"""
Proximal Policy Optimization (PPO) with clip trick
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.learners import Learner
[docs]
class PPO_Learner(Learner):
def __init__(self,
config: Namespace,
policy: Module,
callback):
super(PPO_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.total_iters)
# Parameters
self.mse_loss = nn.MSELoss()
self.vf_coef = config.vf_coef
self.ent_coef = config.ent_coef
self.clip_range = config.clip_range
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 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 forward_fn(self, obs_batch, act_batch, ret_batch, adv_batch, old_log_prob_batch):
if self.is_continuous:
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)
entropy = self.a_dist._entropy(mean=mu, sd=std)
entropy = ops.reduce_sum(x=entropy, axis=-1)
else:
outputs, logits, v_pred = self.policy(obs_batch)
probs = self.softmax(logits)
log_prob = self.a_dist._log_prob(value=act_batch, probs=probs)
entropy = self.a_dist.entropy(probs=probs)
ratio = ops.exp(log_prob - old_log_prob_batch)
surrogate1 = ops.clip_by_value(ratio, 1.0 - self.clip_range, 1.0 + self.clip_range) * adv_batch
surrogate2 = adv_batch * ratio
a_loss = -ops.mean(ops.minimum(surrogate1, surrogate2))
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, v_pred, ratio, log_prob, surrogate1, surrogate2
[docs]
def update(self, **samples):
self.iterations += 1
obs_batch = Tensor(samples['obs'], dtype=ms.float32)
ret_batch = Tensor(samples['returns'], dtype=ms.float32)
adv_batch = Tensor(samples['advantages'], dtype=ms.float32)
old_log_prob_batch = Tensor(samples['aux_batch']['old_logp'], dtype=ms.float32)
if self.is_continuous:
act_batch = Tensor(samples['actions'], dtype=ms.float32)
else:
act_batch = Tensor(samples['actions'], dtype=ms.int32)
info = self.callback.on_update_start(self.iterations,
policy=self.policy, obs=obs_batch, act=act_batch,
returns=ret_batch, advantages=adv_batch, old_logp=old_log_prob_batch)
(loss, a_loss, c_loss, e_loss, outputs, v_pred, ratio, log_prob, surrogate1, surrogate2), grads = self.grad_fn(
obs_batch, act_batch, ret_batch, adv_batch, old_log_prob_batch)
if self.use_grad_clip:
grads = ops.clip_by_norm(grads, self.grad_clip_norm)
self.optimizer(grads)
# Logger
self.scheduler.step()
lr = self.scheduler.get_last_lr()[0]
cr = ((ratio < 1 - self.clip_range).sum() + (ratio > 1 + self.clip_range).sum()) / ratio.shape[0]
info.update({
"actor_loss": a_loss.asnumpy(),
"critic_loss": c_loss.asnumpy(),
"entropy": e_loss.asnumpy(),
"learning_rate": lr.asnumpy(),
"predict_value": v_pred.mean().asnumpy(),
"clip_ratio": cr.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,
ratio=ratio, surrogate1=surrogate1, surrogate2=surrogate2,
a_loss=a_loss, c_loss=c_loss, e_loss=e_loss, loss=loss))
return info