def _loop(self, is_train):
# internal loop for both training and validation
done = False
state_before = self.env.reset() if is_train else self.val_env.reset()
loss_list = []
reward_list = []
while not done:
epsilon = self.agent.epsilon if is_train else self._val_epsilon
action = self.agent.policy(state_before, epsilon)
state_after, reward, done, _ = self.env.step(
action) if is_train else self.val_env.step(action)
if is_train:
self._step += 1
self.memory(
Transition(state_before, action, reward, state_after,
done))
self.agent.parameter_scheduler(self._step)
loss_list.append(self._train_nn())
state_before = state_after
reward_list.append(reward)
if self._step % self.target_update_freq == 0 and is_train:
self.agent.update_target_net()
if self._step % self._val_freq == 0 and is_train:
self.val()
return loss_list, reward_list, state_after
def warm_up(self):
# to populate replay memory
state_before = self.env.reset()
self._warmed = True
for _ in tqdm(range(self.replay_start)):
action = self.env.action_space.sample()
state_after, reward, done, _ = self.env.step(action)
self.memory(Transition(state_before, action, reward, state_after, done))
state_before = self.env.reset() if done else state_after