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 = NStepTransitionBuffer(
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 = PrioritizedReplayBufferfD(
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
with open(self.args.demo_path, "rb") as f:
demo = pickle.load(f)
# replay memory
self.beta = self.hyper_params["PER_BETA"]
self.memory = PrioritizedReplayBufferfD(
self.hyper_params["BUFFER_SIZE"],
self.hyper_params["BATCH_SIZE"],
demo=list(demo),
alpha=self.hyper_params["PER_ALPHA"],
epsilon_d=self.hyper_params["PER_EPS_DEMO"],
)
class Agent(DDPGAgent):
"""ActorCritic interacting with environment.
Attributes:
memory (PrioritizedReplayBufferfD): replay memory
beta (float): beta parameter for prioritized replay buffer
"""
# pylint: disable=attribute-defined-outside-init
def _initialize(self):
"""Initialize non-common things."""
if not self.args.test:
# load demo replay memory
with open(self.args.demo_path, "rb") as f:
demo = pickle.load(f)
# replay memory
self.beta = self.hyper_params["PER_BETA"]
self.memory = PrioritizedReplayBufferfD(
self.hyper_params["BUFFER_SIZE"],
self.hyper_params["BATCH_SIZE"],
demo=list(demo),
alpha=self.hyper_params["PER_ALPHA"],
epsilon_d=self.hyper_params["PER_EPS_DEMO"],
)
def update_model(self) -> Tuple[torch.Tensor, torch.Tensor]:
"""Train the model after each episode."""
experiences = self.memory.sample(self.beta)
states, actions, rewards, next_states, dones, weights, indexes, eps_d = (
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_states_actions = torch.cat((next_states, next_actions), dim=-1)
next_values = self.critic_target(next_states_actions)
curr_returns = rewards + self.hyper_params[
"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)
critic_loss = torch.mean(critic_loss_element_wise * weights)
self.critic_optimizer.zero_grad()
critic_loss.backward()
self.critic_optimizer.step()
# train actor
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()
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(indexes, new_priorities)
# increase beta
fraction = min(
float(self.i_episode) / self.args.max_episode_steps, 1.0)
self.beta = self.beta + fraction * (1.0 - self.beta)
return actor_loss.data, critic_loss.data
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):
loss = self.update_model()
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)
class SACfDAgent(SACAgent):
"""SAC agent interacting with environment.
Attrtibutes:
memory (PrioritizedReplayBufferfD): 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 = NStepTransitionBuffer(
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 = PrioritizedReplayBufferfD(
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):
loss = self.update_model()
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"],
)
print("[INFO] Pre-Train Complete!\n")
class DQfDAgent(DQNAgent):
"""DQN interacting with environment.
Attribute:
memory (PrioritizedReplayBufferfD): 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 = NStepTransitionBuffer(
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 = PrioritizedReplayBufferfD(
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 DDPGfDAgent(DDPGAgent):
"""ActorCritic interacting with environment.
Attributes:
memory (PrioritizedReplayBufferfD): 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 = NStepTransitionBuffer(
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 = PrioritizedReplayBufferfD(
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):
loss = self.update_model()
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)
print("[INFO] Pre-Train Complete!\n")
class Agent(DQNAgent):
"""DQN interacting with environment.
Attribute:
memory (PrioritizedReplayBufferfD): replay memory
"""
# pylint: disable=attribute-defined-outside-init
def _initialize(self):
"""Initialize non-common things."""
if not self.args.test:
# load demo replay memory
with open(self.args.demo_path, "rb") as f:
demo = pickle.load(f)
# replay memory
self.beta = self.hyper_params["PER_BETA"]
self.memory = PrioritizedReplayBufferfD(
self.hyper_params["BUFFER_SIZE"],
self.hyper_params["BATCH_SIZE"],
demo=list(demo),
alpha=self.hyper_params["PER_ALPHA"],
epsilon_d=self.hyper_params["PER_EPS_DEMO"],
)
def update_model(self) -> Tuple[torch.Tensor, ...]:
"""Train the model after each episode."""
