from argparse import Namespace
from typing import List
import torch
from torch import nn
from xuance.torch.learners.multi_agent_rl.ippo_learner import IPPO_Learner
[docs]
class CommNet_Learner(IPPO_Learner):
def __init__(self,
config: Namespace,
model_keys: List[str],
agent_keys: List[str],
policy: nn.Module,
callback):
super(CommNet_Learner, self).__init__(config, model_keys, agent_keys, policy, callback)
[docs]
def build_optimizer(self):
self.optimizer = torch.optim.Adam(self.policy.parameters_model, lr=self.learning_rate, eps=1e-5,
weight_decay=self.config.weight_decay)
self.scheduler = torch.optim.lr_scheduler.LinearLR(self.optimizer,
start_factor=1.0,
end_factor=self.end_factor_lr_decay,
total_iters=self.config.lr_decay_steps)
[docs]
def update_rnn(self, sample):
self.iterations += 1
info = {}
sample_Tensor = self.build_training_data(sample=sample,
use_parameter_sharing=self.use_parameter_sharing,
use_actions_mask=self.use_actions_mask)
batch_size = sample_Tensor['batch_size']
bs_rnn = batch_size * self.n_agents if self.use_parameter_sharing else batch_size
obs = sample_Tensor['obs']
actions = sample_Tensor['actions']
values = sample_Tensor['values']
returns = sample_Tensor['returns']
advantages = sample_Tensor['advantages']
log_pi_old = sample_Tensor['log_pi_old']
avail_actions = sample_Tensor['avail_actions']
agent_mask = sample_Tensor['agent_mask']
filled = sample_Tensor['filled']
seq_len = filled.shape[1]
IDs = sample_Tensor['agent_ids']
if self.use_parameter_sharing:
key = self.model_keys[0]
# agent_mask: [batch_size*self.n_agents, seq_length]
alive_ally = agent_mask[key].view(batch_size, self.n_agents, seq_len).unsqueeze(-1)
alive_ally = {k: alive_ally[:, i] for i, k in enumerate(self.agent_keys)}
else:
alive_ally = {k: agent_mask[k].unsqueeze(-1) for k in self.model_keys}
if self.use_parameter_sharing:
key = self.model_keys[0]
filled = filled.unsqueeze(1).expand(batch_size, self.n_agents, seq_len).reshape(bs_rnn, seq_len)
joint_obs = obs[key].reshape(batch_size, self.n_agents, seq_len, -1).transpose(
1, 2).reshape(batch_size, seq_len, -1)
joint_obs = joint_obs.unsqueeze(1).expand(-1, self.n_agents, -1, -1).reshape(bs_rnn, seq_len, -1)
critic_input = {key: joint_obs}
else:
joint_obs = self.get_joint_input(obs, (batch_size, seq_len, -1))
critic_input = {k: joint_obs for k in self.agent_keys}
# feedfowrd
rnn_hidden_actor = {k: self.policy.actor_representation[k].init_hidden(bs_rnn) for k in self.model_keys}
rnn_hidden_critic = {k: self.policy.critic_representation[k].init_hidden(bs_rnn) for k in self.model_keys}
# feedforward
_, pi_dist_dict = self.policy(obs, agent_ids=IDs, avail_actions=avail_actions, rnn_hidden=rnn_hidden_actor, alive_ally=alive_ally)
_, value_pred_dict = self.policy.get_values(observation=obs, agent_ids=IDs, rnn_hidden=rnn_hidden_critic, alive_ally=alive_ally)
# calculate losses for each agent
loss_a, loss_e, loss_c = [], [], []
for key in self.model_keys:
mask_values = agent_mask[key] * filled
log_pi = pi_dist_dict[key].log_prob(actions[key]).reshape(bs_rnn, seq_len)
ratio = torch.exp(log_pi - log_pi_old[key])
surrogate1 = ratio * advantages[key]
surrogate2 = torch.clip(ratio, 1 - self.clip_range, 1 + self.clip_range) * advantages[key]
loss_a.append(-(torch.min(surrogate1, surrogate2) * mask_values).sum() / mask_values.sum())
# entropy loss
entropy = pi_dist_dict[key].entropy().reshape(bs_rnn, seq_len)
entropy = entropy * mask_values
loss_e.append(entropy.sum() / mask_values.sum())
# critic loss
value_pred_i = value_pred_dict[key].reshape(bs_rnn, seq_len)
value_target = returns[key].reshape(bs_rnn, seq_len)
values_i = values[key].reshape(bs_rnn, seq_len)
if self.use_value_clip:
value_clipped = values_i + (value_pred_i - values_i).clamp(-self.value_clip_range,
self.value_clip_range)
if self.use_value_norm:
self.value_normalizer[key].update(value_target.reshape(-1, 1))
value_target = self.value_normalizer[key].normalize(value_target.reshape(-1, 1))
value_target = value_target.reshape(bs_rnn, seq_len)
if self.use_huber_loss:
loss_v = self.huber_loss(value_pred_i, value_target)
loss_v_clipped = self.huber_loss(value_clipped, value_target)
else:
loss_v = (value_pred_i - value_target) ** 2
loss_v_clipped = (value_clipped - value_target) ** 2
loss_c_ = torch.max(loss_v, loss_v_clipped) * mask_values
loss_c.append(loss_c_.sum() / mask_values.sum())
else:
if self.use_value_norm:
self.value_normalizer[key].update(value_target)
value_target = self.value_normalizer[key].normalize(value_target)
if self.use_huber_loss:
loss_v = self.huber_loss(value_pred_i, value_target)
else:
loss_v = (value_pred_i - value_target) ** 2
loss_c.append((loss_v * mask_values).sum() / mask_values.sum())
info.update({
f"{key}/actor_loss": loss_a[-1].item(),
f"{key}/critic_loss": loss_c[-1].item(),
f"{key}/entropy": loss_e[-1].item(),
f"{key}/predict_value": value_pred_i.mean().item()
})
loss = sum(loss_a) + self.vf_coef * sum(loss_c) - self.ent_coef * sum(loss_e)
self.optimizer.zero_grad()
loss.backward()
if self.use_grad_clip:
grad_norm = torch.nn.utils.clip_grad_norm_(self.policy.parameters_model, self.grad_clip_norm)
info["gradient_norm"] = grad_norm.item()
self.optimizer.step()
if self.scheduler is not None:
self.scheduler.step()
# Logger
lr = self.optimizer.state_dict()['param_groups'][0]['lr']
info.update({
"learning_rate": lr,
"loss": loss.item(),
})
return info