transitions(
self.transitions_epi, self.desired_state
)
self.memory.extend(transitions)
return next_state, reward, done
def update_model(self) -> Tuple[torch.Tensor, ...]:
"""Train the model after each episode."""
experiences = self.memory.sample()
demos = self.demo_memory.sample()
states, actions, rewards, next_states, dones = experiences
demo_states, demo_actions, _, _, _ = demos
new_actions, log_prob, pre_tanh_value, mu, std = self.actor(states)
pred_actions, _, _, _, _ = self.actor(demo_states)
# train alpha
if self.hyper_params["AUTO_ENTROPY_TUNING"]:
alpha_loss = (
-self.log_alpha * (log_prob + self.target_entropy).detach()
).mean()
self.alpha_optimizer.zero_grad()
alpha_loss.backward()
self.alpha_optimizer.step()
alpha = self.log_alpha.exp()
else:
alpha_loss = torch.zeros(1)
alpha = self.hyper_params["W_ENTROPY"]
# Q function loss
masks = 1 - dones
q_1_pred = self.qf_1(states, actions)
q_2_pred = self.qf_2(states, actions)
v_target = self.vf_target(next_states)
q_target = rewards + self.hyper_params["GAMMA"] * v_target * masks
qf_1_loss = F.mse_loss(q_1_pred, q_target.detach())
qf_2_loss = F.mse_loss(q_2_pred, q_target.detach())
# V function loss
v_pred = self.vf(states)
q_pred = torch.min(
self.qf_1(states, new_actions), self.qf_2(states, new_actions)
)
v_target = q_pred - alpha * log_prob
vf_loss = F.mse_loss(v_pred, v_target.detach())
# train Q functions
self.qf_1_optimizer.zero_grad()
qf_1_loss.backward()
self.qf_1_optimizer.step()
self.qf_2_optimizer.zero_grad()
qf_2_loss.backward()
self.qf_2_optimizer.step()
# train V function
self.vf_optimizer.zero_grad()
vf_loss.backward()
self.vf_optimizer.step()
if self.total_step % self.hyper_params["DELAYED_UPDATE"] == 0:
# bc loss
qf_mask = torch.gt(
self.qf_1(demo_states, demo_actions),
self.qf_1(demo_states, pred_actions),
).to(device)
qf_mask = qf_mask.float()
n_qf_mask = int(qf_mask.sum().item())
if n_qf_mask == 0:
bc_loss = torch.zeros(1, device=device)
else:
bc_loss = (
torch.mul(pred_actions, qf_mask) - torch.mul(demo_actions, qf_mask)
).pow(2).sum() / n_qf_mask
# actor loss
advantage = q_pred - v_pred.detach()
actor_loss = (alpha * log_prob - advantage).mean()
actor_loss = self.lambda1 * actor_loss + self.lambda2 * bc_loss
# regularization
if not self.is_discrete: # iff the action is continuous
mean_reg = self.hyper_params["W_MEAN_REG"] * mu.pow(2).mean()
std_reg = self.hyper_params["W_STD_REG"] * std.pow(2).mean()
pre_activation_reg = self.hyper_params["W_PRE_ACTIVATION_REG"] * (
pre_tanh_value.pow(2).sum(dim=-1).mean()
)
actor_reg = mean_reg + std_reg + pre_activation_reg
# actor loss + regularization
actor_loss += actor_reg
# train actor
self.actor_optimizer.zero_grad()
actor_loss.backward()
self.actor_optimizer.step()
# update target networks
common_utils.soft_update(self.vf, self.vf_target, self.hyper_params["TAU"])
else:
actor_loss = torch.zeros(1)
return (
actor_loss.data,
qf_1_loss.data,
qf_2_loss.data,
vf_loss.data,
alpha_loss.data,
)
class Agent(AbstractAgent):
"""ActorCritic interacting with environment.
Attributes:
memory (ReplayBuffer): replay memory
exploration_noise (GaussianNoise): random noise for exploration
target_policy_noise (GaussianNoise): random noise for regularization
hyper_params (dict): hyper-parameters
actor (nn.Module): actor model to select actions
actor_target (nn.Module): target actor model to select actions
critic1 (nn.Module): critic1 model to predict state values
critic2 (nn.Module): critic2 model to predict state values
critic1_target (nn.Module): target critic1 model to predict state values
critic2_target (nn.Module): target critic2 model to predict state values
actor_optim (Optimizer): optimizer for training actor
critic1_optim (Optimizer): optimizer for training critic1
critic2_optim (Optimizer): optimizer for training critic2
curr_state (np.ndarray): temporary storage of the current state
"""
def __init__(self, env, args, hyper_params, models, optims, noises):
"""Initialization.
Args:
env (gym.Env): openAI Gym environment with discrete action space
args (argparse.Namespace): arguments including hyperparameters and training settings
hyper_params (dict): hyper-parameters
models (tuple): models including actor and critics
optims (tuple): optimizers for actor and critics
noises (tuple): noises for exploration and regularization
"""
AbstractAgent.__init__(self, env, args)
self.actor, self.actor_target = models[:2]
self.critic1, self.critic1_target = models[2:4]
self.critic2, self.critic2_target = models[4:]
self.actor_optim, self.critic_optim = optims
self.hyper_params = hyper_params
self.exploration_noise, self.target_policy_noise = noises
self.curr_state = np.zeros((1, ))
self.total_steps = 0
self.episode_steps = 0
# load the optimizer and model parameters
if args.load_from is not None and os.path.exists(args.load_from):
self.load_params(args.load_from)
self._initialize()
def _initialize(self):
"""Initialize non-common things."""
if not self.args.test:
# replay memory
self.memory = ReplayBuffer(self.hyper_params["BUFFER_SIZE"],
self.hyper_params["BATCH_SIZE"])
def select_action(self, state):
"""Select an action from the input space."""
# initial training step, try random action for exploration
random_action_count = self.hyper_params["INITIAL_RANDOM_ACTIONS"]
self.curr_state = state
if self.total_steps < random_action_count and not self.args.test:
return self.env.action_space.sample()
state = torch.FloatTensor(state).to(device)
selected_action = self.actor(state)
if not self.args.test:
noise = torch.FloatTensor(
self.exploration_noise.sample()).to(device)
selected_action = (selected_action + noise).clamp(-1.0, 1.0)
return selected_action.detach().cpu().numpy()
def step(self, action):
"""Take an action and return the response of the env."""
self.total_steps += 1
self.episode_steps += 1
next_state, reward, done, _ = self.env.step(action)
if not self.args.test:
# if the last state is not a terminal state, store done as false
done_bool = (False if self.episode_steps
== self.args.max_episode_steps else done)
transition = (self.curr_state, action, reward, next_state,
done_bool)
self._add_transition_to_memory(transition)
return next_state, reward, done
def _add_transition_to_memory(self, transition):
"""Add 1 step and n step transitions to memory."""
self.memory.add(*transition)
def update_model(self, experiences):
"""Train the model after each episode."""
states, actions, rewards, next_states, dones = experiences
# G_t = r + gamma * v(s_{t+1}) if state != Terminal
# = r otherwise
masks = 1 - dones
noise = torch.FloatTensor(self.target_policy_noise.sample()).to(device)
clipped_noise = torch.clamp(
noise,
-self.hyper_params["TARGET_POLICY_NOISE_CLIP"],
self.hyper_params["TARGET_POLICY_NOISE_CLIP"],
)
next_actions = (self.actor_target(next_states) + clipped_noise).clamp(
-1.0, 1.0)
target_values1 = self.critic1_target(
torch.cat((next_states, next_actions), dim=-1))
target_values2 = self.critic2_target(
torch.cat((next_states, next_actions), dim=-1))
target_values = torch.min(target_values1, target_values2)
target_values = (
rewards +
(self.hyper_params["GAMMA"] * target_values * masks).detach())
# train critic
values1 = self.critic1(torch.cat((states, actions), dim=-1))
critic1_loss = F.mse_loss(values1, target_values)
values2 = self.critic2(torch.cat((states, actions), dim=-1))
critic2_loss = F.mse_loss(values2, target_values)
critic_loss = critic1_loss + critic2_loss
self.critic_optim.zero_grad()
critic_loss.backward()
self.critic_optim.step()
if self.episode_steps % self.hyper_params["POLICY_UPDATE_FREQ"] == 0:
# train actor
actions = self.actor(states)
actor_loss = -self.critic1(torch.cat(
(states, actions), dim=-1)).mean()
self.actor_optim.zero_grad()
actor_loss.backward()
self.actor_optim.step()
# update target networks
tau = self.hyper_params["TAU"]
common_utils.soft_update(self.actor, self.actor_target, tau)
common_utils.soft_update(self.critic1, self.critic1_target, tau)
common_utils.soft_update(self.critic2, self.critic2_target, tau)
else:
actor_loss = torch.zeros(1)
return actor_loss.data, critic1_loss.data, critic2_loss.data
def load_params(self, path):
"""Load model and optimizer parameters."""
if not os.path.exists(path):
print("[ERROR] the input path does not exist. ->", path)
return
params = torch.load(path)
self.actor.load_state_dict(params["actor_state_dict"])
self.actor_target.load_state_dict(params["actor_target_state_dict"])
self.critic1.load_state_dict(params["critic1_state_dict"])
self.critic2.load_state_dict(params["critic2_state_dict"])
self.critic1_target.load_state_dict(
params["critic1_target_state_dict"])
self.critic2_target.load_state_dict(
params["critic2_target_state_dict"])
self.actor_optim.load_state_dict(params["actor_optim_state_dict"])
self.critic_optim.load_state_dict(params["critic_optim_state_dict"])
print("[INFO] loaded the model and optimizer from", path)
def save_params(self, n_episode):
"""Save model and optimizer parameters."""
params = {
"actor_state_dict": self.actor.state_dict(),
"actor_target_state_dict": self.actor_target.state_dict(),
"critic1_state_dict": self.critic1.state_dict(),
"critic2_state_dict": self.critic2.state_dict(),
"critic1_target_state_dict": self.critic1_target.state_dict(),
"critic2_target_state_dict": self.critic2_target.state_dict(),
"actor_optim_state_dict": self.actor_optim.state_dict(),
"critic_optim_state_dict": self.critic_optim.state_dict(),
}
AbstractAgent.save_params(self, params, n_episode)
def write_log(self, i, loss, score):
"""Write log about loss and score"""
total_loss = loss.sum()
print(
"[INFO] total_steps: %d episode: %d total score: %d, total loss: %f\n"
"actor_loss: %.3f critic1_loss: %.3f critic2_loss: %.3f\n" %
(self.total_steps, i, score, total_loss, loss[0], loss[1],
loss[2]))
if self.args.log:
wandb.log({
"total_steps":
self.total_steps,
"score":
score,
"total loss":
total_loss,
"actor loss":
loss[0] * self.hyper_params["POLICY_UPDATE_FREQ"],
"critic1 loss":
loss[1],
"critic2 loss":
loss[2],
})
def train(self):
"""Train the agent."""
# logger
if self.args.log:
wandb.init()
wandb.config.update(self.hyper_params)
wandb.watch([self.actor, self.critic1, self.critic2],
log="parameters")
for i_episode in range(1, self.args.episode_num + 1):
state = self.env.reset()
done = False
score = 0
loss_episode = list()
self.episode_steps = 0
while not done:
if self.args.render and i_episode >= self.args.render_after:
self.env.render()
action = self.select_action(state)
next_state, reward, done = self.step(action)
if len(self.memory) >= self.hyper_params["BATCH_SIZE"]:
experiences = self.memory.sample()
loss = self.update_model(experiences)
loss_episode.append(loss) # for logging
state = next_state
score += reward
# logging
if loss_episode:
avg_loss = np.vstack(loss_episode).mean(axis=0)
self.write_log(i_episode, avg_loss, score)
if i_episode % self.args.save_period == 0:
self.save_params(i_episode)
# termination
self.env.close()
class Agent(DDPGAgent):
"""ActorCritic interacting with environment.
Attributes:
memory (ReplayBuffer): replay memory
demo_memory (ReplayBuffer): replay memory for demo
her (HER): hinsight experience replay
transitions_epi (list): transitions per episode (for HER)
goal_state (np.ndarray): goal state to generate concatenated states
total_step (int): total step numbers
episode_step (int): step number of the current episode
"""
# pylint: disable=attribute-defined-outside-init
def _initialize(self):
"""Initialize non-common things."""
