def train(self):
# Sample
batch = self.replay_buffer.sample()
obs = batch["obs"].to(self.device)
acts = batch["acts"].to(self.device)
rews = batch["rews"].to(self.device)
next_obs = batch["next_obs"].to(self.device)
done = batch["done"].to(self.device)
# Compute target Q value
with torch.no_grad():
next_act = self.target_actor_net(next_obs)
next_Q = self.target_critic_net(next_obs, next_act).squeeze(1)
target_Q = rews + (1. - done) * self.gamma * next_Q
# Compute current Q
current_Q = self.critic_net(obs, acts).squeeze(1)
# Compute critic loss
critic_loss = F.mse_loss(current_Q, target_Q)
# Compute actor loss
actor_loss = -self.critic_net(obs, self.actor_net(obs)).mean()
# Optimize actor net
self.actor_optimizer.zero_grad()
actor_loss.backward()
self.actor_optimizer.step()
# Optimize critic net
self.critic_optimizer.zero_grad()
critic_loss.backward()
self.critic_optimizer.step()
soft_target_update(self.actor_net, self.target_actor_net, tau=self.tau)
soft_target_update(self.critic_net,
self.target_critic_net,
tau=self.tau)
self.train_step += 1
return actor_loss.cpu().item(), critic_loss.cpu().item()
def train(self):
# Sample
batch = self.data_buffer.sample()
obs = batch["obs"].to(self.device)
acts = batch["acts"].to(self.device)
rews = batch["rews"].to(self.device)
next_obs = batch["next_obs"].to(self.device)
done = batch["done"].to(self.device)
"""
Train Critic
"""
with torch.no_grad():
decode_action_next = self.target_actor_net(next_obs,
self.cvae_net.decode)
target_q1 = self.target_critic_net1(next_obs, decode_action_next)
target_q2 = self.target_critic_net2(next_obs, decode_action_next)
target_q = (
self.lmbda * torch.min(target_q1, target_q2) +
(1. - self.lmbda) * torch.max(target_q1, target_q2)).squeeze(1)
target_q = rews + self.gamma * (1. - done) * target_q
current_q1 = self.critic_net1(obs, acts).squeeze(1)
current_q2 = self.critic_net2(obs, acts).squeeze(1)
critic_loss = F.mse_loss(current_q1, target_q) + F.mse_loss(
current_q2, target_q)
self.critic_optimizer1.zero_grad()
self.critic_optimizer2.zero_grad()
critic_loss.backward()
self.critic_optimizer1.step()
self.critic_optimizer2.step()
"""
Train Actor
"""
decode_action = self.actor_net(obs, self.cvae_net.decode)
actor_loss = -self.critic_net1(obs, decode_action).mean()
self.actor_optimizer.zero_grad()
actor_loss.backward()
self.actor_optimizer.step()
"""
Update target networks
"""
soft_target_update(self.critic_net1,
self.target_critic_net1,
tau=self.tau)
soft_target_update(self.critic_net2,
self.target_critic_net2,
tau=self.tau)
soft_target_update(self.actor_net, self.target_actor_net, tau=self.tau)
self.train_step += 1
return critic_loss.cpu().item(), actor_loss.cpu().item()
def train(self):
# Sample
batch = self.replay_buffer.sample()
obs = batch["obs"].to(self.device)
acts = batch["acts"].to(self.device)
rews = batch["rews"].to(self.device)
next_obs = batch["next_obs"].to(self.device)
done = batch["done"].to(self.device)
