def policy(task, step):
policy = Sequential(task)
print('Step: ', step.__name__)
print('Input: ', policy.task.observs)
policy.add(step)
print('Output:', policy.task.actions)
policy.add(Random)
return policy
def train_policies(self):
trainers = []
for _ in range(self.config.learners):
config = AttrDict(self.config.copy())
model = Model(self._create_network, threads=1)
model.weights = self.model.weights
policy = Sequential(self.task)
policy.add(self._create_preprocess())
policy.add(Train, config, self, model)
trainers.append(policy)
return trainers
def _create_preprocess(self):
policy = Sequential(self.task)
# Network preprocessing.
# policy.add(Skip, self.config.frameskip)
# policy.add(Maximum, 2)
# policy.add(Grayscale)
# # policy.add(Subsample, (2, 2))
# # policy.add(Delta)
# policy.add(History, 3)
# policy.add(ClampReward)
# policy.add(EpsilonGreedy, from_=0.5, to=0.5, test=0.5)
policy.add(Normalize)
return policy
def test_policy(self):
policy = Sequential(self.task)
policy.add(self._preprocess)
policy.add(Test, self.model)
return policy
def _create_preprocess(self):
policy = Sequential(self.task)
if self.config.noop_max:
policy.add(RandomStart, self.config.noop_max)
if self.config.frame_skip:
policy.add(Skip, self.config.frame_skip)
if self.config.frame_max:
policy.add(Maximum, self.config.frame_max)
if self.config.history:
policy.add(Grayscale)
if self.config.subsample > 1:
sub = self.config.subsample
amount = (sub, sub) if self.config.history else (sub, sub, 1)
policy.add(Subsample, amount)
if self.config.delta:
policy.add(Delta)
if self.config.history:
policy.add(History, self.config.history)
policy.add(ClampReward)
policy.add(Normalize)
return policy
def _create_preprocess(self):
policy = Sequential(self.task)
policy.add(Image)
if self.config.noop_max:
policy.add(RandomStart, self.config.noop_max)
if self.config.frame_skip > 1:
policy.add(Skip, self.config.frame_skip)
if self.config.frame_max:
policy.add(Maximum, self.config.frame_max)
if self.config.history > 1:
channels = policy.above_task.observs.shape[-1]
policy.add(Grayscale, (0.299, 0.587, 0.114)[:channels])
if self.config.subsample > 1:
sub = self.config.subsample
amount = (sub, sub) if self.config.history > 1 else (sub, sub, 1)
policy.add(Subsample, amount)
if self.config.delta:
policy.add(Delta)
if self.config.history > 1:
policy.add(History, self.config.history)
policy.add(Normalize)
policy.add(ClampReward)
policy.add(EpsilonGreedy, **self.config.epsilon)
return policy
def policy(self):
# TODO: Why doesn't self.task work here?
policy = Sequential(self._preprocess.task)
policy.add(self._preprocess)
policy.add(self)
return policy
def policy(self):
policy = Sequential(self._preprocess.task)
policy.add(self._preprocess)
policy.add(self)
return policy
def _prepend_score_step(self, policy):
combined = Sequential(policy.task)
combined.add(Score)
combined.add(policy)
return combined