experiences = self.memory.sample()
states, actions, rewards, next_states, dones, weights, indexes, eps_d = (
experiences
)
q_values = self.dqn(states, self.epsilon)
next_q_values = self.dqn(next_states, self.epsilon)
next_target_q_values = self.dqn_target(next_states, self.epsilon)
curr_q_values = q_values.gather(1, actions.long().unsqueeze(1))
next_q_values = next_target_q_values.gather( # Double DQN
1, next_q_values.argmax(1).unsqueeze(1)
)
# G_t = r + gamma * v(s_{t+1}) if state != Terminal
# = r otherwise
masks = 1 - dones
target = rewards + self.hyper_params["GAMMA"] * next_q_values * masks
target = target.to(device)
# calculate dq loss
dq_loss_element_wise = (target - curr_q_values).pow(2)
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)
if demo_idxs[0].size != 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(indexes, new_priorities)
# increase beta
fraction = min(float(self.i_episode) / self.args.max_episode_steps, 1.0)
self.beta = self.beta + fraction * (1.0 - self.beta)
return loss.data, dq_loss.data, supervised_loss.data
def write_log(self, i: int, avg_loss: np.ndarray, score: int = 0):
"""Write log about loss and score"""
print(
"[INFO] episode %d, episode step: %d, total step: %d, total score: %d\n"
"epsilon: %f, total loss: %f, dq loss: %f, supervised loss: %f\n"
"at %s\n"
% (
i,
self.episode_steps[0],
self.total_steps.sum(),
score,
self.epsilon,
avg_loss[0],
avg_loss[1],
avg_loss[2],
datetime.datetime.now(),
)
)
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],
}
)
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):
loss = self.update_model()
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)
class Agent(TD3Agent):
"""TD3 agent interacting with environment.
Attrtibutes:
memory (PrioritizedReplayBufferfD): 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
# TODO: should make new demo to set protocol 2
# e.g. pickle.dump(your_object, your_file, protocol=2)
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 = NStepTransitionBuffer(
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 = PrioritizedReplayBufferfD(
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):
"""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, gamma):
"""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
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 + (gamma * target_values * masks).detach()
# train critic
values1 = self.critic1(torch.cat((states, actions), dim=-1))
critic1_loss_element_wise = (values1 - target_values.detach()).pow(2)
values2 = self.critic2(torch.cat((states, actions), dim=-1))
critic2_loss_element_wise = (values2 - target_values.detach()).pow(2)
return critic1_loss_element_wise, critic2_loss_element_wise
# pylint: disable=too-many-statements
def update_model(self, experiences):
"""Train the model after each episode."""
states, actions, rewards, next_states, dones, weights, indices, eps_d = (
experiences)
gamma = self.hyper_params["GAMMA"]
critic1_loss_element_wise, critic2_loss_element_wise = self._get_critic_loss(
experiences, gamma)
critic_loss_element_wise = critic1_loss_element_wise + critic2_loss_element_wise
critic1_loss = torch.mean(critic1_loss_element_wise * weights)
critic2_loss = torch.mean(critic2_loss_element_wise * weights)
critic_loss = critic1_loss + critic2_loss
if self.use_n_step:
experiences_n = self.memory_n.sample(indices)
gamma = self.hyper_params["GAMMA"]**self.hyper_params["N_STEP"]
critic1_loss_n_element_wise, critic2_loss_n_element_wise = self._get_critic_loss(
experiences_n, gamma)
critic_loss_n_element_wise = (critic1_loss_n_element_wise +
critic2_loss_n_element_wise)
critic1_loss_n = torch.mean(critic1_loss_n_element_wise * weights)
critic2_loss_n = torch.mean(critic2_loss_n_element_wise * weights)
critic_loss_n = critic1_loss_n + critic2_loss_n
lambda1 = self.hyper_params["LAMBDA1"]
critic_loss_element_wise += lambda1 * critic_loss_n_element_wise
critic_loss += lambda1 * critic_loss_n
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_element_wise = -self.critic1(
torch.cat((states, actions), dim=-1))
actor_loss = torch.mean(actor_loss_element_wise * weights)
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)
# 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)
else:
actor_loss = torch.zeros(1)
return actor_loss.data, critic1_loss.data, critic2_loss.data
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):
loss = self.update_model()
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,
delayed_update=self.hyper_params["DELAYED_UPDATE"])