# Target Policy Smoothing. Add clipped noise to next actions when computing target Q.
with torch.no_grad():
noise = torch.normal(mean=0, std=self.policy_noise, size=acts.size()).to(self.device)
noise = noise.clamp(-self.noise_clip, self.noise_clip)
next_act = self.target_actor_net(next_obs) + noise
next_act = next_act.clamp(-self.action_bound, self.action_bound)
# Clipped Double Q-Learning. Compute the min of target Q1 and target Q2
min_target_q = torch.min(self.target_critic_net1(next_obs, next_act),
self.target_critic_net2(next_obs, next_act)).squeeze(1)
y = rews + self.gamma * (1. - done) * min_target_q
current_q1 = self.critic_net1(obs, acts).squeeze(1)
current_q2 = self.critic_net2(obs, acts).squeeze(1)
# TD3 Loss
critic_loss1 = F.mse_loss(current_q1, y)
critic_loss2 = F.mse_loss(current_q2, y)
# Optimize critic net
self.critic_optimizer1.zero_grad()
critic_loss1.backward()
self.critic_optimizer1.step()
self.critic_optimizer2.zero_grad()
critic_loss2.backward()
self.critic_optimizer2.step()
if (self.train_step+1) % self.policy_delay == 0:
# Compute actor loss
actor_loss = -self.critic_net1(obs, self.actor_net(obs)).mean()
# Optimize actor net
self.actor_optimizer.zero_grad()
actor_loss.backward()
self.actor_optimizer.step()
soft_target_update(self.actor_net, self.target_actor_net, tau=self.tau)
soft_target_update(self.critic_net1, self.target_critic_net1, tau=self.tau)
soft_target_update(self.critic_net2, self.target_critic_net2, tau=self.tau)
else:
actor_loss = torch.tensor(0)
self.train_step += 1
return actor_loss.cpu().item(), critic_loss1.cpu().item(), critic_loss2.cpu().item()
def train(self):
# Sample
batch = self.data_buffer.sample()
obs = batch["obs"].to(self.device)
acts = batch["acts"].to(self.device)
rews = batch["rews"].to(self.device)
next_obs = batch["next_obs"].to(self.device)
done = batch["done"].to(self.device)
"""
Train the Behaviour cloning policy to be able to take more than 1 sample for MMD.
Conditional VAE is used as Behaviour cloning policy in BEAR.
"""
recon_action, mu, log_std = self.cvae_net(obs, acts)
cvae_loss = self.cvae_net.loss_function(recon_action, acts, mu,
log_std)
self.cvae_optimizer.zero_grad()
cvae_loss.backward()
self.cvae_optimizer.step()
"""
Critic Training
"""
with torch.no_grad():
# generate 10 actions for every next_obs(Same as BCQ)
next_obs = torch.repeat_interleave(next_obs,
repeats=self.n_target_samples,
dim=0).to(self.device)
# compute target Q value of generated action
target_q1 = self.target_q_net1(next_obs,
self.policy_net(next_obs)[0])
target_q2 = self.target_q_net2(next_obs,
self.policy_net(next_obs)[0])
# soft clipped double q-learning
target_q = self.lmbda * torch.min(target_q1, target_q2) + (
1. - self.lmbda) * torch.max(target_q1, target_q2)
# take max over each action sampled from the generation and perturbation model
target_q = target_q.reshape(obs.shape[0], self.n_target_samples,
1).max(1)[0].squeeze(1)
target_q = rews + self.gamma * (1. - done) * target_q
# compute current Q
current_q1 = self.q_net1(obs, acts).squeeze(1)
current_q2 = self.q_net2(obs, acts).squeeze(1)
# compute critic loss
critic_loss1 = F.mse_loss(current_q1, target_q)
critic_loss2 = F.mse_loss(current_q2, target_q)
self.q_optimizer1.zero_grad()
critic_loss1.backward()
self.q_optimizer1.step()
self.q_optimizer2.zero_grad()
critic_loss2.backward()
self.q_optimizer2.step()
# MMD Loss
# sample actions from dataset and current policy(B x N x D)
raw_sampled_actions = self.cvae_net.decode_multiple_without_squash(
obs, decode_num=self.n_mmd_action_samples, z_device=self.device)
raw_actor_actions = self.policy_net.sample_multiple_without_squash(
obs, sample_num=self.n_mmd_action_samples)
if self.kernel_type == 'gaussian':
mmd_loss = self.mmd_loss_gaussian(raw_sampled_actions,
raw_actor_actions,
sigma=self.mmd_sigma)
else:
mmd_loss = self.mmd_loss_laplacian(raw_sampled_actions,
raw_actor_actions,
sigma=self.mmd_sigma)
"""
Alpha prime training(lagrangian parameter update for MMD loss weight)
"""
alpha_prime_loss = -(self.log_alpha_prime.exp() *
(mmd_loss - self.lagrange_thresh)).mean()
self.alpha_prime_optimizer.zero_grad()
alpha_prime_loss.backward(retain_graph=True)
self.alpha_prime_optimizer.step()
self.log_alpha_prime.data.clamp_(min=-5.0,
max=10.0) # clip for stability
"""
Actor Training
Actor Loss = alpha_prime * MMD Loss + -minQ(s,a)
"""
a, log_prob, _ = self.policy_net(obs)
min_q = torch.min(self.q_net1(obs, a), self.q_net2(obs, a)).squeeze(1)
# policy_loss = (self.alpha * log_prob - min_q).mean() # SAC Type
policy_loss = -(min_q.mean())
# BEAR Actor Loss
actor_loss = (self.log_alpha_prime.exp() * mmd_loss).mean()
if self.train_step > self.warmup_step:
actor_loss = policy_loss + actor_loss
self.policy_optimizer.zero_grad()
actor_loss.backward(
) # the mmd_loss will backward again in alpha_prime_loss.
