def process_fn(self, batch: Batch, buffer: ReplayBuffer,
indice: np.ndarray) -> Batch:
if self._rew_norm:
mean, std = batch.rew.mean(), batch.rew.std()
if not np.isclose(std, 0, 1e-2):
batch.rew = (batch.rew - mean) / std
v, v_, old_log_prob = [], [], []
with torch.no_grad():
for b in batch.split(self._batch, shuffle=False):
v_.append(self.critic(b.obs_next))
v.append(self.critic(b.obs))
old_log_prob.append(self(b).dist.log_prob(
to_torch_as(b.act, v[0])))
v_ = to_numpy(torch.cat(v_, dim=0))
batch = self.compute_episodic_return(
batch, v_, gamma=self._gamma, gae_lambda=self._lambda,
rew_norm=self._rew_norm)
batch.v = torch.cat(v, dim=0).flatten() # old value
batch.act = to_torch_as(batch.act, v[0])
batch.logp_old = torch.cat(old_log_prob, dim=0)
batch.returns = to_torch_as(batch.returns, v[0])
batch.adv = batch.returns - batch.v
if self._rew_norm:
mean, std = batch.adv.mean(), batch.adv.std()
if not np.isclose(std.item(), 0, 1e-2):
batch.adv = (batch.adv - mean) / std
return batch
def _compute_returns(self, batch: Batch, buffer: ReplayBuffer,
indice: np.ndarray) -> Batch:
v_s, v_s_ = [], []
with torch.no_grad():
for b in batch.split(self._batch, shuffle=False, merge_last=True):
v_s.append(self.critic(b.obs))
v_s_.append(self.critic(b.obs_next))
batch.v_s = torch.cat(v_s, dim=0).flatten() # old value
v_s = batch.v_s.cpu().numpy()
v_s_ = torch.cat(v_s_, dim=0).flatten().cpu().numpy()
# when normalizing values, we do not minus self.ret_rms.mean to be numerically
# consistent with OPENAI baselines' value normalization pipeline. Emperical
# study also shows that "minus mean" will harm performances a tiny little bit
# due to unknown reasons (on Mujoco envs, not confident, though).
if self._rew_norm: # unnormalize v_s & v_s_
v_s = v_s * np.sqrt(self.ret_rms.var + self._eps)
v_s_ = v_s_ * np.sqrt(self.ret_rms.var + self._eps)
unnormalized_returns, advantages = self.compute_episodic_return(
batch,
buffer,
indice,
v_s_,
v_s,
gamma=self._gamma,
gae_lambda=self._lambda)
if self._rew_norm:
batch.returns = unnormalized_returns / \
np.sqrt(self.ret_rms.var + self._eps)
self.ret_rms.update(unnormalized_returns)
else:
batch.returns = unnormalized_returns
batch.returns = to_torch_as(batch.returns, batch.v_s)
batch.adv = to_torch_as(advantages, batch.v_s)
return batch
def process_fn(
self, batch: Batch, buffer: ReplayBuffer, indice: np.ndarray
) -> Batch:
v_s, v_s_, old_log_prob = [], [], []
with torch.no_grad():
for b in batch.split(self._batch, shuffle=False, merge_last=True):
v_s.append(self.critic(b.obs))
v_s_.append(self.critic(b.obs_next))
old_log_prob.append(self(b).dist.log_prob(to_torch_as(b.act, v_s[0])))
batch.v_s = torch.cat(v_s, dim=0).flatten() # old value
v_s = to_numpy(batch.v_s)
v_s_ = to_numpy(torch.cat(v_s_, dim=0).flatten())
if self._rew_norm: # unnormalize v_s & v_s_
v_s = v_s * np.sqrt(self.ret_rms.var + self._eps) + self.ret_rms.mean
v_s_ = v_s_ * np.sqrt(self.ret_rms.var + self._eps) + self.ret_rms.mean
unnormalized_returns, advantages = self.compute_episodic_return(
batch, buffer, indice, v_s_, v_s,
gamma=self._gamma, gae_lambda=self._lambda)