# load demo replay memory
with open(self.args.demo_path, "rb") as f:
demo = list(pickle.load(f))
# HER
if self.hyper_params["USE_HER"]:
self.her = HER(self.args.demo_path)
self.transitions_epi: list = list()
self.desired_state = np.zeros((1, ))
self.hook_transition = True
demo = self.her.generate_demo_transitions(demo)
if not self.args.test:
# Replay buffers
demo_batch_size = self.hyper_params["DEMO_BATCH_SIZE"]
self.demo_memory = ReplayBuffer(len(demo), demo_batch_size)
self.demo_memory.extend(demo)
self.memory = ReplayBuffer(self.hyper_params["BUFFER_SIZE"],
self.hyper_params["BATCH_SIZE"])
# set hyper parameters
self.lambda1 = self.hyper_params["LAMBDA1"]
self.lambda2 = self.hyper_params["LAMBDA2"] / demo_batch_size
def select_action(self, state: np.ndarray) -> np.ndarray:
"""Select an action from the input space."""
state_ = state
if self.hyper_params["USE_HER"]:
self.desired_state = self.her.sample_desired_state()
state = np.concatenate((state, self.desired_state), axis=-1)
selected_action = DDPGAgent.select_action(self, state)
self.curr_state = state_
return selected_action
def step(self, action: np.ndarray) -> Tuple[np.ndarray, np.float64, bool]:
"""Take an action and return the response of the env."""
next_state, reward, done = DDPGAgent.step(self, action)
if not self.args.test and self.hyper_params["USE_HER"]:
self.transitions_epi.append(self.hooked_transition)
if done:
# insert generated transitions if the episode is done
transitions = self.her.generate_transitions(
self.transitions_epi, self.desired_state)
self.memory.extend(transitions)
return next_state, reward, done
def update_model(self) -> Tuple[torch.Tensor, torch.Tensor]:
"""Train the model after each episode."""
experiences = self.memory.sample()
demos = self.demo_memory.sample()
exp_states, exp_actions, exp_rewards, exp_next_states, exp_dones = experiences
demo_states, demo_actions, demo_rewards, demo_next_states, demo_dones = demos
states = torch.cat((exp_states, demo_states), dim=0)
actions = torch.cat((exp_actions, demo_actions), dim=0)
rewards = torch.cat((exp_rewards, demo_rewards), dim=0)
next_states = torch.cat((exp_next_states, demo_next_states), dim=0)
dones = torch.cat((exp_dones, demo_dones), dim=0)
# G_t = r + gamma * v(s_{t+1}) if state != Terminal
# = r otherwise
masks = 1 - dones
next_actions = self.actor_target(next_states)
next_values = self.critic_target(
torch.cat((next_states, next_actions), dim=-1))
curr_returns = rewards + (self.hyper_params["GAMMA"] * next_values *
masks)
curr_returns = curr_returns.to(device)
# critic loss
values = self.critic(torch.cat((states, actions), dim=-1))
critic_loss = F.mse_loss(values, curr_returns)
# train critic
self.critic_optimizer.zero_grad()
critic_loss.backward()
self.critic_optimizer.step()
# policy loss
actions = self.actor(states)
policy_loss = -self.critic(torch.cat((states, actions), dim=-1)).mean()
# bc loss
pred_actions = self.actor(demo_states)
qf_mask = torch.gt(
self.critic(torch.cat((demo_states, demo_actions), dim=-1)),
self.critic(torch.cat((demo_states, pred_actions), dim=-1)),
).to(device)
qf_mask = qf_mask.float()
n_qf_mask = int(qf_mask.sum().item())
if n_qf_mask == 0:
bc_loss = torch.zeros(1, device=device)
else:
bc_loss = (torch.mul(pred_actions, qf_mask) - torch.mul(
demo_actions, qf_mask)).pow(2).sum() / n_qf_mask
# train actor: pg loss + BC loss
actor_loss = self.lambda1 * policy_loss + self.lambda2 * bc_loss
self.actor_optimizer.zero_grad()
actor_loss.backward()
self.actor_optimizer.step()
# update target networks
tau = self.hyper_params["TAU"]
common_utils.soft_update(self.actor, self.actor_target, tau)
common_utils.soft_update(self.critic, self.critic_target, tau)
return actor_loss.data, critic_loss.data
class DDPGfDAgent(DDPGAgent):
"""ActorCritic interacting with environment.
Attributes:
memory (PrioritizedReplayBuffer): replay memory
beta (float): beta parameter for prioritized replay buffer
"""
# pylint: disable=attribute-defined-outside-init
def _initialize(self):
"""Initialize non-common things."""
self.use_n_step = self.hyper_params["N_STEP"] > 1
if not self.args.test:
# load demo replay memory
with open(self.args.demo_path, "rb") as f:
demos = pickle.load(f)
if self.use_n_step:
demos, demos_n_step = common_utils.get_n_step_info_from_demo(
demos, self.hyper_params["N_STEP"], self.hyper_params["GAMMA"]
)
# replay memory for multi-steps
self.memory_n = ReplayBuffer(
buffer_size=self.hyper_params["BUFFER_SIZE"],
n_step=self.hyper_params["N_STEP"],
gamma=self.hyper_params["GAMMA"],
demo=demos_n_step,
)
# replay memory for a single step
self.beta = self.hyper_params["PER_BETA"]
self.memory = PrioritizedReplayBuffer(
self.hyper_params["BUFFER_SIZE"],
self.hyper_params["BATCH_SIZE"],
demo=demos,
alpha=self.hyper_params["PER_ALPHA"],
epsilon_d=self.hyper_params["PER_EPS_DEMO"],
)
def _add_transition_to_memory(self, transition: Tuple[np.ndarray, ...]):
"""Add 1 step and n step transitions to memory."""
# add n-step transition
if self.use_n_step:
transition = self.memory_n.add(transition)
# add a single step transition
# if transition is not an empty tuple
if transition:
self.memory.add(transition)
def _get_critic_loss(
self, experiences: Tuple[torch.Tensor, ...], gamma: float
) -> torch.Tensor:
"""Return element-wise critic loss."""
states, actions, rewards, next_states, dones = experiences[:5]
# G_t = r + gamma * v(s_{t+1}) if state != Terminal
# = r otherwise
masks = 1 - dones
next_actions = self.actor_target(next_states)
next_states_actions = torch.cat((next_states, next_actions), dim=-1)
next_values = self.critic_target(next_states_actions)
curr_returns = rewards + gamma * next_values * masks
curr_returns = curr_returns.to(device).detach()
# train critic
values = self.critic(torch.cat((states, actions), dim=-1))
critic_loss_element_wise = (values - curr_returns).pow(2)
return critic_loss_element_wise
def update_model(self) -> Tuple[torch.Tensor, torch.Tensor]:
"""Train the model after each episode."""
experiences_1 = self.memory.sample(self.beta)
states, actions = experiences_1[:2]
weights, indices, eps_d = experiences_1[-3:]
gamma = self.hyper_params["GAMMA"]
# train critic
gradient_clip_cr = self.hyper_params["GRADIENT_CLIP_CR"]
critic_loss_element_wise = self._get_critic_loss(experiences_1, gamma)
critic_loss = torch.mean(critic_loss_element_wise * weights)
if self.use_n_step:
experiences_n = self.memory_n.sample(indices)
gamma = gamma ** self.hyper_params["N_STEP"]
critic_loss_n_element_wise = self._get_critic_loss(experiences_n, gamma)
# to update loss and priorities
lambda1 = self.hyper_params["LAMBDA1"]
critic_loss_element_wise += critic_loss_n_element_wise * lambda1
critic_loss = torch.mean(critic_loss_element_wise * weights)
self.critic_optimizer.zero_grad()
critic_loss.backward()
nn.utils.clip_grad_norm_(self.critic.parameters(), gradient_clip_cr)
self.critic_optimizer.step()
# train actor
gradient_clip_ac = self.hyper_params["GRADIENT_CLIP_AC"]
actions = self.actor(states)
actor_loss_element_wise = -self.critic(torch.cat((states, actions), dim=-1))
actor_loss = torch.mean(actor_loss_element_wise * weights)
self.actor_optimizer.zero_grad()
actor_loss.backward()
nn.utils.clip_grad_norm_(self.actor.parameters(), gradient_clip_ac)
self.actor_optimizer.step()
# update target networks
tau = self.hyper_params["TAU"]
common_utils.soft_update(self.actor, self.actor_target, tau)
common_utils.soft_update(self.critic, self.critic_target, tau)
# update priorities
new_priorities = critic_loss_element_wise
new_priorities += self.hyper_params["LAMBDA3"] * actor_loss_element_wise.pow(2)
new_priorities += self.hyper_params["PER_EPS"]
new_priorities = new_priorities.data.cpu().numpy().squeeze()
new_priorities += eps_d
self.memory.update_priorities(indices, new_priorities)
# increase beta
fraction = min(float(self.i_episode) / self.args.episode_num, 1.0)
self.beta = self.beta + fraction * (1.0 - self.beta)
return actor_loss.item(), critic_loss.item()
def pretrain(self):
"""Pretraining steps."""
pretrain_loss = list()
print("[INFO] Pre-Train %d step." % self.hyper_params["PRETRAIN_STEP"])
for i_step in range(1, self.hyper_params["PRETRAIN_STEP"] + 1):
t_begin = time.time()
loss = self.update_model()
t_end = time.time()
pretrain_loss.append(loss) # for logging
# logging
if i_step == 1 or i_step % 100 == 0:
avg_loss = np.vstack(pretrain_loss).mean(axis=0)
pretrain_loss.clear()
self.write_log(0, avg_loss, 0, t_end - t_begin)
print("[INFO] Pre-Train Complete!\n")
class Agent(AbstractAgent):
"""SAC agent interacting with environment.
Attrtibutes:
memory (ReplayBuffer): replay memory
actor (nn.Module): actor model to select actions
actor_target (nn.Module): target actor model to select actions
actor_optimizer (Optimizer): optimizer for training actor
critic_1 (nn.Module): critic model to predict state values
critic_2 (nn.Module): critic model to predict state values
critic_target1 (nn.Module): target critic model to predict state values
critic_target2 (nn.Module): target critic model to predict state values
critic_optimizer1 (Optimizer): optimizer for training critic_1
critic_optimizer2 (Optimizer): optimizer for training critic_2
curr_state (np.ndarray): temporary storage of the current state
target_entropy (int): desired entropy used for the inequality constraint
alpha (torch.Tensor): weight for entropy
alpha_optimizer (Optimizer): optimizer for alpha
hyper_params (dict): hyper-parameters
total_step (int): total step numbers
episode_step (int): step number of the current episode
i_episode (int): current episode number
her (HER): hinsight experience replay
"""
def __init__(self, env, args, hyper_params, models, optims, target_entropy,
her):
"""Initialization.
Args:
env (gym.Env): openAI Gym environment
args (argparse.Namespace): arguments including hyperparameters and training settings
hyper_params (dict): hyper-parameters
models (tuple): models including actor and critic
optims (tuple): optimizers for actor and critic
target_entropy (float): target entropy for the inequality constraint
her (HER): hinsight experience replay
"""
AbstractAgent.__init__(self, env, args)
self.actor, self.vf, self.vf_target, self.qf_1, self.qf_2 = models
self.actor_optimizer, self.vf_optimizer = optims[0:2]
self.qf_1_optimizer, self.qf_2_optimizer = optims[2:4]
self.hyper_params = hyper_params
self.curr_state = np.zeros((1, ))
self.total_step = 0
self.episode_step = 0
self.i_episode = 0
self.her = her
# automatic entropy tuning
if self.hyper_params["AUTO_ENTROPY_TUNING"]:
self.target_entropy = target_entropy
self.log_alpha = torch.zeros(1, requires_grad=True, device=device)
self.alpha_optimizer = optim.Adam(
[self.log_alpha], lr=self.hyper_params["LR_ENTROPY"])
# load the optimizer and model parameters
if args.load_from is not None and os.path.exists(args.load_from):
self.load_params(args.load_from)
self._initialize()
def _initialize(self):
"""Initialize non-common things."""
if not self.args.test:
# replay memory
self.memory = ReplayBuffer(self.hyper_params["BUFFER_SIZE"],
self.hyper_params["BATCH_SIZE"])
# HER
if self.hyper_params["USE_HER"]:
# load demo replay memory
with open(self.args.demo_path, "rb") as f:
demo = pickle.load(f)
if self.hyper_params["DESIRED_STATES_FROM_DEMO"]:
self.her.fetch_desired_states_from_demo(demo)
self.transitions_epi = list()
self.desired_state = np.zeros((1, ))
demo = self.her.generate_demo_transitions(demo)
if not self.args.test:
# Replay buffers
self.memory = ReplayBuffer(self.hyper_params["BUFFER_SIZE"],
self.hyper_params["BATCH_SIZE"])
def _preprocess_state(self, state):
"""Preprocess state so that actor selects an action."""
if self.hyper_params["USE_HER"]:
self.desired_state = self.her.get_desired_state()
state = np.concatenate((state, self.desired_state), axis=-1)
state = torch.FloatTensor(state).to(device)
return state
def _add_transition_to_memory(self, transition):
"""Add 1 step and n step transitions to memory."""
