"""
Proximal Policy Optimization with KL divergence (PPO-KL)
Paper link: https://arxiv.org/pdf/1707.06347.pdf
Implementation: Pytorch
"""
import torch
import numpy as np
from torch import nn
from xuance.torch.learners import Learner
from argparse import Namespace
from xuance.torch.utils import merge_distributions
[docs]
class PPOKL_Learner(Learner):
def __init__(self,
config: Namespace,
policy: nn.Module,
callback):
super(PPOKL_Learner, self).__init__(config, policy, callback)
self.optimizer = torch.optim.Adam(self.policy.parameters(), self.config.learning_rate, eps=1e-5)
self.scheduler = torch.optim.lr_scheduler.LinearLR(self.optimizer,
start_factor=1.0,
end_factor=self.end_factor_lr_decay,
total_iters=self.total_iters)
self.mse_loss = nn.MSELoss()
self.vf_coef = config.vf_coef
self.ent_coef = config.ent_coef
self.target_kl = config.target_kl
self.kl_coef = config.kl_coef
[docs]
def estimate_total_iterations(self):
"""Estimated total number of training iterations"""
buffer_size = self.config.horizon_size * self.config.parallels
update_times = self.config.running_steps // buffer_size
total_iters = update_times * self.config.n_epochs * self.config.n_minibatch
return total_iters
[docs]
def update(self, **samples):
self.iterations += 1
obs_batch = torch.as_tensor(samples['obs'], device=self.device)
act_batch = torch.as_tensor(samples['actions'], device=self.device)
ret_batch = torch.as_tensor(samples['returns'], device=self.device)
adv_batch = torch.as_tensor(samples['advantages'], device=self.device)
old_dists = samples['aux_batch']['old_dist']
info = self.callback.on_update_start(self.iterations,
policy=self.policy, obs=obs_batch, act=act_batch,
returns=ret_batch, advantages=adv_batch, old_dists=old_dists)
outputs, a_dist, v_pred = self.policy(obs_batch)
log_prob = a_dist.log_prob(act_batch)
old_dist = merge_distributions(old_dists)
kl = a_dist.kl_divergence(old_dist).mean()
old_logp_batch = old_dist.log_prob(act_batch)
# ppo-clip core implementations
ratio = (log_prob - old_logp_batch).exp().float()
a_loss = -(ratio * adv_batch).mean() + self.kl_coef * kl
c_loss = self.mse_loss(v_pred, ret_batch)
e_loss = a_dist.entropy().mean()
loss = a_loss - self.ent_coef * e_loss + self.vf_coef * c_loss
if kl > self.target_kl * 1.5:
self.kl_coef = self.kl_coef * 2.
elif kl < self.target_kl * 0.5:
self.kl_coef = self.kl_coef / 2.
self.kl_coef = np.clip(self.kl_coef, 0.1, 20)
self.optimizer.zero_grad()
loss.backward()
if self.use_grad_clip:
torch.nn.utils.clip_grad_norm_(self.policy.parameters(), self.grad_clip_norm)
self.optimizer.step()
if self.scheduler is not None:
self.scheduler.step()
# Logger
lr = self.optimizer.state_dict()['param_groups'][0]['lr']
if self.distributed_training:
info.update({
f"actor-loss/rank_{self.rank}": a_loss.item(),
f"critic-loss/rank_{self.rank}": c_loss.item(),
f"entropy/rank_{self.rank}": e_loss.item(),
f"learning_rate/rank_{self.rank}": lr,
f"kl/rank_{self.rank}": kl.item(),
f"predict_value/rank_{self.rank}": v_pred.mean().item()
})
else:
info.update({
"actor-loss": a_loss.item(),
"critic-loss": c_loss.item(),
"entropy": e_loss.item(),
"learning_rate": lr,
"kl": kl.item(),
"predict_value": v_pred.mean().item()
})
info.update(self.callback.on_update_end(self.iterations,
policy=self.policy, info=info, rep_output=outputs,
a_dist=a_dist, v_pred=v_pred, log_prob=log_prob, kl=kl, ratio=ratio,
a_loss=a_loss, c_loss=c_loss, e_loss=e_loss, loss=loss))
return info