self.policy_optimizer.step()
soft_target_update(self.q_net1, self.target_q_net1, tau=self.tau)
soft_target_update(self.q_net2, self.target_q_net2, tau=self.tau)
self.train_step += 1
return critic_loss1.cpu().item(), critic_loss2.cpu().item(
), policy_loss.cpu().item(), alpha_prime_loss.cpu().item()
def train(self):
# Sample
batch = self.replay_buffer.sample()
obs = batch["obs"]
acts = batch["acts"]
rews = batch["rews"]
next_obs = batch["next_obs"]
done = batch["done"]
# compute policy Loss
a, log_prob = self.policy_net(obs)
min_q = torch.min(self.q_net1(obs, a), self.q_net2(obs, a)).squeeze(1)
policy_loss = (self.alpha * log_prob - min_q).mean()
# compute Q Loss
q1 = self.q_net1(obs, acts).squeeze(1)
q2 = self.q_net2(obs, acts).squeeze(1)
with torch.no_grad():
next_a, next_log_prob = self.policy_net(next_obs)
min_target_next_q = torch.min(self.target_q_net1(next_obs, next_a),
self.target_q_net2(
next_obs, next_a)).squeeze(1)
y = rews + self.gamma * (1. - done) * (min_target_next_q -
self.alpha * next_log_prob)
q_loss1 = F.mse_loss(q1, y)
q_loss2 = F.mse_loss(q2, y)
# Update policy network parameter(应该先更新策略网络，否则梯度不对)
self.policy_optimizer.zero_grad()
policy_loss.backward()
self.policy_optimizer.step()
# Update q network1 parameter
self.q_optimizer1.zero_grad()
q_loss1.backward()
self.q_optimizer1.step()
# Update q network2 parameter
self.q_optimizer2.zero_grad()
q_loss2.backward()
self.q_optimizer2.step()
if self.automatic_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()
self.alpha = self.log_alpha.exp()
else:
alpha_loss = torch.tensor(0)
self.train_step += 1
soft_target_update(self.q_net1, self.target_q_net1, tau=self.tau)
soft_target_update(self.q_net2, self.target_q_net2, tau=self.tau)
return q_loss1.item(), q_loss2.item(), policy_loss.item(
), alpha_loss.item()
def train(self):
# Sample
batch = self.data_buffer.sample()
obs = batch["obs"].to(self.device)
acts = batch["acts"].to(self.device)
rews = batch["rews"].to(self.device)
next_obs = batch["next_obs"].to(self.device)
done = batch["done"].to(self.device)
"""
SAC Loss
"""
# compute policy Loss
a, log_prob, _ = self.policy_net(obs)
min_q = torch.min(self.q_net1(obs, a), self.q_net2(obs, a)).squeeze(1)
policy_loss = (self.alpha * log_prob - min_q).mean()
# compute Q Loss
q1 = self.q_net1(obs, acts).squeeze(1)
q2 = self.q_net2(obs, acts).squeeze(1)
with torch.no_grad():
if not self.max_q_backup:
next_a, next_log_prob, _ = self.policy_net(next_obs)
min_target_next_q = torch.min(self.target_q_net1(next_obs, next_a),
self.target_q_net2(next_obs, next_a)).squeeze(1)
if self.entropy_backup:
# y = rews + self.gamma * (1. - done) * (min_target_next_q - self.alpha * next_log_prob)
min_target_next_q = min_target_next_q - self.alpha * next_log_prob
else:
"""when using max q backup"""
next_a_temp, _ = self.get_policy_actions(next_obs, n_action_samples=10)
target_qf1_values = self.get_actions_values(next_obs, next_a_temp, self.n_action_samples, self.q_net1).max(1)[0]
target_qf2_values = self.get_actions_values(next_obs, next_a_temp, self.n_action_samples, self.q_net2).max(1)[0]
min_target_next_q = torch.min(target_qf1_values, target_qf2_values).squeeze(1)
y = rews + self.gamma * (1. - done) * min_target_next_q
q_loss1 = F.mse_loss(q1, y)
q_loss2 = F.mse_loss(q2, y)
"""
CQL Loss
Total Loss = SAC loss + min_q_weight * CQL loss
"""