if self.hyper_params["USE_HER"]:
self.transitions_epi.append(transition)
done = transition[
-1] or self.episode_step == self.args.max_episode_steps
if done:
# insert generated transitions if the episode is done
transitions = self.her.generate_transitions(
self.transitions_epi,
self.desired_state,
self.hyper_params["SUCCESS_SCORE"],
)
self.memory.extend(transitions)
self.transitions_epi = list()
else:
self.memory.add(*transition)
def select_action(self, state):
"""Select an action from the input space."""
self.curr_state = state
state = self._preprocess_state(state)
# if initial random action should be conducted
if (self.total_step < self.hyper_params["INITIAL_RANDOM_ACTION"]
and not self.args.test):
return self.env.action_space.sample()
if self.args.test:
_, _, _, selected_action, _ = self.actor(state)
else:
selected_action, _, _, _, _ = self.actor(state)
return selected_action.detach().cpu().numpy()
def step(self, action):
"""Take an action and return the response of the env."""
self.total_step += 1
self.episode_step += 1
next_state, reward, done, _ = self.env.step(action)
if not self.args.test:
# if the last state is not a terminal state, store done as false
done_bool = (False if self.episode_step
== self.args.max_episode_steps else done)
transition = (self.curr_state, action, reward, next_state,
done_bool)
self._add_transition_to_memory(transition)
return next_state, reward, done
def update_model(self, experiences):
"""Train the model after each episode."""
states, actions, rewards, next_states, dones = experiences
new_actions, log_prob, pre_tanh_value, mu, std = self.actor(states)
# train alpha
if self.hyper_params["AUTO_ENTROPY_TUNING"]:
alpha_loss = (-self.log_alpha *
(log_prob + self.target_entropy).detach()).mean()
self.alpha_optimizer.zero_grad()
alpha_loss.backward()
self.alpha_optimizer.step()
alpha = self.log_alpha.exp()
else:
alpha_loss = torch.zeros(1)
alpha = self.hyper_params["W_ENTROPY"]
# Q function loss
masks = 1 - dones
q_1_pred = self.qf_1(states, actions)
q_2_pred = self.qf_2(states, actions)
v_target = self.vf_target(next_states)
q_target = rewards + self.hyper_params["GAMMA"] * v_target * masks
qf_1_loss = F.mse_loss(q_1_pred, q_target.detach())
qf_2_loss = F.mse_loss(q_2_pred, q_target.detach())
# V function loss
v_pred = self.vf(states)
q_pred = torch.min(self.qf_1(states, new_actions),
self.qf_2(states, new_actions))
v_target = q_pred - alpha * log_prob
vf_loss = F.mse_loss(v_pred, v_target.detach())
# train Q functions
self.qf_1_optimizer.zero_grad()
qf_1_loss.backward()
self.qf_1_optimizer.step()
self.qf_2_optimizer.zero_grad()
qf_2_loss.backward()
self.qf_2_optimizer.step()
# train V function
self.vf_optimizer.zero_grad()
vf_loss.backward()
self.vf_optimizer.step()
if self.total_step % self.hyper_params["DELAYED_UPDATE"] == 0:
# actor loss
advantage = q_pred - v_pred.detach()
actor_loss = (alpha * log_prob - advantage).mean()
# regularization
mean_reg = self.hyper_params["W_MEAN_REG"] * mu.pow(2).mean()
std_reg = self.hyper_params["W_STD_REG"] * std.pow(2).mean()
pre_activation_reg = self.hyper_params["W_PRE_ACTIVATION_REG"] * (
pre_tanh_value.pow(2).sum(dim=-1).mean())
actor_reg = mean_reg + std_reg + pre_activation_reg
# actor loss + regularization
actor_loss += actor_reg
# train actor
self.actor_optimizer.zero_grad()
actor_loss.backward()
self.actor_optimizer.step()
# update target networks
common_utils.soft_update(self.vf, self.vf_target,
self.hyper_params["TAU"])
else:
actor_loss = torch.zeros(1)
return (
actor_loss.data,
qf_1_loss.data,
qf_2_loss.data,
vf_loss.data,
alpha_loss.data,
)
def load_params(self, path):
"""Load model and optimizer parameters."""
if not os.path.exists(path):
print("[ERROR] the input path does not exist. ->", path)
return
params = torch.load(path)
self.actor.load_state_dict(params["actor"])
self.qf_1.load_state_dict(params["qf_1"])
self.qf_2.load_state_dict(params["qf_2"])
self.vf.load_state_dict(params["vf"])
self.vf_target.load_state_dict(params["vf_target"])
self.actor_optimizer.load_state_dict(params["actor_optim"])
self.qf_1_optimizer.load_state_dict(params["qf_1_optim"])
self.qf_2_optimizer.load_state_dict(params["qf_2_optim"])
self.vf_optimizer.load_state_dict(params["vf_optim"])
if self.hyper_params["AUTO_ENTROPY_TUNING"]:
self.alpha_optimizer.load_state_dict(params["alpha_optim"])
print("[INFO] loaded the model and optimizer from", path)
def save_params(self, n_episode):
"""Save model and optimizer parameters."""
params = {
"actor": self.actor.state_dict(),
"qf_1": self.qf_1.state_dict(),
"qf_2": self.qf_2.state_dict(),
"vf": self.vf.state_dict(),
"vf_target": self.vf_target.state_dict(),
"actor_optim": self.actor_optimizer.state_dict(),
"qf_1_optim": self.qf_1_optimizer.state_dict(),
"qf_2_optim": self.qf_2_optimizer.state_dict(),
"vf_optim": self.vf_optimizer.state_dict(),
}
if self.hyper_params["AUTO_ENTROPY_TUNING"]:
params["alpha_optim"] = self.alpha_optimizer.state_dict()
AbstractAgent.save_params(self, params, n_episode)
def write_log(self, i, loss, score=0.0, delayed_update=1):
"""Write log about loss and score"""
total_loss = loss.sum()
print(
"[INFO] episode %d, episode_step %d, total step %d, total score: %d\n"
"total loss: %.3f actor_loss: %.3f qf_1_loss: %.3f qf_2_loss: %.3f "
"vf_loss: %.3f alpha_loss: %.3f\n" % (
i,
self.episode_step,
self.total_step,
score,
total_loss,
loss[0] * delayed_update, # actor loss
loss[1], # qf_1 loss
loss[2], # qf_2 loss
loss[3], # vf loss
loss[4], # alpha loss
))
if self.args.log:
wandb.log({
"score": score,
"total loss": total_loss,
"actor loss": loss[0] * delayed_update,
"qf_1 loss": loss[1],
"qf_2 loss": loss[2],
"vf loss": loss[3],
"alpha loss": loss[4],
})
def train(self):
"""Train the agent."""
# logger
if self.args.log:
wandb.init()
wandb.config.update(self.hyper_params)
wandb.config.update(vars(self.args))
wandb.watch([self.actor, self.vf, self.qf_1, self.qf_2],
log="parameters")
for self.i_episode in range(1, self.args.episode_num + 1):
state = self.env.reset()
done = False
score = 0
self.episode_step = 0
loss_episode = list()
while not done:
if self.args.render and self.i_episode >= self.args.render_after:
self.env.render()
action = self.select_action(state)
next_state, reward, done = self.step(action)
state = next_state
score += reward
# training
if len(self.memory) >= self.hyper_params["BATCH_SIZE"]:
experiences = self.memory.sample()
loss = self.update_model(experiences)
loss_episode.append(loss) # for logging
# logging
if loss_episode:
avg_loss = np.vstack(loss_episode).mean(axis=0)
self.write_log(self.i_episode, avg_loss, score,
self.hyper_params["DELAYED_UPDATE"])
if self.i_episode % self.args.save_period == 0:
self.save_params(self.i_episode)
# termination
self.env.close()
class Agent(AbstractAgent):
"""ActorCritic interacting with environment.
Attributes:
memory (ReplayBuffer): replay memory
noise (OUNoise): random noise for exploration
hyper_params (dict): hyper-parameters
actor (nn.Module): actor model to select actions
actor_target (nn.Module): target actor model to select actions
critic (nn.Module): critic model to predict state values
critic_target (nn.Module): target critic model to predict state values
actor_optimizer (Optimizer): optimizer for training actor
critic_optimizer (Optimizer): optimizer for training critic
curr_state (np.ndarray): temporary storage of the current state
"""
def __init__(
self,
env: gym.Env,
args: argparse.Namespace,
hyper_params: dict,
models: tuple,
optims: tuple,
noise: OUNoise,
):
"""Initialization.
Args:
env (gym.Env): openAI Gym environment with discrete action space
args (argparse.Namespace): arguments including hyperparameters and training settings
hyper_params (dict): hyper-parameters
models (tuple): models including actor and critic
optims (tuple): optimizers for actor and critic
noise (OUNoise): random noise for exploration
"""
AbstractAgent.__init__(self, env, args)
self.actor, self.actor_target, self.critic, self.critic_target = models
self.actor_optimizer, self.critic_optimizer = optims
self.hyper_params = hyper_params
self.curr_state = np.zeros((1, ))
self.noise = noise
# load the optimizer and model parameters
if args.load_from is not None and os.path.exists(args.load_from):
self.load_params(args.load_from)
# replay memory
self.memory = ReplayBuffer(hyper_params["BUFFER_SIZE"],
hyper_params["BATCH_SIZE"], self.args.seed)
def select_action(self, state: np.ndarray) -> torch.Tensor:
"""Select an action from the input space."""
self.curr_state = state
state = torch.FloatTensor(state).to(device)
selected_action = self.actor(state)
selected_action += torch.FloatTensor(self.noise.sample()).to(device)
selected_action = torch.clamp(selected_action, -1.0, 1.0)
return selected_action
def step(self,
action: torch.Tensor) -> Tuple[np.ndarray, np.float64, bool]:
"""Take an action and return the response of the env."""
action = action.detach().cpu().numpy()
next_state, reward, done, _ = self.env.step(action)
self.memory.add(self.curr_state, action, reward, next_state, done)
return next_state, reward, done
def update_model(
self,
experiences: Tuple[torch.Tensor, torch.Tensor, torch.Tensor,
torch.Tensor, torch.Tensor],
) -> Tuple[torch.Tensor, torch.Tensor]:
"""Train the model after each episode."""
states, actions, rewards, next_states, dones = experiences
# G_t = r + gamma * v(s_{t+1}) if state != Terminal
# = r otherwise
masks = 1 - dones
next_actions = self.actor_target(next_states)
next_values = self.critic_target(
torch.cat((next_states, next_actions), dim=-1))
curr_returns = rewards + self.hyper_params[
"GAMMA"] * next_values * masks
curr_returns = curr_returns.to(device)
# train critic
values = self.critic(torch.cat((states, actions), dim=-1))
critic_loss = F.mse_loss(values, curr_returns)
self.critic_optimizer.zero_grad()
critic_loss.backward()
self.critic_optimizer.step()
# train actor
actions = self.actor(states)
actor_loss = -self.critic(torch.cat((states, actions), dim=-1)).mean()
self.actor_optimizer.zero_grad()
actor_loss.backward()
self.actor_optimizer.step()
# update target networks
tau = self.hyper_params["TAU"]
common_utils.soft_update(self.actor, self.actor_target, tau)
common_utils.soft_update(self.critic, self.critic_target, tau)
return actor_loss.data, critic_loss.data
def load_params(self, path: str):
"""Load model and optimizer parameters."""
if not os.path.exists(path):
print("[ERROR] the input path does not exist. ->", path)
return
params = torch.load(path)
self.actor.load_state_dict(params["actor_state_dict"])
self.actor_target.load_state_dict(params["actor_target_state_dict"])
self.critic.load_state_dict(params["critic_state_dict"])
self.critic_target.load_state_dict(params["critic_target_state_dict"])
self.actor_optimizer.load_state_dict(params["actor_optim_state_dict"])
self.critic_optimizer.load_state_dict(
params["critic_optim_state_dict"])
print("[INFO] loaded the model and optimizer from", path)
def save_params(self, n_episode: int):
"""Save model and optimizer parameters."""
params = {
"actor_state_dict": self.actor.state_dict(),
"actor_target_state_dict": self.actor_target.state_dict(),
"critic_state_dict": self.critic.state_dict(),
"critic_target_state_dict": self.critic_target.state_dict(),
"actor_optim_state_dict": self.actor_optimizer.state_dict(),
"critic_optim_state_dict": self.critic_optimizer.state_dict(),
}
AbstractAgent.save_params(self, self.args.algo, params, n_episode)
def write_log(self, i: int, loss: np.ndarray, score: int):
"""Write log about loss and score"""
total_loss = loss.sum()
print("[INFO] episode %d total score: %d, total loss: %f\n"
"actor_loss: %.3f critic_loss: %.3f\n" %
(i, score, total_loss, loss[0], loss[1]
) # actor loss # critic loss
)
if self.args.log:
wandb.log({
"score": score,
"total loss": total_loss,
"actor loss": loss[0],
"critic loss": loss[1],
})
def train(self):
"""Train the agent."""
