"""
Proximal Policy Optimization with KL divergence (PPO-KL)
Paper link: https://arxiv.org/pdf/1707.06347.pdf
Implementation: MindSpore
"""
from argparse import Namespace
from mindspore import nn
from xuance.mindspore import ms, msd, ops, Module, Tensor, optim
from xuance.mindspore.utils import merge_distributions
from xuance.mindspore.learners import Learner
[docs]
class PPOKL_Learner(Learner):
def __init__(self,
config: Namespace,
policy: Module,
callback):
super(PPOKL_Learner, self).__init__(config, policy, callback)
self.optimizer = optim.Adam(params=self.policy.trainable_params(), lr=self.config.learning_rate, eps=1e-5)
self.scheduler = optim.lr_scheduler.LinearLR(self.optimizer, start_factor=1.0, end_factor=self.end_factor_lr_decay,
total_iters=self.config.running_steps)
self.vf_coef = config.vf_coef
self.ent_coef = config.ent_coef
self.target_kl = config.target_kl
self.kl_coef = Tensor(config.kl_coef)
self.clip_range = config.clip_range
self.mse_loss = nn.MSELoss()
self.softmax = nn.Softmax(axis=-1)
self.is_continuous = self.policy.is_continuous
self.a_dist = msd.Normal(dtype=ms.float32) if self.is_continuous else msd.Categorical()
# Get gradient function
self.grad_fn = ms.value_and_grad(self.forward_fn, None, self.optimizer.parameters, has_aux=True)
self.policy.set_train()
[docs]
def forward_fn(self, *args):
if self.is_continuous:
obs_batch, act_batch, ret_batch, adv_batch, old_mu, old_std = args
outputs, mu, std, v_pred = self.policy(obs_batch)
log_prob = self.a_dist._log_prob(value=act_batch, mean=mu, sd=std)
log_prob = ops.reduce_sum(x=log_prob, axis=-1)
old_log_prob = self.a_dist._log_prob(value=act_batch, mean=old_mu, sd=old_std)
old_log_prob = old_log_prob.squeeze(-1)
entropy = self.a_dist._entropy(mean=mu, sd=std)
entropy = ops.reduce_sum(x=entropy, axis=-1)
kl = self.a_dist._kl_loss("Normal", mean_b=old_mu, sd_b=old_std, mean=mu, sd=std)
else:
obs_batch, act_batch, ret_batch, adv_batch, old_logits = args
outputs, logits, v_pred = self.policy(obs_batch)
probs = self.softmax(logits)
log_prob = self.a_dist._log_prob(value=act_batch, probs=probs)
old_probs = self.softmax(old_logits)
old_log_prob = self.a_dist._log_prob(value=act_batch, probs=old_probs)
entropy = self.a_dist.entropy(probs=probs)
kl = self.a_dist._kl_loss("Categorical", probs_b=old_probs, probs=probs)
ratio = ops.exp(log_prob - old_log_prob)
kl = ops.reduce_mean(kl)
a_loss = -ops.mean(ratio * adv_batch) + self.kl_coef * kl
c_loss = self.mse_loss(logits=v_pred, labels=ops.stop_gradient(ret_batch))
e_loss = ops.mean(entropy)
loss = a_loss - self.ent_coef * e_loss + self.vf_coef * c_loss
return loss, a_loss, c_loss, e_loss, outputs, log_prob, ratio, kl, v_pred
[docs]
def update(self, **samples):
self.iterations += 1
obs_batch = Tensor(samples['obs'])
ret_batch = Tensor(samples['returns'])
adv_batch = Tensor(samples['advantages'])
if self.is_continuous:
act_batch = Tensor(samples['actions'], dtype=ms.float32)
else:
act_batch = Tensor(samples['actions'], dtype=ms.int32)
old_dists = merge_distributions(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)
if self.is_continuous:
old_mu = old_dists.mu
old_std = old_dists.std
(loss, a_loss, c_loss, e_loss, outputs, log_prob, ratio, kl, v_pred), grads = self.grad_fn(
obs_batch, act_batch, ret_batch, adv_batch, old_mu, old_std)
else:
old_logits = old_dists.logits
(loss, a_loss, c_loss, e_loss, outputs, log_prob, ratio, kl, v_pred), grads = self.grad_fn(
obs_batch, act_batch, ret_batch, adv_batch, old_logits)
if self.use_grad_clip:
grads = ops.clip_by_norm(grads, self.grad_clip_norm)
self.optimizer(grads)
self.scheduler.step()
lr = self.scheduler.get_last_lr()[0]
if kl.asnumpy() > self.target_kl * 1.5:
self.kl_coef = self.kl_coef * 2.
elif kl.asnumpy() < self.target_kl * 0.5:
self.kl_coef = self.kl_coef / 2.
self.kl_coef = ops.clip_by_value(self.kl_coef, 0.1, 20)
info.update({
"actor-loss": a_loss.asnumpy(),
"critic-loss": c_loss.asnumpy(),
"entropy": e_loss.asnumpy(),
"learning_rate": lr.asnumpy(),
"kl": kl.asnumpy(),
"predict_value": v_pred.mean().asnumpy()
})
info.update(self.callback.on_update_end(self.iterations,
policy=self.policy, info=info, rep_output=outputs,
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