# Use importance sampling to compute log sum exp of Q(s, a), which is shown in paper's Appendix F.
random_sampled_actions = torch.FloatTensor(obs.shape[0] * self.n_action_samples, acts.shape[-1]).uniform_(-1, 1).to(self.device)
curr_sampled_actions, curr_log_probs = self.get_policy_actions(obs, self.n_action_samples)
# This is different from the paper because it samples not only from the current state, but also from the next state
next_sampled_actions, next_log_probs = self.get_policy_actions(next_obs, self.n_action_samples)
q1_rand = self.get_actions_values(obs, random_sampled_actions, self.n_action_samples, self.q_net1)
q2_rand = self.get_actions_values(obs, random_sampled_actions, self.n_action_samples, self.q_net2)
q1_curr = self.get_actions_values(obs, curr_sampled_actions, self.n_action_samples, self.q_net1)
q2_curr = self.get_actions_values(obs, curr_sampled_actions, self.n_action_samples, self.q_net2)
q1_next = self.get_actions_values(obs, next_sampled_actions, self.n_action_samples, self.q_net1)
q2_next = self.get_actions_values(obs, next_sampled_actions, self.n_action_samples, self.q_net2)
random_density = np.log(0.5 ** acts.shape[-1])
cat_q1 = torch.cat([q1_rand - random_density, q1_next - next_log_probs, q1_curr - curr_log_probs], dim=1)
cat_q2 = torch.cat([q2_rand - random_density, q2_next - next_log_probs, q2_curr - curr_log_probs], dim=1)
min_qf1_loss = torch.logsumexp(cat_q1, dim=1).mean()
min_qf2_loss = torch.logsumexp(cat_q2, dim=1).mean()
min_qf1_loss = self.min_q_weight * (min_qf1_loss - q1.mean())
min_qf2_loss = self.min_q_weight * (min_qf2_loss - q2.mean())
if self.with_lagrange:
alpha_prime = torch.clamp(self.log_alpha_prime.exp(), min=0.0, max=1e6)
# the lagrange_thresh has no effect on the gradient of policy,
# but it has an effect on the gradient of alpha_prime
min_qf1_loss = alpha_prime * (min_qf1_loss - self.lagrange_thresh)
min_qf2_loss = alpha_prime * (min_qf2_loss - self.lagrange_thresh)
alpha_prime_loss = -(min_qf1_loss + min_qf2_loss) * 0.5
self.alpha_prime_optimizer.zero_grad()
alpha_prime_loss.backward(retain_graph=True) # the min_qf_loss will backward again latter, so retain graph.
self.alpha_prime_optimizer.step()
else:
alpha_prime_loss = torch.tensor(0)
q_loss1 = q_loss1 + min_qf1_loss
q_loss2 = q_loss2 + min_qf2_loss
"""
Update networks
"""
# Update policy network parameter
# policy network's update should be done before updating q network, or there will make some errors
self.policy_optimizer.zero_grad()
policy_loss.backward()
self.policy_optimizer.step()
# Update q network1 parameter
self.q_optimizer1.zero_grad()
q_loss1.backward(retain_graph=True)
self.q_optimizer1.step()
# Update q network2 parameter
self.q_optimizer2.zero_grad()
q_loss2.backward(retain_graph=True)
self.q_optimizer2.step()
if self.auto_alpha_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()
self.alpha = self.log_alpha.exp()
else:
alpha_loss = torch.tensor(0)
soft_target_update(self.q_net1, self.target_q_net1, tau=self.tau)
soft_target_update(self.q_net2, self.target_q_net2, tau=self.tau)
self.train_step += 1
return q_loss1.cpu().item(), q_loss2.cpu().item(), policy_loss.cpu().item(), alpha_loss.cpu().item(), alpha_prime_loss.cpu().item()
def train(self):
# Sample
batch = self.data_buffer.sample()
obs = batch["obs"].to(self.device)
acts = batch["acts"].to(self.device)
rews = batch["rews"].to(self.device)
next_obs = batch["next_obs"].to(self.device)
done = batch["done"].to(self.device)
"""
CVAE Loss (the generation model)
"""
recon_action, mu, log_std = self.cvae_net(obs, acts)
cvae_loss = self.cvae_net.loss_function(recon_action, acts, mu,
log_std)
self.cvae_optimizer.zero_grad()
cvae_loss.backward()
self.cvae_optimizer.step()