# logger
if self.args.log:
wandb.init()
wandb.config.update(self.hyper_params)
wandb.watch([self.actor, self.critic], log="parameters")
for i_episode in range(1, self.args.episode_num + 1):
state = self.env.reset()
done = False
score = 0
loss_episode = list()
while not done:
if self.args.render and i_episode >= self.args.render_after:
self.env.render()
action = self.select_action(state)
next_state, reward, done = self.step(action)
if len(self.memory) >= self.hyper_params["BATCH_SIZE"]:
experiences = self.memory.sample()
loss = self.update_model(experiences)
loss_episode.append(loss) # for logging
state = next_state
score += reward
# logging
if loss_episode:
avg_loss = np.vstack(loss_episode).mean(axis=0)
self.write_log(i_episode, avg_loss, score)
if i_episode % self.args.save_period == 0:
self.save_params(i_episode)
# termination
self.env.close()
class DQNAgent(Agent):
"""DQN interacting with environment.
Attribute:
memory (PrioritizedReplayBuffer): replay memory
dqn (nn.Module): actor model to select actions
dqn_target (nn.Module): target actor model to select actions
dqn_optimizer (Optimizer): optimizer for training actor
hyper_params (dict): hyper-parameters
beta (float): beta parameter for prioritized replay buffer
curr_state (np.ndarray): temporary storage of the current state
total_step (int): total step number
episode_step (int): step number of the current episode
epsilon (float): parameter for epsilon greedy policy
i_episode (int): current episode number
n_step_buffer (deque): n-size buffer to calculate n-step returns
use_n_step (bool): whether or not to use n-step returns
"""
def __init__(
self,
env: gym.Env,
args: argparse.Namespace,
hyper_params: dict,
models: tuple,
optim: torch.optim.Adam,
):
"""Initialization.
Args:
env (gym.Env): openAI Gym environment
args (argparse.Namespace): arguments including hyperparameters and training settings
hyper_params (dict): hyper-parameters
models (tuple): models including main network and target
optim (torch.optim.Adam): optimizers for dqn
"""
Agent.__init__(self, env, args)
self.use_n_step = hyper_params["N_STEP"] > 1
self.epsilon = hyper_params["MAX_EPSILON"]
self.dqn, self.dqn_target = models
self.hyper_params = hyper_params
self.curr_state = np.zeros(1)
self.dqn_optimizer = optim
self.episode_step = 0
self.total_step = 0
self.i_episode = 0
# load the optimizer and model parameters
if args.load_from is not None and os.path.exists(args.load_from):
self.load_params(args.load_from)
self._initialize()
# pylint: disable=attribute-defined-outside-init
def _initialize(self):
"""Initialize non-common things."""
if not self.args.test:
# replay memory for a single step
self.beta = self.hyper_params["PER_BETA"]
self.memory = PrioritizedReplayBuffer(
self.hyper_params["BUFFER_SIZE"],
self.hyper_params["BATCH_SIZE"],
alpha=self.hyper_params["PER_ALPHA"],
)
# replay memory for multi-steps
if self.use_n_step:
self.memory_n = ReplayBuffer(
self.hyper_params["BUFFER_SIZE"],
n_step=self.hyper_params["N_STEP"],
gamma=self.hyper_params["GAMMA"],
)
def select_action(self, state: np.ndarray) -> np.ndarray:
"""Select an action from the input space."""
self.curr_state = state
# epsilon greedy policy
# pylint: disable=comparison-with-callable
if not self.args.test and self.epsilon > np.random.random():
selected_action = np.array(self.env.action_space.sample())
else:
state = self._preprocess_state(state)
selected_action = self.dqn(state).argmax()
selected_action = selected_action.detach().cpu().numpy()
return selected_action
# pylint: disable=no-self-use
def _preprocess_state(self, state: np.ndarray) -> torch.Tensor:
"""Preprocess state so that actor selects an action."""
state = torch.FloatTensor(state).to(device)
return state
def step(self, action: np.ndarray) -> Tuple[np.ndarray, np.float64, bool, dict]:
"""Take an action and return the response of the env."""
next_state, reward, done, info = self.env.step(action)
if not self.args.test:
# if the last state is not a terminal state, store done as false
done_bool = (
False if self.episode_step == self.args.max_episode_steps else done
)
transition = (self.curr_state, action, reward, next_state, done_bool)
self._add_transition_to_memory(transition)
return next_state, reward, done, info
def _add_transition_to_memory(self, transition: Tuple[np.ndarray, ...]):
"""Add 1 step and n step transitions to memory."""
# add n-step transition
if self.use_n_step:
transition = self.memory_n.add(transition)
# add a single step transition
# if transition is not an empty tuple
if transition:
self.memory.add(transition)
def _get_dqn_loss(
self, experiences: Tuple[torch.Tensor, ...], gamma: float
) -> Tuple[torch.Tensor, torch.Tensor]:
"""Return element-wise dqn loss and Q-values."""
if self.hyper_params["USE_DIST_Q"] == "IQN":
return dqn_utils.calculate_iqn_loss(
model=self.dqn,
target_model=self.dqn_target,
experiences=experiences,
gamma=gamma,
batch_size=self.hyper_params["BATCH_SIZE"],
n_tau_samples=self.hyper_params["N_TAU_SAMPLES"],
n_tau_prime_samples=self.hyper_params["N_TAU_PRIME_SAMPLES"],
kappa=self.hyper_params["KAPPA"],
)
elif self.hyper_params["USE_DIST_Q"] == "C51":
return dqn_utils.calculate_c51_loss(
model=self.dqn,
target_model=self.dqn_target,
experiences=experiences,
gamma=gamma,
batch_size=self.hyper_params["BATCH_SIZE"],
v_min=self.hyper_params["V_MIN"],
v_max=self.hyper_params["V_MAX"],
atom_size=self.hyper_params["ATOMS"],
)
else:
return dqn_utils.calculate_dqn_loss(
model=self.dqn,
target_model=self.dqn_target,
experiences=experiences,
gamma=gamma,
)
def update_model(self) -> Tuple[torch.Tensor, torch.Tensor]:
"""Train the model after each episode."""
# 1 step loss
experiences_1 = self.memory.sample(self.beta)
weights, indices = experiences_1[-3:-1]
gamma = self.hyper_params["GAMMA"]
dq_loss_element_wise, q_values = self._get_dqn_loss(experiences_1, gamma)
dq_loss = torch.mean(dq_loss_element_wise * weights)
# n step loss
if self.use_n_step:
experiences_n = self.memory_n.sample(indices)
gamma = self.hyper_params["GAMMA"] ** self.hyper_params["N_STEP"]
dq_loss_n_element_wise, q_values_n = self._get_dqn_loss(
experiences_n, gamma
)
# to update loss and priorities
q_values = 0.5 * (q_values + q_values_n)
dq_loss_element_wise += (
dq_loss_n_element_wise * self.hyper_params["W_N_STEP"]
)
dq_loss = torch.mean(dq_loss_element_wise * weights)
# q_value regularization
q_regular = torch.norm(q_values, 2).mean() * self.hyper_params["W_Q_REG"]
# total loss
loss = dq_loss + q_regular
self.dqn_optimizer.zero_grad()
loss.backward()
clip_grad_norm_(self.dqn.parameters(), self.hyper_params["GRADIENT_CLIP"])
self.dqn_optimizer.step()
# update target networks
tau = self.hyper_params["TAU"]
common_utils.soft_update(self.dqn, self.dqn_target, tau)
# update priorities in PER
loss_for_prior = dq_loss_element_wise.detach().cpu().numpy()
new_priorities = loss_for_prior + self.hyper_params["PER_EPS"]
self.memory.update_priorities(indices, new_priorities)
# increase beta
fraction = min(float(self.i_episode) / self.args.episode_num, 1.0)
self.beta = self.beta + fraction * (1.0 - self.beta)
if self.hyper_params["USE_NOISY_NET"]:
self.dqn.reset_noise()
self.dqn_target.reset_noise()
return loss.item(), q_values.mean().item()
def load_params(self, path: str):
"""Load model and optimizer parameters."""
if not os.path.exists(path):
print("[ERROR] the input path does not exist. ->", path)
return
params = torch.load(path)
self.dqn.load_state_dict(params["dqn_state_dict"])
self.dqn_target.load_state_dict(params["dqn_target_state_dict"])
self.dqn_optimizer.load_state_dict(params["dqn_optim_state_dict"])
print("[INFO] loaded the model and optimizer from", path)
def save_params(self, n_episode: int):
"""Save model and optimizer parameters."""
params = {
"dqn_state_dict": self.dqn.state_dict(),
"dqn_target_state_dict": self.dqn_target.state_dict(),
"dqn_optim_state_dict": self.dqn_optimizer.state_dict(),
}
Agent.save_params(self, params, n_episode)
def write_log(self, i: int, loss: np.ndarray, score: float, avg_time_cost: float):
"""Write log about loss and score"""
print(
"[INFO] episode %d, episode step: %d, total step: %d, total score: %f\n"
"epsilon: %f, loss: %f, avg q-value: %f (spent %.6f sec/step)\n"
% (
i,
self.episode_step,
self.total_step,
score,
self.epsilon,
loss[0],
loss[1],
avg_time_cost,
)
)
if self.args.log:
wandb.log(
{
"score": score,
"epsilon": self.epsilon,
"dqn loss": loss[0],
"avg q values": loss[1],
"time per each step": avg_time_cost,
}
)
# pylint: disable=no-self-use, unnecessary-pass
def pretrain(self):
"""Pretraining steps."""
pass
def train(self):
"""Train the agent."""
# logger
if self.args.log:
wandb.init()
wandb.config.update(self.hyper_params)
# wandb.watch([self.dqn], log="parameters")
# pre-training if needed
self.pretrain()
max_epsilon, min_epsilon, epsilon_decay = (
self.hyper_params["MAX_EPSILON"],
self.hyper_params["MIN_EPSILON"],
self.hyper_params["EPSILON_DECAY"],
)
for self.i_episode in range(1, self.args.episode_num + 1):
state = self.env.reset()
self.episode_step = 0
losses = list()
done = False
score = 0
t_begin = time.time()
while not done:
if self.args.render and self.i_episode >= self.args.render_after:
self.env.render()
action = self.select_action(state)
next_state, reward, done, _ = self.step(action)
self.total_step += 1
self.episode_step += 1
if len(self.memory) >= self.hyper_params["UPDATE_STARTS_FROM"]:
if self.total_step % self.hyper_params["TRAIN_FREQ"] == 0:
for _ in range(self.hyper_params["MULTIPLE_LEARN"]):
loss = self.update_model()
losses.append(loss) # for logging
# decrease epsilon
self.epsilon = max(
self.epsilon - (max_epsilon - min_epsilon) * epsilon_decay,
min_epsilon,
)
state = next_state
score += reward
t_end = time.time()
avg_time_cost = (t_end - t_begin) / self.episode_step
if losses:
avg_loss = np.vstack(losses).mean(axis=0)
self.write_log(self.i_episode, avg_loss, score, avg_time_cost)
if self.i_episode % self.args.save_period == 0:
self.save_params(self.i_episode)
self.interim_test()
# termination
self.env.close()
self.save_params(self.i_episode)
self.interim_test()
class TD3Agent(Agent):
"""ActorCritic interacting with environment.
Attributes:
memory (ReplayBuffer): replay memory
exploration_noise (GaussianNoise): random noise for exploration
target_policy_noise (GaussianNoise): random noise for target values
actor (nn.Module): actor model to select actions
critic1 (nn.Module): critic model to predict state values
critic2 (nn.Module): critic model to predict state values
critic_target1 (nn.Module): target critic model to predict state values
critic_target2 (nn.Module): target critic model to predict state values
actor_target (nn.Module): target actor model to select actions
critic_optim (Optimizer): optimizer for training critic
actor_optim (Optimizer): optimizer for training actor
hyper_params (dict): hyper-parameters
curr_state (np.ndarray): temporary storage of the current state
total_steps (int): total step numbers
episode_steps (int): step number of the current episode
"""
def __init__(
self,
env: gym.Env,
args: argparse.Namespace,
hyper_params: dict,
models: tuple,
optims: tuple,
exploration_noise: GaussianNoise,
target_policy_noise: GaussianNoise,
):
"""Initialization.
Args:
env (gym.Env): openAI Gym environment
args (argparse.Namespace): arguments including hyperparameters and training settings
hyper_params (dict): hyper-parameters
models (tuple): models including actor and critic
optims (tuple): optimizers for actor and critic
exploration_noise (GaussianNoise): random noise for exploration
target_policy_noise (GaussianNoise): random noise for target values
"""
Agent.__init__(self, env, args)
self.actor, self.actor_target = models[0:2]
self.critic1, self.critic2 = models[2:4]
self.critic_target1, self.critic_target2 = models[4:6]
self.actor_optim = optims[0]
self.critic_optim = optims[1]
self.hyper_params = hyper_params
self.curr_state = np.zeros((1, ))
self.exploration_noise = exploration_noise
self.target_policy_noise = target_policy_noise
self.total_step = 0
self.episode_step = 0
self.update_step = 0
self.i_episode = 0
# load the optimizer and model parameters
if args.load_from is not None and os.path.exists(args.load_from):
self.load_params(args.load_from)
if not self.args.test:
# replay memory
self.memory = ReplayBuffer(hyper_params["BUFFER_SIZE"],
hyper_params["BATCH_SIZE"])
def select_action(self, state: np.ndarray) -> np.ndarray:
"""Select an action from the input space."""