if self._rew_norm:
batch.returns = (unnormalized_returns - self.ret_rms.mean) / \
np.sqrt(self.ret_rms.var + self._eps)
self.ret_rms.update(unnormalized_returns)
mean, std = np.mean(advantages), np.std(advantages)
advantages = (advantages - mean) / std # per-batch norm
else:
batch.returns = unnormalized_returns
batch.act = to_torch_as(batch.act, batch.v_s)
batch.logp_old = torch.cat(old_log_prob, dim=0)
batch.returns = to_torch_as(batch.returns, batch.v_s)
batch.adv = to_torch_as(advantages, batch.v_s)
return batch
def process_fn(self, batch: Batch, buffer: ReplayBuffer,
indices: np.ndarray) -> Batch:
batch = super().process_fn(batch, buffer, indices)
old_log_prob = []
with torch.no_grad():
for b in batch.split(self._batch, shuffle=False, merge_last=True):
old_log_prob.append(self(b).dist.log_prob(b.act))
batch.logp_old = torch.cat(old_log_prob, dim=0)
if self._norm_adv:
batch.adv = (batch.adv - batch.adv.mean()) / batch.adv.std()
return batch
def learn(self, batch: Batch, batch_size: int, repeat: int,
**kwargs) -> Dict[str, List[float]]:
self._batch = batch_size
losses, clip_losses, vf_losses, ent_losses = [], [], [], []
v = []
old_log_prob = []
with torch.no_grad():
for b in batch.split(batch_size, shuffle=False):
v.append(self.critic(b.obs))
old_log_prob.append(
self(b).dist.log_prob(
torch.tensor(b.act, device=v[0].device)))
batch.v = torch.cat(v, dim=0) # old value
dev = batch.v.device
batch.act = torch.tensor(batch.act, dtype=torch.float, device=dev)
batch.logp_old = torch.cat(old_log_prob, dim=0)
batch.returns = torch.tensor(batch.returns,
dtype=torch.float,
device=dev).reshape(batch.v.shape)
if self._rew_norm:
mean, std = batch.returns.mean(), batch.returns.std()
if std > self.__eps:
batch.returns = (batch.returns - mean) / std
batch.adv = batch.returns - batch.v
if self._rew_norm:
mean, std = batch.adv.mean(), batch.adv.std()
if std > self.__eps:
batch.adv = (batch.adv - mean) / std
for _ in range(repeat):
for b in batch.split(batch_size):
dist = self(b).dist
value = self.critic(b.obs)
ratio = (dist.log_prob(b.act) - b.logp_old).exp().float()
surr1 = ratio * b.adv
surr2 = ratio.clamp(1. - self._eps_clip,
1. + self._eps_clip) * b.adv
if self._dual_clip:
clip_loss = -torch.max(torch.min(surr1, surr2),
self._dual_clip * b.adv).mean()
else:
clip_loss = -torch.min(surr1, surr2).mean()
clip_losses.append(clip_loss.item())
if self._value_clip:
v_clip = b.v + (value - b.v).clamp(-self._eps_clip,
self._eps_clip)
vf1 = (b.returns - value).pow(2)
vf2 = (b.returns - v_clip).pow(2)
vf_loss = .5 * torch.max(vf1, vf2).mean()
else:
vf_loss = .5 * (b.returns - value).pow(2).mean()
vf_losses.append(vf_loss.item())
e_loss = dist.entropy().mean()
ent_losses.append(e_loss.item())
loss = clip_loss + self._w_vf * vf_loss - self._w_ent * e_loss
losses.append(loss.item())
self.optim.zero_grad()
loss.backward()
nn.utils.clip_grad_norm_(
list(self.actor.parameters()) +
list(self.critic.parameters()), self._max_grad_norm)
self.optim.step()
return {
'loss': losses,
'loss/clip': clip_losses,
'loss/vf': vf_losses,
'loss/ent': ent_losses,
}