"""
Critic Loss
"""
with torch.no_grad():
# generate 10 actions for every next_obs
next_obs = torch.repeat_interleave(next_obs, repeats=10,
dim=0).to(self.device)
generated_action = self.cvae_net.decode(next_obs,
z_device=self.device)
# perturb the generated action
perturbed_action = self.target_perturbation_net(
next_obs, generated_action)
# compute target Q value of perturbed action
target_q1 = self.target_critic_net1(next_obs, perturbed_action)
target_q2 = self.target_critic_net2(next_obs, perturbed_action)
# soft clipped double q-learning
target_q = self.lmbda * torch.min(target_q1, target_q2) + (
1. - self.lmbda) * torch.max(target_q1, target_q2)
# take max over each action sampled from the generation and perturbation model
target_q = target_q.reshape(obs.shape[0], 10,
1).max(1)[0].squeeze(1)
target_q = rews + self.gamma * (1. - done) * target_q
# compute current Q
current_q1 = self.critic_net1(obs, acts).squeeze(1)
current_q2 = self.critic_net2(obs, acts).squeeze(1)
# compute critic loss
critic_loss1 = F.mse_loss(current_q1, target_q)
critic_loss2 = F.mse_loss(current_q2, target_q)
self.critic_optimizer1.zero_grad()
critic_loss1.backward()
self.critic_optimizer1.step()
self.critic_optimizer2.zero_grad()
critic_loss2.backward()
self.critic_optimizer2.step()
"""
Perturbation Loss
"""
generated_action_ = self.cvae_net.decode(obs, z_device=self.device)
perturbed_action_ = self.perturbation_net(obs, generated_action_)
perturbation_loss = -self.critic_net1(obs, perturbed_action_).mean()
self.perturbation_optimizer.zero_grad()
perturbation_loss.backward()
self.perturbation_optimizer.step()
"""
Update target networks
"""
soft_target_update(self.critic_net1,
self.target_critic_net1,
tau=self.tau)
soft_target_update(self.critic_net2,
self.target_critic_net2,
tau=self.tau)
soft_target_update(self.perturbation_net,
self.target_perturbation_net,
tau=self.tau)
self.train_step += 1
return cvae_loss.cpu().item(), (
critic_loss1 +
critic_loss2).cpu().item(), perturbation_loss.cpu().item()
def train(self):
# Sample
batch = self.data_buffer.sample()
obs = batch["obs"].to(self.device)
acts = batch["acts"].to(self.device)
rews = batch["rews"].to(self.device)
next_obs = batch["next_obs"].to(self.device)
done = batch["done"].to(self.device)
# compute policy Loss
a, log_prob, _ = self.policy_net(obs)
min_q = torch.min(self.q_net1(obs, a), self.q_net2(obs, a)).squeeze(1)
policy_loss = (self.alpha * log_prob - min_q).mean()
# compute Q Loss
q1 = self.q_net1(obs, acts).squeeze(1)
q2 = self.q_net2(obs, acts).squeeze(1)
with torch.no_grad():
next_a, next_log_prob, _ = self.policy_net(next_obs)
min_target_next_q = torch.min(self.target_q_net1(next_obs, next_a),
self.target_q_net2(
next_obs, next_a)).squeeze(1)
y = rews + self.gamma * (1. - done) * (min_target_next_q -
self.alpha * next_log_prob)
q_loss1 = F.mse_loss(q1, y)
q_loss2 = F.mse_loss(q2, y)
# Update policy network parameter
# policy network's update should be done before updating q network, or there will make some errors
self.policy_optimizer.zero_grad()
policy_loss.backward()
self.policy_optimizer.step()
# Update q network1 parameter
self.q_optimizer1.zero_grad()
q_loss1.backward()
self.q_optimizer1.step()
# Update q network2 parameter
self.q_optimizer2.zero_grad()
q_loss2.backward()
self.q_optimizer2.step()
if self.auto_alpha_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()
self.alpha = self.log_alpha.exp()
else:
alpha_loss = torch.tensor(0)
self.train_step += 1
soft_target_update(self.q_net1, self.target_q_net1, tau=self.tau)
soft_target_update(self.q_net2, self.target_q_net2, tau=self.tau)
return q_loss1.cpu().item(), q_loss2.cpu().item(), policy_loss.cpu(
).item(), alpha_loss.cpu().item()