# initial training step, try random action for exploration
random_action_count = self.hyper_params["INITIAL_RANDOM_ACTION"]
self.curr_state = state
if self.total_step < random_action_count and not self.args.test:
return self.env.action_space.sample()
state = torch.FloatTensor(state).to(device)
selected_action = self.actor(state).detach().cpu().numpy()
if not self.args.test:
noise = self.exploration_noise.sample()
selected_action = np.clip(selected_action + noise, -1.0, 1.0)
return selected_action
def step(self,
action: np.ndarray) -> Tuple[np.ndarray, np.float64, bool, dict]:
"""Take an action and return the response of the env."""
next_state, reward, done, info = self.env.step(action)
if not self.args.test:
# if last state is not terminal state in episode, done is false
done_bool = (False if self.episode_step
== self.args.max_episode_steps else done)
self.memory.add(self.curr_state, action, reward, next_state,
done_bool)
return next_state, reward, done, info
def update_model(
self, experiences: Tuple[torch.Tensor, ...]
) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
"""Train the model after each episode."""
self.update_step += 1
states, actions, rewards, next_states, dones = experiences
masks = 1 - dones
# get actions with noise
noise = torch.FloatTensor(self.target_policy_noise.sample()).to(device)
clipped_noise = torch.clamp(
noise,
-self.hyper_params["TARGET_POLICY_NOISE_CLIP"],
self.hyper_params["TARGET_POLICY_NOISE_CLIP"],
)
next_actions = (self.actor_target(next_states) + clipped_noise).clamp(
-1.0, 1.0)
# min (Q_1', Q_2')
next_values1 = self.critic_target1(next_states, next_actions)
next_values2 = self.critic_target2(next_states, next_actions)
next_values = torch.min(next_values1, next_values2)
# G_t = r + gamma * v(s_{t+1}) if state != Terminal
# = r otherwise
curr_returns = rewards + self.hyper_params[
"GAMMA"] * next_values * masks
curr_returns = curr_returns.detach()
# critic loss
values1 = self.critic1(states, actions)
values2 = self.critic2(states, actions)
critic1_loss = F.mse_loss(values1, curr_returns)
critic2_loss = F.mse_loss(values2, curr_returns)
# train critic
critic_loss = critic1_loss + critic2_loss
self.critic_optim.zero_grad()
critic_loss.backward()
self.critic_optim.step()
if self.update_step % self.hyper_params["POLICY_UPDATE_FREQ"] == 0:
# policy loss
actions = self.actor(states)
actor_loss = -self.critic1(states, actions).mean()
# train actor
self.actor_optim.zero_grad()
actor_loss.backward()
self.actor_optim.step()
# update target networks
tau = self.hyper_params["TAU"]
common_utils.soft_update(self.critic1, self.critic_target1, tau)
common_utils.soft_update(self.critic2, self.critic_target2, tau)
common_utils.soft_update(self.actor, self.actor_target, tau)
else:
actor_loss = torch.zeros(1)
return actor_loss.item(), critic1_loss.item(), critic2_loss.item()
def load_params(self, path: str):
"""Load model and optimizer parameters."""
if not os.path.exists(path):
print("[ERROR] the input path does not exist. ->", path)
return
params = torch.load(path)
self.critic1.load_state_dict(params["critic1"])
self.critic2.load_state_dict(params["critic2"])
self.critic_target1.load_state_dict(params["critic_target1"])
self.critic_target2.load_state_dict(params["critic_target2"])
self.critic_optim.load_state_dict(params["critic_optim"])
self.actor.load_state_dict(params["actor"])
self.actor_target.load_state_dict(params["actor_target"])
self.actor_optim.load_state_dict(params["actor_optim"])
print("[INFO] loaded the model and optimizer from", path)
def save_params(self, n_episode: int):
"""Save model and optimizer parameters."""
params = {
"actor": self.actor.state_dict(),
"actor_target": self.actor_target.state_dict(),
"actor_optim": self.actor_optim.state_dict(),
"critic1": self.critic1.state_dict(),
"critic2": self.critic2.state_dict(),
"critic_target1": self.critic_target1.state_dict(),
"critic_target2": self.critic_target2.state_dict(),
"critic_optim": self.critic_optim.state_dict(),
}
Agent.save_params(self, params, n_episode)
def write_log(self,
i: int,
loss: np.ndarray,
score: float = 0.0,
policy_update_freq: int = 1):
"""Write log about loss and score"""
total_loss = loss.sum()
print(
"[INFO] episode %d, episode_step: %d, total_step: %d, total score: %d\n"
"total loss: %f actor_loss: %.3f critic1_loss: %.3f critic2_loss: %.3f\n"
% (
i,
self.episode_step,
self.total_step,
score,
total_loss,
loss[0] * policy_update_freq, # actor loss
loss[1], # critic1 loss
loss[2], # critic2 loss
))
if self.args.log:
wandb.log({
"score": score,
"total loss": total_loss,
"actor loss": loss[0] * policy_update_freq,
"critic1 loss": loss[1],
"critic2 loss": loss[2],
})
def train(self):
"""Train the agent."""
# logger
if self.args.log:
wandb.init(project=self.args.wandb_project)
wandb.config.update(self.hyper_params)
# wandb.watch([self.actor, self.critic1, self.critic2], log="parameters")
for self.i_episode in range(1, self.args.episode_num + 1):
state = self.env.reset()
done = False
score = 0
loss_episode = list()
self.episode_step = 0
while not done:
if self.args.render and self.i_episode >= self.args.render_after:
self.env.render()
action = self.select_action(state)
next_state, reward, done, _ = self.step(action)
self.total_step += 1
self.episode_step += 1
state = next_state
score += reward
if len(self.memory) >= self.hyper_params["BATCH_SIZE"]:
experiences = self.memory.sample()
loss = self.update_model(experiences)
loss_episode.append(loss) # for logging
# logging
if loss_episode:
avg_loss = np.vstack(loss_episode).mean(axis=0)
self.write_log(
self.i_episode,
avg_loss,
score,
self.hyper_params["POLICY_UPDATE_FREQ"],
)
if self.i_episode % self.args.save_period == 0:
self.save_params(self.i_episode)
self.interim_test()
# termination
self.env.close()
self.save_params(self.i_episode)
self.interim_test()
class DQfDAgent(DQNAgent):
"""DQN interacting with environment.
Attribute:
memory (PrioritizedReplayBuffer): replay memory
"""
# pylint: disable=attribute-defined-outside-init
def _initialize(self):
"""Initialize non-common things."""
if not self.args.test:
# load demo replay memory
demos = self._load_demos()
if self.use_n_step:
demos, demos_n_step = common_utils.get_n_step_info_from_demo(
demos, self.hyper_params["N_STEP"],
self.hyper_params["GAMMA"])
self.memory_n = ReplayBuffer(
buffer_size=self.hyper_params["BUFFER_SIZE"],
n_step=self.hyper_params["N_STEP"],
gamma=self.hyper_params["GAMMA"],
demo=demos_n_step,
)
# replay memory
self.beta = self.hyper_params["PER_BETA"]
self.memory = PrioritizedReplayBuffer(
self.hyper_params["BUFFER_SIZE"],
self.hyper_params["BATCH_SIZE"],
demo=demos,
alpha=self.hyper_params["PER_ALPHA"],
epsilon_d=self.hyper_params["PER_EPS_DEMO"],
)
def _load_demos(self) -> list:
"""Load expert's demonstrations."""
# load demo replay memory
with open(self.args.demo_path, "rb") as f:
demos = pickle.load(f)
return demos
def update_model(self) -> Tuple[torch.Tensor, ...]:
"""Train the model after each episode."""
experiences_1 = self.memory.sample()
weights, indices, eps_d = experiences_1[-3:]
actions = experiences_1[1]
# 1 step loss
gamma = self.hyper_params["GAMMA"]
dq_loss_element_wise, q_values = self._get_dqn_loss(
experiences_1, gamma)
dq_loss = torch.mean(dq_loss_element_wise * weights)
# n step loss
if self.use_n_step:
experiences_n = self.memory_n.sample(indices)
gamma = self.hyper_params["GAMMA"]**self.hyper_params["N_STEP"]
dq_loss_n_element_wise, q_values_n = self._get_dqn_loss(
experiences_n, gamma)
# to update loss and priorities
q_values = 0.5 * (q_values + q_values_n)
dq_loss_element_wise += (dq_loss_n_element_wise *
self.hyper_params["LAMBDA1"])
dq_loss = torch.mean(dq_loss_element_wise * weights)
# supervised loss using demo for only demo transitions
demo_idxs = np.where(eps_d != 0.0)
n_demo = demo_idxs[0].size
if n_demo != 0: # if 1 or more demos are sampled
# get margin for each demo transition
action_idxs = actions[demo_idxs].long()
margin = torch.ones(q_values.size()) * self.hyper_params["MARGIN"]
margin[demo_idxs, action_idxs] = 0.0 # demo actions have 0 margins
margin = margin.to(device)
# calculate supervised loss
demo_q_values = q_values[demo_idxs, action_idxs].squeeze()
supervised_loss = torch.max(q_values + margin, dim=-1)[0]
supervised_loss = supervised_loss[demo_idxs] - demo_q_values
supervised_loss = torch.mean(
supervised_loss) * self.hyper_params["LAMBDA2"]
else: # no demo sampled
supervised_loss = torch.zeros(1, device=device)
# q_value regularization
q_regular = torch.norm(q_values,
2).mean() * self.hyper_params["W_Q_REG"]
# total loss
loss = dq_loss + supervised_loss + q_regular
# train dqn
self.dqn_optimizer.zero_grad()
loss.backward()
clip_grad_norm_(self.dqn.parameters(),
self.hyper_params["GRADIENT_CLIP"])
self.dqn_optimizer.step()
# update target networks
tau = self.hyper_params["TAU"]
common_utils.soft_update(self.dqn, self.dqn_target, tau)
# update priorities in PER
loss_for_prior = dq_loss_element_wise.detach().cpu().numpy().squeeze()
new_priorities = loss_for_prior + self.hyper_params["PER_EPS"]
new_priorities += eps_d
self.memory.update_priorities(indices, new_priorities)
# increase beta
fraction = min(float(self.i_episode) / self.args.episode_num, 1.0)
self.beta = self.beta + fraction * (1.0 - self.beta)
if self.hyper_params["USE_NOISY_NET"]:
self.dqn.reset_noise()
self.dqn_target.reset_noise()
return (
loss.item(),
dq_loss.item(),
supervised_loss.item(),
q_values.mean().item(),
n_demo,
)
def write_log(self, i: int, avg_loss: np.ndarray, score: float,
avg_time_cost: float):
"""Write log about loss and score"""
print(
"[INFO] episode %d, episode step: %d, total step: %d, total score: %f\n"
"epsilon: %f, total loss: %f, dq loss: %f, supervised loss: %f\n"
"avg q values: %f, demo num in minibatch: %d (spent %.6f sec/step)\n"
% (
i,
self.episode_step,
self.total_step,
score,
self.epsilon,
avg_loss[0],
avg_loss[1],
avg_loss[2],
avg_loss[3],
avg_loss[4],
avg_time_cost,
))
if self.args.log:
wandb.log({
"score": score,
"epsilon": self.epsilon,
"total loss": avg_loss[0],
"dq loss": avg_loss[1],
"supervised loss": avg_loss[2],
"avg q values": avg_loss[3],
"demo num in minibatch": avg_loss[4],
"time per each step": avg_time_cost,
})
def pretrain(self):
"""Pretraining steps."""
pretrain_loss = list()
print("[INFO] Pre-Train %d step." % self.hyper_params["PRETRAIN_STEP"])
for i_step in range(1, self.hyper_params["PRETRAIN_STEP"] + 1):
t_begin = time.time()
loss = self.update_model()
t_end = time.time()
pretrain_loss.append(loss) # for logging
# logging
if i_step == 1 or i_step % 100 == 0:
avg_loss = np.vstack(pretrain_loss).mean(axis=0)
pretrain_loss.clear()
self.write_log(0, avg_loss, 0.0, t_end - t_begin)
print("[INFO] Pre-Train Complete!\n")
class SACfDAgent(SACAgent):
"""SAC agent interacting with environment.
