# MIT License
# Copyright (c) 2021 MARL @ SJTU
# Author: Ming Zhou
# Permission is hereby granted, free of charge, to any person obtaining a copy
# of this software and associated documentation files (the "Software"), to deal
# in the Software without restriction, including without limitation the rights
# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
# copies of the Software, and to permit persons to whom the Software is
# furnished to do so, subject to the following conditions:
# The above copyright notice and this permission notice shall be included in all
# copies or substantial portions of the Software.
# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
# SOFTWARE.
from argparse import Namespace
from typing import Dict, Any, Sequence
import itertools
import numpy as np
import torch
from torch import optim
from torch import nn
from torch.nn import functional as F
from malib.utils.typing import AgentID
from malib.utils.data import to_torch
from malib.rl.common.trainer import Trainer
from malib.utils.data import Postprocessor
from malib.utils.tianshou_batch import Batch
from malib.utils.statistic import RunningMeanStd
[docs]class A2CTrainer(Trainer):
[docs] def setup(self):
parameter_dict = self.policy.parameters()
# concate parameters
parameters = set(itertools.chain(*parameter_dict.values()))
self.optimizer = getattr(optim, self.training_config["optimizer"])(
parameters, lr=self.training_config["lr"]
)
self.parameters = parameters
self.lr_scheduler: torch.optim.lr_scheduler.LambdaLR = None
# runtime return averaging
self.ret_rms = RunningMeanStd()
[docs] def post_process(self, batch: Batch, agent_filter: Sequence[AgentID]) -> Batch:
state_value, next_state_value = [], []
with torch.no_grad():
for minibatch in batch.split(
self.training_config.get("max_gae_batchsize", 256),
shuffle=False,
merge_last=True,
):
state_value.append(self.policy.critic(minibatch.obs))
next_state_value.append(self.policy.critic(minibatch.obs_next))
batch["state_value"] = (
torch.cat(state_value, dim=0).flatten().cpu().numpy()
) # old value
state_value = batch["state_value"]
next_state_value = torch.cat(next_state_value, dim=0).flatten().cpu().numpy()
if self.training_config[
"reward_norm"
]: # unnormalize state_value & next_state_value
eps = self.training_config["reward_norm"]
state_value = state_value * np.sqrt(self.ret_rms.var + eps)
next_state_value = next_state_value * np.sqrt(self.ret_rms.var + eps)
unnormalized_returns, advantages = Postprocessor.compute_episodic_return(
batch,
state_value,
next_state_value,
self.training_config["gamma"],
self.training_config["gae_lambda"],
)
if self.training_config["reward_norm"]:
batch["returns"] = unnormalized_returns / np.sqrt(self.ret_rms.var + eps)
self.ret_rms.update(unnormalized_returns)
else:
batch["returns"] = unnormalized_returns
# batch.returns = to_torch_as(batch.returns, batch.state_value)
batch["advantage"] = advantages # to_torch_as(advantages, batch.state_value)
assert (
batch.advantage.shape == batch.state_value.shape == batch.returns.shape
), (batch.advantage.shape, batch.state_value.shape, batch.returns.shape)
batch["logits"], _ = self.policy.actor(
batch.obs, state=batch.get("state", None)
)
return batch
[docs] def train(self, batch: Batch) -> Dict[str, float]:
batch = Namespace(
**{k: to_torch(v, device=self.policy.device) for k, v in batch.items()}
)
# calculate loss for actor
logits = batch.logits
dist = self.policy.dist_fn.proba_distribution(logits)
log_prob = dist.log_prob(batch.act)
log_prob = log_prob.reshape(len(batch.advantage), -1).transpose(0, 1)
actor_loss = -(log_prob * batch.advantage).mean()
# calculate loss for critic
value = self.policy.critic(batch.obs).flatten()
vf_loss = F.mse_loss(batch.returns, value)
# calculate regularization and overall loss
ent_loss = dist.entropy().mean()
loss = (
actor_loss
+ self.training_config["value_coef"] * vf_loss
- self.training_config["entropy_coef"] * ent_loss
)
self.optimizer.zero_grad()
loss.backward()
if self.training_config.get("grad_norm", 0): # clip large gradient
nn.utils.clip_grad_norm_(
self.parameters, max_norm=self.training_config["grad_norm"]
)
self.optimizer.step()
if self.lr_scheduler is not None:
self.lr_scheduler.step()
return {
"actor_loss": actor_loss.item(),
"vf_loss": vf_loss.item(),
"ent_loss": ent_loss.item(),
"total_loss": loss.item(),
}