"""
Deep Q-Network (DQN)
Paper link: https://www.nature.com/articles/nature14236
Implementation: MindSpore
"""
from argparse import Namespace
from mindspore import nn
from xuance.mindspore import ms, ops, Module, Tensor, optim
from xuance.mindspore.learners import Learner
[docs]
class DQN_Learner(Learner):
def __init__(self,
config: Namespace,
policy: Module,
callback):
super(DQN_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.gamma = config.gamma
self.sync_frequency = config.sync_frequency
self.mse_loss = nn.MSELoss()
self.gather = ops.Gather(batch_dims=-1)
self.n_actions = int(self.policy.action_dim)
# Get gradient function
self.grad_fn = ms.value_and_grad(self.forward_fn, None, self.optimizer.parameters, has_aux=True)
self.policy.set_train()
self.grad_reducer = self.get_grad_reducer(self.optimizer)
[docs]
def forward_fn(self, obs_batch, act_batch, next_batch, rew_batch, ter_batch):
_, _, evalQ = self.policy(obs_batch)
_, _, targetQ = self.policy.target(next_batch)
targetQ = targetQ.max(axis=-1)
targetQ = rew_batch + self.gamma * (1 - ter_batch) * targetQ
predictQ = self.gather(evalQ, act_batch, axis=-1).reshape(-1)
loss = self.mse_loss(logits=predictQ, labels=ops.stop_gradient(targetQ))
return loss, evalQ, targetQ, predictQ
[docs]
def update(self, **samples):
self.iterations += 1
obs_batch = Tensor(samples['obs'])
act_batch = Tensor(samples['actions'].reshape(-1, 1), dtype=ms.int32)
rew_batch = Tensor(samples['rewards'])
next_batch = Tensor(samples['obs_next'])
ter_batch = Tensor(samples['terminals'])
info = self.callback.on_update_start(self.iterations,
policy=self.policy, obs=obs_batch, act=act_batch,
next_obs=next_batch, rew=rew_batch, termination=ter_batch)
(loss, evalQ, targetQ, predictQ), grads = self.grad_fn(obs_batch, act_batch, next_batch, rew_batch, ter_batch)
if self.distributed_training:
grads = self.grad_reducer(grads)
self.optimizer(grads)
# hard update for target network
if self.iterations % self.sync_frequency == 0:
self.policy.copy_target()
self.scheduler.step()
lr = self.scheduler.get_last_lr()[0]
info.update({
"Qloss": loss.asnumpy(),
"predictQ": predictQ.mean().asnumpy(),
"learning_rate": lr.asnumpy(),
})
info.update(self.callback.on_update_end(self.iterations,
policy=self.policy, info=info,
evalQ=evalQ, predictQ=predictQ, targetQ=targetQ, loss=loss))
return info