Attrtibutes:
memory (PrioritizedReplayBuffer): replay memory
beta (float): beta parameter for prioritized replay buffer
"""
# pylint: disable=attribute-defined-outside-init
def _initialize(self):
"""Initialize non-common things."""
self.use_n_step = self.hyper_params["N_STEP"] > 1
if not self.args.test:
# load demo replay memory
with open(self.args.demo_path, "rb") as f:
demos = pickle.load(f)
if self.use_n_step:
demos, demos_n_step = common_utils.get_n_step_info_from_demo(
demos, self.hyper_params["N_STEP"],
self.hyper_params["GAMMA"])
# replay memory for multi-steps
self.memory_n = ReplayBuffer(
buffer_size=self.hyper_params["BUFFER_SIZE"],
n_step=self.hyper_params["N_STEP"],
gamma=self.hyper_params["GAMMA"],
demo=demos_n_step,
)
# replay memory
self.beta = self.hyper_params["PER_BETA"]
self.memory = PrioritizedReplayBuffer(
self.hyper_params["BUFFER_SIZE"],
self.hyper_params["BATCH_SIZE"],
demo=demos,
alpha=self.hyper_params["PER_ALPHA"],
epsilon_d=self.hyper_params["PER_EPS_DEMO"],
)
def _add_transition_to_memory(self, transition: Tuple[np.ndarray, ...]):
"""Add 1 step and n step transitions to memory."""
# add n-step transition
if self.use_n_step:
transition = self.memory_n.add(transition)
# add a single step transition
# if transition is not an empty tuple
if transition:
self.memory.add(transition)
# pylint: disable=too-many-statements
def update_model(self) -> Tuple[torch.Tensor, ...]:
"""Train the model after each episode."""
self.update_step += 1
experiences = self.memory.sample(self.beta)
states, actions, rewards, next_states, dones, weights, indices, eps_d = (
experiences)
new_actions, log_prob, pre_tanh_value, mu, std = self.actor(states)
# train alpha
if self.hyper_params["AUTO_ENTROPY_TUNING"]:
alpha_loss = torch.mean(
(-self.log_alpha *
(log_prob + self.target_entropy).detach()) * weights)
self.alpha_optimizer.zero_grad()
alpha_loss.backward()
self.alpha_optimizer.step()
alpha = self.log_alpha.exp()
else:
alpha_loss = torch.zeros(1)
alpha = self.hyper_params["W_ENTROPY"]
# Q function loss
masks = 1 - dones
gamma = self.hyper_params["GAMMA"]
q_1_pred = self.qf_1(states, actions)
q_2_pred = self.qf_2(states, actions)
v_target = self.vf_target(next_states)
q_target = rewards + self.hyper_params["GAMMA"] * v_target * masks
qf_1_loss = torch.mean((q_1_pred - q_target.detach()).pow(2) * weights)
qf_2_loss = torch.mean((q_2_pred - q_target.detach()).pow(2) * weights)
if self.use_n_step:
experiences_n = self.memory_n.sample(indices)
_, _, rewards, next_states, dones = experiences_n
gamma = gamma**self.hyper_params["N_STEP"]
lambda1 = self.hyper_params["LAMBDA1"]
masks = 1 - dones
v_target = self.vf_target(next_states)
q_target = rewards + gamma * v_target * masks
qf_1_loss_n = torch.mean(
(q_1_pred - q_target.detach()).pow(2) * weights)
qf_2_loss_n = torch.mean(
(q_2_pred - q_target.detach()).pow(2) * weights)
# to update loss and priorities
qf_1_loss = qf_1_loss + qf_1_loss_n * lambda1
qf_2_loss = qf_2_loss + qf_2_loss_n * lambda1
# V function loss
v_pred = self.vf(states)
q_pred = torch.min(self.qf_1(states, new_actions),
self.qf_2(states, new_actions))
v_target = (q_pred - alpha * log_prob).detach()
vf_loss_element_wise = (v_pred - v_target).pow(2)
vf_loss = torch.mean(vf_loss_element_wise * weights)
# train Q functions
self.qf_1_optimizer.zero_grad()
qf_1_loss.backward()
self.qf_1_optimizer.step()
self.qf_2_optimizer.zero_grad()
qf_2_loss.backward()
self.qf_2_optimizer.step()
# train V function
self.vf_optimizer.zero_grad()
vf_loss.backward()
self.vf_optimizer.step()
if self.update_step % self.hyper_params["POLICY_UPDATE_FREQ"] == 0:
# actor loss
advantage = q_pred - v_pred.detach()
actor_loss_element_wise = alpha * log_prob - advantage
actor_loss = torch.mean(actor_loss_element_wise * weights)
# regularization
if not self.is_discrete: # iff the action is continuous
mean_reg = self.hyper_params["W_MEAN_REG"] * mu.pow(2).mean()
std_reg = self.hyper_params["W_STD_REG"] * std.pow(2).mean()
pre_activation_reg = self.hyper_params[
"W_PRE_ACTIVATION_REG"] * (pre_tanh_value.pow(2).sum(
dim=-1).mean())
actor_reg = mean_reg + std_reg + pre_activation_reg
# actor loss + regularization
actor_loss += actor_reg
# train actor
self.actor_optimizer.zero_grad()
actor_loss.backward()
self.actor_optimizer.step()
# update target networks
common_utils.soft_update(self.vf, self.vf_target,
self.hyper_params["TAU"])
# update priorities
new_priorities = vf_loss_element_wise
new_priorities += self.hyper_params[
"LAMBDA3"] * actor_loss_element_wise.pow(2)
new_priorities += self.hyper_params["PER_EPS"]
new_priorities = new_priorities.data.cpu().numpy().squeeze()
new_priorities += eps_d
self.memory.update_priorities(indices, new_priorities)
# increase beta
fraction = min(float(self.i_episode) / self.args.episode_num, 1.0)
self.beta = self.beta + fraction * (1.0 - self.beta)
else:
actor_loss = torch.zeros(1)
return (
actor_loss.item(),
qf_1_loss.item(),
qf_2_loss.item(),
vf_loss.item(),
alpha_loss.item(),
)
def pretrain(self):
"""Pretraining steps."""
pretrain_loss = list()
print("[INFO] Pre-Train %d steps." %
self.hyper_params["PRETRAIN_STEP"])
for i_step in range(1, self.hyper_params["PRETRAIN_STEP"] + 1):
t_begin = time.time()
loss = self.update_model()
t_end = time.time()
pretrain_loss.append(loss) # for logging
# logging
if i_step == 1 or i_step % 100 == 0:
avg_loss = np.vstack(pretrain_loss).mean(axis=0)
pretrain_loss.clear()
self.write_log(
0,
avg_loss,
0,
policy_update_freq=self.hyper_params["POLICY_UPDATE_FREQ"],
avg_time_cost=t_end - t_begin,
)
print("[INFO] Pre-Train Complete!\n")
class Agent(AbstractAgent):
"""ActorCritic interacting with environment.
Attributes:
memory (ReplayBuffer): replay memory
noise (OUNoise): random noise for exploration
hyper_params (dict): hyper-parameters
actor (nn.Module): actor model to select actions
actor_target (nn.Module): target actor model to select actions
critic (nn.Module): critic model to predict state values
critic_target (nn.Module): target critic model to predict state values
actor_optimizer (Optimizer): optimizer for training actor
critic_optimizer (Optimizer): optimizer for training critic
curr_state (np.ndarray): temporary storage of the current state
total_step (int): total step numbers
episode_step (int): step number of the current episode
i_episode (int): current episode number
"""
def __init__(
self,
env: gym.Env,
args: argparse.Namespace,
hyper_params: dict,
models: tuple,
optims: tuple,
noise: OUNoise,
):
"""Initialization.
Args:
env (gym.Env): openAI Gym environment
args (argparse.Namespace): arguments including hyperparameters and training settings
hyper_params (dict): hyper-parameters
models (tuple): models including actor and critic
optims (tuple): optimizers for actor and critic
noise (OUNoise): random noise for exploration
"""
AbstractAgent.__init__(self, env, args)
self.actor, self.actor_target, self.critic, self.critic_target = models
self.actor_optimizer, self.critic_optimizer = optims
self.hyper_params = hyper_params
self.curr_state = np.zeros((1, ))
self.noise = noise
self.total_step = 0
self.episode_step = 0
self.i_episode = 0
# load the optimizer and model parameters
if args.load_from is not None and os.path.exists(args.load_from):
self.load_params(args.load_from)
self._initialize()
def _initialize(self):
"""Initialize non-common things."""
if not self.args.test:
# replay memory
self.memory = ReplayBuffer(self.hyper_params["BUFFER_SIZE"],
self.hyper_params["BATCH_SIZE"])
def select_action(self, state: np.ndarray) -> np.ndarray:
"""Select an action from the input space."""
self.curr_state = state
state = self._preprocess_state(state)
# if initial random action should be conducted
if (self.total_step < self.hyper_params["INITIAL_RANDOM_ACTION"]
and not self.args.test):
return self.env.action_space.sample()
selected_action = self.actor(state)
if not self.args.test:
selected_action += torch.FloatTensor(
self.noise.sample()).to(device)
selected_action = torch.clamp(selected_action, -1.0, 1.0)
return selected_action.detach().cpu().numpy()
# pylint: disable=no-self-use
def _preprocess_state(self, state: np.ndarray) -> torch.Tensor:
"""Preprocess state so that actor selects an action."""
state = torch.FloatTensor(state).to(device)
return state
def step(self, action: np.ndarray) -> Tuple[torch.Tensor, ...]:
"""Take an action and return the response of the env."""
self.total_step += 1
self.episode_step += 1
next_state, reward, done, _ = self.env.step(action)
if not self.args.test:
# if the last state is not a terminal state, store done as false
done_bool = (False if self.episode_step
== self.args.max_episode_steps else done)
transition = (self.curr_state, action, reward, next_state,
done_bool)
self._add_transition_to_memory(transition)
return next_state, reward, done
def _add_transition_to_memory(self, transition: Tuple[np.ndarray, ...]):
"""Add 1 step and n step transitions to memory."""
self.memory.add(*transition)
def update_model(self) -> Tuple[torch.Tensor, torch.Tensor]:
"""Train the model after each episode."""
experiences = self.memory.sample()
states, actions, rewards, next_states, dones = experiences
# G_t = r + gamma * v(s_{t+1}) if state != Terminal
# = r otherwise
masks = 1 - dones
next_actions = self.actor_target(next_states)
next_values = self.critic_target(
torch.cat((next_states, next_actions), dim=-1))
curr_returns = rewards + self.hyper_params[
"GAMMA"] * next_values * masks
curr_returns = curr_returns.to(device)
# train critic
values = self.critic(torch.cat((states, actions), dim=-1))
critic_loss = F.mse_loss(values, curr_returns)
self.critic_optimizer.zero_grad()
critic_loss.backward()
self.critic_optimizer.step()
# train actor
actions = self.actor(states)
actor_loss = -self.critic(torch.cat((states, actions), dim=-1)).mean()
self.actor_optimizer.zero_grad()
actor_loss.backward()
self.actor_optimizer.step()
# update target networks
tau = self.hyper_params["TAU"]
common_utils.soft_update(self.actor, self.actor_target, tau)
common_utils.soft_update(self.critic, self.critic_target, tau)
return actor_loss.data, critic_loss.data
def load_params(self, path: str):
"""Load model and optimizer parameters."""
if not os.path.exists(path):
print("[ERROR] the input path does not exist. ->", path)
return
params = torch.load(path)
self.actor.load_state_dict(params["actor_state_dict"])
self.actor_target.load_state_dict(params["actor_target_state_dict"])
self.critic.load_state_dict(params["critic_state_dict"])
self.critic_target.load_state_dict(params["critic_target_state_dict"])
self.actor_optimizer.load_state_dict(params["actor_optim_state_dict"])
self.critic_optimizer.load_state_dict(
params["critic_optim_state_dict"])
print("[INFO] loaded the model and optimizer from", path)
def save_params(self, n_episode: int):
"""Save model and optimizer parameters."""
params = {
"actor_state_dict": self.actor.state_dict(),
"actor_target_state_dict": self.actor_target.state_dict(),
"critic_state_dict": self.critic.state_dict(),
"critic_target_state_dict": self.critic_target.state_dict(),
"actor_optim_state_dict": self.actor_optimizer.state_dict(),
"critic_optim_state_dict": self.critic_optimizer.state_dict(),
}
AbstractAgent.save_params(self, params, n_episode)
def write_log(self, i: int, loss: np.ndarray, score: int):
"""Write log about loss and score"""
total_loss = loss.sum()
print(
"[INFO] episode %d, episode step: %d, total step: %d, total score: %d\n"
"total loss: %f actor_loss: %.3f critic_loss: %.3f\n" % (
i,
self.episode_step,
self.total_step,
score,
total_loss,
loss[0],
loss[1],
) # actor loss # critic loss
)
if self.args.log:
wandb.log({
"score": score,
"total loss": total_loss,
"actor loss": loss[0],
"critic loss": loss[1],
})
# pylint: disable=no-self-use, unnecessary-pass
def pretrain(self):
"""Pretraining steps."""
pass
def train(self):
"""Train the agent."""
# logger
if self.args.log:
wandb.init()
wandb.config.update(self.hyper_params)
# wandb.watch([self.actor, self.critic], log="parameters")
# pre-training if needed
self.pretrain()
for self.i_episode in range(1, self.args.episode_num + 1):
state = self.env.reset()
done = False
score = 0
self.episode_step = 0
losses = list()
while not done:
if self.args.render and self.i_episode >= self.args.render_after:
self.env.render()
action = self.select_action(state)
next_state, reward, done = self.step(action)
if len(self.memory) >= self.hyper_params["BATCH_SIZE"]:
for _ in range(self.hyper_params["MULTIPLE_LEARN"]):
loss = self.update_model()
losses.append(loss) # for logging
state = next_state
score += reward
# logging
if losses:
avg_loss = np.vstack(losses).mean(axis=0)
self.write_log(self.i_episode, avg_loss, score)
losses.clear()
if self.i_episode % self.args.save_period == 0:
self.save_params(self.i_episode)
self.interim_test()
# termination
self.env.close()
self.save_params(self.i_episode)
self.interim_test()
class BCDDPGAgent(DDPGAgent):
"""BC with DDPG agent interacting with environment.
Attributes:
her (HER): hinsight experience replay
transitions_epi (list): transitions per episode (for HER)
desired_state (np.ndarray): desired state of current episode
memory (ReplayBuffer): replay memory
demo_memory (ReplayBuffer): replay memory for demo
lambda1 (float): proportion of policy loss
lambda2 (float): proportion of BC loss
"""
def __init__(
self,
env: gym.Env,
args: argparse.Namespace,
hyper_params: dict,
models: tuple,
optims: tuple,
noise: OUNoise,
her: HER,
):
"""Initialization.
Args:
her (HER): hinsight experience replay
"""
self.her = her
DDPGAgent.__init__(self, env, args, hyper_params, models, optims, noise)
# pylint: disable=attribute-defined-outside-init
def _initialize(self):
"""Initialize non-common things."""
# load demo replay memory
with open(self.args.demo_path, "rb") as f:
demo = list(pickle.load(f))
# HER
if self.hyper_params["USE_HER"]:
if self.hyper_params["DESIRED_STATES_FROM_DEMO"]:
self.her.fetch_desired_states_from_demo(demo)
self.transitions_epi: list = list()
self.desired_state = np.zeros((1,))
demo = self.her.generate_demo_transitions(demo)
if not self.args.test:
# Replay buffers
demo_batch_size = self.hyper_params["DEMO_BATCH_SIZE"]
self.demo_memory = ReplayBuffer(len(demo), demo_batch_size)
self.demo_memory.extend(demo)
self.memory = ReplayBuffer(
self.hyper_params["BUFFER_SIZE"], self.hyper_params["BATCH_SIZE"]
)
# set hyper parameters
self.lambda1 = self.hyper_params["LAMBDA1"]
self.lambda2 = self.hyper_params["LAMBDA2"] / demo_batch_size
def _preprocess_state(self, state: np.ndarray) -> torch.Tensor:
"""Preprocess state so that actor selects an action."""
if self.hyper_params["USE_HER"]:
self.desired_state = self.her.get_desired_state()
state = np.concatenate((state, self.desired_state), axis=-1)
state = torch.FloatTensor(state).to(device)
return state
def _add_transition_to_memory(self, transition: Tuple[np.ndarray, ...]):
"""Add 1 step and n step transitions to memory."""
if self.hyper_params["USE_HER"]:
self.transitions_epi.append(transition)
done = transition[-1] or self.episode_step == self.args.max_episode_steps
if done:
# insert generated transitions if the episode is done
transitions = self.her.generate_transitions(
self.transitions_epi,
self.desired_state,
self.hyper_params["SUCCESS_SCORE"],
)
self.memory.extend(transitions)
self.transitions_epi.clear()
else:
self.memory.add(transition)
def update_model(self) -> Tuple[torch.Tensor, ...]:
"""Train the model after each episode."""
experiences = self.memory.sample()
demos = self.demo_memory.sample()
exp_states, exp_actions, exp_rewards, exp_next_states, exp_dones = experiences
demo_states, demo_actions, demo_rewards, demo_next_states, demo_dones = demos
states = torch.cat((exp_states, demo_states), dim=0)
actions = torch.cat((exp_actions, demo_actions), dim=0)
rewards = torch.cat((exp_rewards, demo_rewards), dim=0)
next_states = torch.cat((exp_next_states, demo_next_states), dim=0)
dones = torch.cat((exp_dones, demo_dones), dim=0)
# G_t = r + gamma * v(s_{t+1}) if state != Terminal
# = r otherwise
masks = 1 - dones
next_actions = self.actor_target(next_states)
next_values = self.critic_target(torch.cat((next_states, next_actions), dim=-1))
curr_returns = rewards + (self.hyper_params["GAMMA"] * next_values * masks)
curr_returns = curr_returns.to(device)
# critic loss
values = self.critic(torch.cat((states, actions), dim=-1))
critic_loss = F.mse_loss(values, curr_returns)
# train critic
gradient_clip_cr = self.hyper_params["GRADIENT_CLIP_CR"]
self.critic_optimizer.zero_grad()
critic_loss.backward()
nn.utils.clip_grad_norm_(self.critic.parameters(), gradient_clip_cr)
self.critic_optimizer.step()
# policy loss
actions = self.actor(states)
policy_loss = -self.critic(torch.cat((states, actions), dim=-1)).mean()
# bc loss
pred_actions = self.actor(demo_states)
qf_mask = torch.gt(
self.critic(torch.cat((demo_states, demo_actions), dim=-1)),
self.critic(torch.cat((demo_states, pred_actions), dim=-1)),
).to(device)
qf_mask = qf_mask.float()
n_qf_mask = int(qf_mask.sum().item())
if n_qf_mask == 0:
bc_loss = torch.zeros(1, device=device)
else:
bc_loss = (
torch.mul(pred_actions, qf_mask) - torch.mul(demo_actions, qf_mask)
).pow(2).sum() / n_qf_mask
# train actor: pg loss + BC loss
actor_loss = self.lambda1 * policy_loss + self.lambda2 * bc_loss
gradient_clip_ac = self.hyper_params["GRADIENT_CLIP_AC"]
self.actor_optimizer.zero_grad()
actor_loss.backward()
nn.utils.clip_grad_norm_(self.actor.parameters(), gradient_clip_ac)
self.actor_optimizer.step()
# update target networks
tau = self.hyper_params["TAU"]
common_utils.soft_update(self.actor, self.actor_target, tau)
common_utils.soft_update(self.critic, self.critic_target, tau)
return actor_loss.item(), critic_loss.item(), n_qf_mask
def write_log(self, i: int, loss: np.ndarray, score: int, avg_time_cost):
"""Write log about loss and score"""
total_loss = loss.sum()
print(
"[INFO] episode %d, episode step: %d, total step: %d, total score: %d\n"
"total loss: %f actor_loss: %.3f critic_loss: %.3f, n_qf_mask: %d "
"(spent %.6f sec/step)\n"
% (
i,
self.episode_step,
self.total_step,
score,
total_loss,
loss[0],
loss[1],
loss[2],
avg_time_cost,
) # actor loss # critic loss
)
if self.args.log:
wandb.log(
{
"score": score,
"total loss": total_loss,
"actor loss": loss[0],
"critic loss": loss[1],
"time per each step": avg_time_cost,
}
)
class SACAgent(Agent):
"""SAC agent interacting with environment.
Attrtibutes:
memory (ReplayBuffer): replay memory
actor (nn.Module): actor model to select actions
actor_target (nn.Module): target actor model to select actions
actor_optimizer (Optimizer): optimizer for training actor
critic_1 (nn.Module): critic model to predict state values
critic_2 (nn.Module): critic model to predict state values
critic_target1 (nn.Module): target critic model to predict state values
critic_target2 (nn.Module): target critic model to predict state values
critic_optimizer1 (Optimizer): optimizer for training critic_1
critic_optimizer2 (Optimizer): optimizer for training critic_2
curr_state (np.ndarray): temporary storage of the current state
target_entropy (int): desired entropy used for the inequality constraint
beta (float): beta parameter for prioritized replay buffer
alpha (torch.Tensor): weight for entropy
alpha_optimizer (Optimizer): optimizer for alpha
hyper_params (dict): hyper-parameters
total_step (int): total step numbers
episode_step (int): step number of the current episode
update_step (int): step number of updates
i_episode (int): current episode number
"""
def __init__(
self,
env: gym.Env,
args: argparse.Namespace,
hyper_params: dict,
models: tuple,
optims: tuple,
target_entropy: float,
):
"""Initialization.
Args:
env (gym.Env): openAI Gym environment
args (argparse.Namespace): arguments including hyperparameters and training settings
hyper_params (dict): hyper-parameters
models (tuple): models including actor and critic
optims (tuple): optimizers for actor and critic
target_entropy (float): target entropy for the inequality constraint
"""
Agent.__init__(self, env, args)
self.actor, self.vf, self.vf_target, self.qf_1, self.qf_2 = models
self.actor_optimizer, self.vf_optimizer = optims[0:2]
self.qf_1_optimizer, self.qf_2_optimizer = optims[2:4]
self.hyper_params = hyper_params
self.curr_state = np.zeros((1,))
self.total_step = 0
self.episode_step = 0
self.update_step = 0
self.i_episode = 0
# automatic entropy tuning
if self.hyper_params["AUTO_ENTROPY_TUNING"]:
self.target_entropy = target_entropy
self.log_alpha = torch.zeros(1, requires_grad=True, device=device)
self.alpha_optimizer = optim.Adam(
[self.log_alpha], lr=self.hyper_params["LR_ENTROPY"]
)
# load the optimizer and model parameters
if args.load_from is not None and os.path.exists(args.load_from):
self.load_params(args.load_from)
self._initialize()
# pylint: disable=attribute-defined-outside-init
def _initialize(self):
"""Initialize non-common things."""
if not self.args.test:
# replay memory
self.memory = ReplayBuffer(
self.hyper_params["BUFFER_SIZE"], self.hyper_params["BATCH_SIZE"]
)
def select_action(self, state: np.ndarray) -> np.ndarray:
"""Select an action from the input space."""
self.curr_state = state
state = self._preprocess_state(state)
# if initial random action should be conducted
if (
self.total_step < self.hyper_params["INITIAL_RANDOM_ACTION"]
and not self.args.test
):
return self.env.action_space.sample()
if self.args.test and not self.is_discrete:
_, _, _, selected_action, _ = self.actor(state)
else:
selected_action, _, _, _, _ = self.actor(state)
return selected_action.detach().cpu().numpy()
# pylint: disable=no-self-use
def _preprocess_state(self, state: np.ndarray) -> torch.Tensor:
"""Preprocess state so that actor selects an action."""
state = torch.FloatTensor(state).to(device)
return state
def step(self, action: np.ndarray) -> Tuple[np.ndarray, np.float64, bool]:
"""Take an action and return the response of the env."""
next_state, reward, done, _ = self.env.step(action)
if not self.args.test:
# if the last state is not a terminal state, store done as false
done_bool = (
False if self.episode_step == self.args.max_episode_steps else done
)
transition = (self.curr_state, action, reward, next_state, done_bool)
self._add_transition_to_memory(transition)
return next_state, reward, done
def _add_transition_to_memory(self, transition: Tuple[np.ndarray, ...]):
"""Add 1 step and n step transitions to memory."""
self.memory.add(*transition)
def update_model(self) -> Tuple[torch.Tensor, ...]:
"""Train the model after each episode."""
self.update_step += 1
experiences = self.memory.sample()
states, actions, rewards, next_states, dones = experiences
new_actions, log_prob, pre_tanh_value, mu, std = self.actor(states)
# train alpha
if self.hyper_params["AUTO_ENTROPY_TUNING"]:
alpha_loss = (
-self.log_alpha * (log_prob + self.target_entropy).detach()
).mean()
self.alpha_optimizer.zero_grad()
alpha_loss.backward()
self.alpha_optimizer.step()
alpha = self.log_alpha.exp()
else:
alpha_loss = torch.zeros(1)
alpha = self.hyper_params["W_ENTROPY"]
# Q function loss
masks = 1 - dones
q_1_pred = self.qf_1(states, actions)
q_2_pred = self.qf_2(states, actions)
v_target = self.vf_target(next_states)
q_target = rewards + self.hyper_params["GAMMA"] * v_target * masks
qf_1_loss = F.mse_loss(q_1_pred, q_target.detach())
qf_2_loss = F.mse_loss(q_2_pred, q_target.detach())
# V function loss
v_pred = self.vf(states)
q_pred = torch.min(
self.qf_1(states, new_actions), self.qf_2(states, new_actions)
)
v_target = q_pred - alpha * log_prob
vf_loss = F.mse_loss(v_pred, v_target.detach())
# train Q functions
self.qf_1_optimizer.zero_grad()
qf_1_loss.backward()
self.qf_1_optimizer.step()
self.qf_2_optimizer.zero_grad()
qf_2_loss.backward()
self.qf_2_optimizer.step()
# train V function
self.vf_optimizer.zero_grad()
vf_loss.backward()
self.vf_optimizer.step()
if self.update_step % self.hyper_params["POLICY_UPDATE_FREQ"] == 0:
# actor loss
advantage = q_pred - v_pred.detach()
actor_loss = (alpha * log_prob - advantage).mean()
# regularization
if not self.is_discrete: # iff the action is continuous
mean_reg = self.hyper_params["W_MEAN_REG"] * mu.pow(2).mean()
std_reg = self.hyper_params["W_STD_REG"] * std.pow(2).mean()
pre_activation_reg = self.hyper_params["W_PRE_ACTIVATION_REG"] * (
pre_tanh_value.pow(2).sum(dim=-1).mean()
)
actor_reg = mean_reg + std_reg + pre_activation_reg
# actor loss + regularization
actor_loss += actor_reg
# train actor
self.actor_optimizer.zero_grad()
actor_loss.backward()
self.actor_optimizer.step()
# update target networks
common_utils.soft_update(self.vf, self.vf_target, self.hyper_params["TAU"])
else:
actor_loss = torch.zeros(1)
return (
actor_loss.item(),
qf_1_loss.item(),
qf_2_loss.item(),
vf_loss.item(),
alpha_loss.item(),
)
def load_params(self, path: str):
"""Load model and optimizer parameters."""
if not os.path.exists(path):
print("[ERROR] the input path does not exist. ->", path)
return
params = torch.load(path)
self.actor.load_state_dict(params["actor"])
self.qf_1.load_state_dict(params["qf_1"])
self.qf_2.load_state_dict(params["qf_2"])
self.vf.load_state_dict(params["vf"])
self.vf_target.load_state_dict(params["vf_target"])
self.actor_optimizer.load_state_dict(params["actor_optim"])
self.qf_1_optimizer.load_state_dict(params["qf_1_optim"])
self.qf_2_optimizer.load_state_dict(params["qf_2_optim"])
self.vf_optimizer.load_state_dict(params["vf_optim"])
if self.hyper_params["AUTO_ENTROPY_TUNING"]:
self.alpha_optimizer.load_state_dict(params["alpha_optim"])
print("[INFO] loaded the model and optimizer from", path)
def save_params(self, n_episode: int):
"""Save model and optimizer parameters."""
params = {
"actor": self.actor.state_dict(),
"qf_1": self.qf_1.state_dict(),
"qf_2": self.qf_2.state_dict(),
"vf": self.vf.state_dict(),
"vf_target": self.vf_target.state_dict(),
"actor_optim": self.actor_optimizer.state_dict(),
"qf_1_optim": self.qf_1_optimizer.state_dict(),
"qf_2_optim": self.qf_2_optimizer.state_dict(),
"vf_optim": self.vf_optimizer.state_dict(),
}
if self.hyper_params["AUTO_ENTROPY_TUNING"]:
params["alpha_optim"] = self.alpha_optimizer.state_dict()
Agent.save_params(self, params, n_episode)
def write_log(
self, i: int, loss: np.ndarray, score: float = 0.0, policy_update_freq: int = 1
):
"""Write log about loss and score"""
total_loss = loss.sum()
print(
"[INFO] episode %d, episode_step %d, total step %d, total score: %d\n"
"total loss: %.3f actor_loss: %.3f qf_1_loss: %.3f qf_2_loss: %.3f "
"vf_loss: %.3f alpha_loss: %.3f\n"
% (
i,
self.episode_step,
self.total_step,
score,
total_loss,
loss[0] * policy_update_freq, # actor loss
loss[1], # qf_1 loss
loss[2], # qf_2 loss
loss[3], # vf loss
loss[4], # alpha loss
)
)
if self.args.log:
wandb.log(
{
"score": score,
"total loss": total_loss,
"actor loss": loss[0] * policy_update_freq,
"qf_1 loss": loss[1],
"qf_2 loss": loss[2],
"vf loss": loss[3],
"alpha loss": loss[4],
}
)
# pylint: disable=no-self-use, unnecessary-pass
def pretrain(self):
"""Pretraining steps."""
pass
def train(self):
"""Train the agent."""
# logger
if self.args.log:
wandb.init(project=self.args.wandb_project)
wandb.config.update(self.hyper_params)
# wandb.watch([self.actor, self.vf, self.qf_1, self.qf_2], log="parameters")
# pre-training if needed
self.pretrain()
for self.i_episode in range(1, self.args.episode_num + 1):
state = self.env.reset()
done = False
score = 0
self.episode_step = 0
loss_episode = list()
while not done:
if self.args.render and self.i_episode >= self.args.render_after:
self.env.render()
action = self.select_action(state)
next_state, reward, done = self.step(action)
self.total_step += 1
self.episode_step += 1
state = next_state
score += reward
# training
if len(self.memory) >= self.hyper_params["BATCH_SIZE"]:
for _ in range(self.hyper_params["MULTIPLE_LEARN"]):
loss = self.update_model()
loss_episode.append(loss) # for logging
# logging
if loss_episode:
avg_loss = np.vstack(loss_episode).mean(axis=0)
self.write_log(
self.i_episode,
avg_loss,
score,
self.hyper_params["POLICY_UPDATE_FREQ"],
)
if self.i_episode % self.args.save_period == 0:
self.save_params(self.i_episode)
self.interim_test()
# termination
self.env.close()
self.save_params(self.i_episode)
self.interim_test()
class Agent(SACAgent):
"""BC with SAC agent interacting with environment.
Attrtibutes:
HER (AbstractHER): hinsight experience replay
transitions_epi (list): transitions per episode (for HER)
desired_state (np.ndarray): desired state of current episode
memory (ReplayBuffer): replay memory
demo_memory (ReplayBuffer): replay memory for demo
lambda1 (float): proportion of policy loss
lambda2 (float): proportion of BC loss
"""
def __init__(
self,
env: gym.Env,
args: argparse.Namespace,
hyper_params: dict,
models: tuple,
optims: tuple,
target_entropy: float,
HER: AbstractHER,
):
"""Initialization.
Args:
HER (AbstractHER): hinsight experience replay
"""
self.HER = HER
SACAgent.__init__(self, env, args, hyper_params, models, optims,
target_entropy)
# pylint: disable=attribute-defined-outside-init
def _initialize(self):
"""Initialize non-common things."""
# load demo replay memory
with open(self.args.demo_path, "rb") as f:
demo = list(pickle.load(f))
# HER
if self.hyper_params["USE_HER"]:
self.her = self.HER()
if self.hyper_params["DESIRED_STATES_FROM_DEMO"]:
self.her.fetch_desired_states_from_demo(demo)
self.transitions_epi: list = list()
self.desired_state = np.zeros((1, ))
demo = self.her.generate_demo_transitions(demo)
if not self.args.test:
# Replay buffers
demo_batch_size = self.hyper_params["DEMO_BATCH_SIZE"]
self.demo_memory = ReplayBuffer(len(demo), demo_batch_size)
self.demo_memory.extend(demo)
self.memory = ReplayBuffer(self.hyper_params["BUFFER_SIZE"],
self.hyper_params["BATCH_SIZE"])
# set hyper parameters
self.lambda1 = self.hyper_params["LAMBDA1"]
self.lambda2 = self.hyper_params["LAMBDA2"] / demo_batch_size
def _preprocess_state(self, state: np.ndarray) -> torch.Tensor:
"""Preprocess state so that actor selects an action."""
if self.hyper_params["USE_HER"]:
self.desired_state = self.her.get_desired_state()
state = np.concatenate((state, self.desired_state), axis=-1)
state = torch.FloatTensor(state).to(device)
return state
def _add_transition_to_memory(self, transition: Tuple[np.ndarray, ...]):
"""Add 1 step and n step transitions to memory."""
if self.hyper_params["USE_HER"]:
self.transitions_epi.append(transition)
done = transition[
-1] or self.episode_step == self.args.max_episode_steps
if done:
# insert generated transitions if the episode is done
transitions = self.her.generate_transitions(
self.transitions_epi,
self.desired_state,
self.hyper_params["SUCCESS_SCORE"],
)
self.memory.extend(transitions)
self.transitions_epi.clear()
else:
self.memory.add(*transition)
def update_model(self) -> Tuple[torch.Tensor, ...]:
"""Train the model after each episode."""
experiences = self.memory.sample()
demos = self.demo_memory.sample()
states, actions, rewards, next_states, dones = experiences
demo_states, demo_actions, _, _, _ = demos
new_actions, log_prob, pre_tanh_value, mu, std = self.actor(states)
pred_actions, _, _, _, _ = self.actor(demo_states)
# train alpha
if self.hyper_params["AUTO_ENTROPY_TUNING"]:
alpha_loss = (-self.log_alpha *
(log_prob + self.target_entropy).detach()).mean()
self.alpha_optimizer.zero_grad()
alpha_loss.backward()
self.alpha_optimizer.step()
alpha = self.log_alpha.exp()
else:
alpha_loss = torch.zeros(1)
alpha = self.hyper_params["W_ENTROPY"]
# Q function loss
masks = 1 - dones
q_1_pred = self.qf_1(states, actions)
q_2_pred = self.qf_2(states, actions)
v_target = self.vf_target(next_states)
q_target = rewards + self.hyper_params["GAMMA"] * v_target * masks
qf_1_loss = F.mse_loss(q_1_pred, q_target.detach())
qf_2_loss = F.mse_loss(q_2_pred, q_target.detach())
# V function loss
v_pred = self.vf(states)
q_pred = torch.min(self.qf_1(states, new_actions),
self.qf_2(states, new_actions))
v_target = q_pred - alpha * log_prob
vf_loss = F.mse_loss(v_pred, v_target.detach())
# train Q functions
self.qf_1_optimizer.zero_grad()
qf_1_loss.backward()
self.qf_1_optimizer.step()
self.qf_2_optimizer.zero_grad()
qf_2_loss.backward()
self.qf_2_optimizer.step()
# train V function
self.vf_optimizer.zero_grad()
vf_loss.backward()
self.vf_optimizer.step()
if self.total_step % self.hyper_params["DELAYED_UPDATE"] == 0:
# bc loss
qf_mask = torch.gt(
self.qf_1(demo_states, demo_actions),
self.qf_1(demo_states, pred_actions),
).to(device)
qf_mask = qf_mask.float()
n_qf_mask = int(qf_mask.sum().item())
if n_qf_mask == 0:
bc_loss = torch.zeros(1, device=device)
else:
bc_loss = (torch.mul(pred_actions, qf_mask) - torch.mul(
demo_actions, qf_mask)).pow(2).sum() / n_qf_mask
# actor loss
advantage = q_pred - v_pred.detach()
actor_loss = (alpha * log_prob - advantage).mean()
actor_loss = self.lambda1 * actor_loss + self.lambda2 * bc_loss
# regularization
if not self.is_discrete: # iff the action is continuous
mean_reg = self.hyper_params["W_MEAN_REG"] * mu.pow(2).mean()
std_reg = self.hyper_params["W_STD_REG"] * std.pow(2).mean()
pre_activation_reg = self.hyper_params[
"W_PRE_ACTIVATION_REG"] * (pre_tanh_value.pow(2).sum(
dim=-1).mean())
actor_reg = mean_reg + std_reg + pre_activation_reg
# actor loss + regularization
actor_loss += actor_reg
# train actor
self.actor_optimizer.zero_grad()
actor_loss.backward()
self.actor_optimizer.step()
# update target networks
common_utils.soft_update(self.vf, self.vf_target,
self.hyper_params["TAU"])
else:
actor_loss = torch.zeros(1)
n_qf_mask = 0
return (
actor_loss.data,
qf_1_loss.data,
qf_2_loss.data,
vf_loss.data,
alpha_loss.data,
n_qf_mask,
)
def write_log(self,
i: int,
loss: np.ndarray,
score: float = 0.0,
delayed_update: int = 1):
"""Write log about loss and score"""
total_loss = loss.sum()
print(
"[INFO] episode %d, episode_step %d, total step %d, total score: %d\n"
"total loss: %.3f actor_loss: %.3f qf_1_loss: %.3f qf_2_loss: %.3f "
"vf_loss: %.3f alpha_loss: %.3f n_qf_mask: %d\n" % (
i,
self.episode_step,
self.total_step,
score,
total_loss,
loss[0] * delayed_update, # actor loss
loss[1], # qf_1 loss
loss[2], # qf_2 loss
loss[3], # vf loss
loss[4], # alpha loss
loss[5], # n_qf_mask
))
if self.args.log:
wandb.log({
"score": score,
"total loss": total_loss,
"actor loss": loss[0] * delayed_update,
"qf_1 loss": loss[1],
"qf_2 loss": loss[2],
"vf loss": loss[3],
"alpha loss": loss[4],
})
